Legal teams face two problems with AI at the same time. The technology is moving fast, and the conversation around it is moving faster. Most of what gets published swings between breathless enthusiasm and existential alarm. Neither framing helps legal professionals who need to make sound decisions about specific tools in specific workflows with real clients and real consequences, whether they work in a law firm, a corporate legal department, an alternative legal service provider, or a legal technology company. The actual challenge is not whether to use AI. Legal teams are already using it, often embedded in contract platforms, research tools, and workflow systems they rely on daily. The challenge is how to use it responsibly given specific, well-documented risks. Three concerns consistently surface across bar guidance, ethics opinions, and experience using these tools: accuracy and hallucinations, confidentiality and data security, and ethics, accountability, and bias. These are not theoretical. They have already produced court sanctions, ethics complaints, and client harm. This article explains what each risk involves, where current professional guidance stands, and what responsible adoption looks like in practice. How Generative AI Actually Works Most concerns about AI in legal work trace back to a single mechanical reality. Generative AI systems produce content by predicting likely continuations based on patterns learned from vast amounts of data. They do not understand meaning, verify facts, or reason from first principles. They estimate what plausible output looks like, and they are remarkably good at it. This distinction explains nearly every concern that follows. Accuracy problems arise because the system optimizes for plausibility, not truth. Confidentiality exposure arises because data must travel to the system for processing. Accountability questions arise because the output looks authoritative even when it is wrong. Understanding these mechanics is not about becoming a technologist. It is about knowing enough to ask the right questions and recognize when something requires closer scrutiny. Accuracy, Reliability, and Hallucinations Generative models do not retrieve information the way a search engine queries a database. They generate responses by predicting what is most likely to come next. When they encounter gaps in their training data or ambiguous questions, they do not pause or flag uncertainty. They generate plausible output anyway. This is hallucination: confident, well-structured content that is factually wrong. In legal work, it manifests as fabricated case citations, misstated holdings, invented statutory provisions, incorrect regulatory references, and overstated factual claims. The term "hallucination" can be misleading because it implies something unusual. In practice, it is an inherent characteristic of how these systems operate. Every output is a prediction. Some predictions are accurate. Others are not. The system itself cannot tell the difference. The reliability problem compounds the accuracy problem. The same prompt can produce different outputs on different occasions. A contract clause that an AI drafts correctly on Monday may contain a material error on Tuesday. Past accuracy is not a reliable indicator of future performance, which makes spot-checking an inadequate substitute for systematic verification. Several well-publicized court sanctions have resulted from lawyers filing briefs containing AI-generated citations to cases that do not exist. But the accuracy risk extends well beyond litigation. In-house teams relying on AI-drafted contract language may miss nonstandard terms. Compliance professionals using AI to interpret regulatory requirements may act on incorrect guidance. Legal operations teams automating intake or triage with AI may misroute matters based on flawed analysis. What makes this particularly dangerous is the quality of the output. Hallucinated content often reads as entirely credible. The citation format is correct. The reasoning sounds coherent. Catching errors requires the same verification effort as checking accurate work, which means efficiency gains can evaporate if verification is not built into the workflow from the start. Courts are responding with certification requirements for AI-assisted filings. Clients are asking how their legal service providers use AI and what quality controls are in place. Where Professional Guidance Stands Multiple bar associations now treat AI output the way legal organizations have always treated work from a junior professional: useful for drafts and initial research, but requiring independent verification before use. ABA Formal Opinion 512, along with emerging state-level guidance, makes the obligation explicit: lawyers bear full responsibility for accuracy regardless of whether AI generated the content. The duty of competence now extends to understanding how these systems function and where they are likely to fail. For organizations providing legal services, this means building institutional competence, not just individual awareness. What Responsible Practice Looks Like Treat every AI output as a draft. This is not a hedge. It is the operational reality of working with systems that optimize for plausibility rather than correctness. Build verification into the workflow at the point of creation, not as an afterthought. Require independent confirmation of citations, factual assertions, and legal conclusions before client delivery, whether that takes the form of a court filing, a contract, a compliance memo, or a matter recommendation. Use AI for what it does well: organizing information, producing first passes, and accelerating early-stage work. Apply human judgment where it is indispensable: verifying accuracy, assessing risk, and deciding what to trust. Confidentiality and Data Security When legal professionals input client information into AI systems, that data travels to third-party infrastructure. Depending on the vendor's architecture, client facts, drafts, and documents may be processed on external servers, stored in system logs, used to train or improve models, or accessible to vendor employees and subprocessors. This creates a confidentiality exposure that most legal professionals would never accept in other contexts. Pasting a sensitive fact pattern into a consumer AI tool is functionally equivalent to emailing it to an unknown third party with no confidentiality agreement in place. Yet this happens routinely because the interface feels private even when the infrastructure is not. Some systems retain conversation history. Others use inputs to improve model performance, meaning client data could influence outputs generated for other users. For in-house teams, ALSPs, and legal technology providers handling data from multiple clients or business units, the cross-contamination risk is particularly acute. The duty of confidentiality encompasses reasonable measures to prevent inadvertent exposure, not just intentional disclosure. Using AI tools without understanding their data handling practices can breach this duty even when no one intended to share anything. The risk scales with sensitivity. Regulated data carries heightened obligations. Attorney-client privileged communications require particular care because privilege can be waived through disclosure to third parties. Deal-related information demands strict controls because premature disclosure can have material consequences. Organizations handling data subject to HIPAA, GDPR, or sector-specific regulations face additional compliance layers that general-purpose AI tools do not account for. Where Professional Guidance Stands Ethics guidance now frames vendor due diligence for AI tools as part of the duty of confidentiality and competence. Lawyers and legal organizations must understand, before using a tool, where data goes, who can access it, how long it is retained, and whether deletion rights exist. The key questions: 1) Does the vendor use inputs to train models? Are there data localization requirements? 2) What subprocessors handle the data? 3) Can the organization audit data handling practices? 4) What happens to data after the engagement ends? These are not optional inquiries. They are part of the professional obligation to protect client information. What Responsible Practice Looks Like Conduct vendor security assessments before deploying AI tools in legal workflows. If a vendor cannot answer basic questions about data handling, retention, and deletion, that tells you something important. Establish clear policies governing what types of information can be entered into which tools. A tool approved for general research may not be appropriate for drafting involving specific client facts. A consumer-grade system should never touch privileged communications. Make these distinctions explicit and communicate them to everyone with access to your systems, including contractors and service providers. For sensitive or regulated workflows, consider enterprise deployments where the organization controls the infrastructure. Train teams on data hygiene: strip identifying details when possible and understand the difference between consumer and enterprise AI environments. Ethics, Accountability, and Bias Two distinct concerns converge here. The first is bias. AI systems trained on historical data can embed patterns reflecting existing inequities. In legal contexts, this matters for case outcome prediction, risk scoring, candidate screening, litigation strategy, and settlement valuation. A model that learned from biased data will produce biased outputs, and those outputs can drive decisions with real consequences for real people. For organizations serving diverse client populations or operating across jurisdictions, the stakes are compounded. The second is accountability. When AI generates a legal analysis, the responsible party under current guidance is the lawyer or supervising professional. But AI systems can make it easy to defer judgment without realizing it. The output looks polished, arrives quickly, and the temptation to forward it without careful review is real, particularly in high-volume environments where speed is a competitive advantage. These concerns are connected. Biased outputs that go unreviewed create both the harm and the accountability gap simultaneously. Professional guidance draws a clear distinction between AI as a tool that assists legal professionals and AI as a system that effectively provides legal analysis without adequate oversight. When a professional reviews AI output, evaluates its reasoning, and exercises independent judgment, the system functions as a tool. When someone forwards AI output without meaningful review, the system is functioning as the advisor. This line becomes harder to maintain with agentic AI systems that take actions across connected platforms. An agent that drafts a response, routes a matter, and updates a record creates multiple decision points where oversight can lapse if checkpoints are not designed in advance. The speed and autonomy that make agents valuable are the same characteristics that make governance essential. There is also the question of unauthorized practice. When AI systems provide legal analysis directly to non-lawyers without meaningful attorney oversight, they may cross from permissible automation into territory that raises unauthorized practice concerns. Organizations deploying AI in client-facing or self-service contexts need to think carefully about where this boundary falls and design accordingly. Where Professional Guidance Stands ABA Formal Opinion 512, along with state-level guidance, emphasizes that lawyers remain fully responsible for AI-assisted work. The duty of supervision applies to AI systems the same way it applies to non-lawyer staff. Emerging guidance addresses bias directly. Legal professionals and organizations that use AI tools producing biased or discriminatory outputs may face liability for downstream consequences, whether in litigation strategy or operational decisions like hiring and vendor selection. The standard is not perfection. It is diligence: understanding the tools, mitigating bias, and maintaining substantive oversight. What Responsible Practice Looks Like Maintain meaningful human review at every material decision point. Meaningful review means engaging with substance, not skimming formatting. Test AI tools for bias before deploying them in sensitive contexts. If a tool is scoring risk or screening candidates, validate its outputs against diverse scenarios. If you cannot evaluate a tool's outputs for bias, reconsider using it for that purpose. Document AI use in your workflows. Clear records of human review and professional judgment protect both the client and the organization. Documentation also creates accountability, which is the point. Define which tasks are appropriate for AI assistance and which require unaided professional judgment. Putting It Together The three concerns are interconnected in practice. Accuracy failures create accountability exposure. Confidentiality lapses erode client trust. Biased outputs generate liability. A responsible AI strategy addresses all three together. Evaluate  before adopting any tool. Assess accuracy characteristics, data handling practices, and potential for biased outputs. Ask vendors specific questions and expect specific answers. Design  workflows with verification checkpoints, data handling controls, and clear accountability at each step. Build oversight in from the start, not after deployment. Monitor  outcomes, not just usage. Track error rates. Audit for bias periodically. Governance is ongoing, not a one-time implementation. Train  everyone who uses AI tools on the risks, the organization's policies, and their professional responsibilities. Training is not a formality. It is how organizations close the gap between policy and practice. Judgment Is the Point AI changes what legal teams can accomplish. It accelerates drafting, organizes information, and compresses the time between question and first answer. These are real benefits. Responsible teams should capture them. But the value of legal work has never been speed alone. It is judgment: knowing what matters, what to trust, what to question, what to advise. AI systems cannot do this. They can make the work surrounding judgment more efficient, freeing professionals for the decisions that require expertise. The concerns about accuracy, confidentiality, and accountability are not reasons to avoid AI. They are reasons to adopt it deliberately, with clear processes, genuine oversight, and an honest understanding of what these systems can and cannot do. AI should extend judgment, not replace it.

We're drowning in information. Every legal tech vendor promises transformative results. Every AI tool claims to revolutionize practice. Meanwhile, clients are overwhelmed by options, partners are paralyzed by choices, and associates are buried in data they don't know how to interpret. Judgment is becoming one of the scarcest resources in the legal profession. Not the only thing that matters. Not a replacement for knowledge or experience. But increasingly, the differentiator between lawyers who add value and those who don't. The problem: we don't really know what we mean when we talk about judgment. We use it as a catchall for "good lawyering" without examining what it actually entails. In a world where algorithms can analyze cases faster than any human, where AI can spot patterns we'd miss, we need to get specific. If we can't define judgment precisely, we can't develop it systematically. And if we can't develop it, we'll struggle to explain why our judgment is worth paying for. Judgment Isn't What We Think It Is Ask most lawyers what judgment means and you'll get vague answers about "knowing the right thing to do" or "understanding the law deeply." These aren't wrong. They're incomplete. Judgment isn't primarily about legal knowledge, though it requires legal knowledge as a foundation. You can't exercise good judgment on securities issues without understanding securities law. But knowing securities law doesn't automatically give you good judgment about when to push a disclosure issue versus when to let it go. Judgment isn't about experience either, at least not directly. Experience informs judgment. But I've seen senior lawyers repeat the same mistakes for thirty years, and young lawyers who grasp strategic nuance that escapes their more seasoned colleagues. Experience creates opportunities to develop judgment. It doesn't guarantee it. So what is judgment? Judgment is the capacity to discern what matters from what doesn't in a specific context. To separate signal from noise. To know which facts are relevant and which are distractions. To understand when the law is the answer and when it's beside the point. Notice the qualifier: in a specific context. Good judgment isn't abstract wisdom. It's situation-specific discernment informed by knowledge, experience, and understanding of the particular client, industry, and circumstance. The Filtering Function In a world saturated with information, judgment functions primarily as a filter. When a client comes to you with a problem, they give you facts. Some relevant, some not. Some they think matter that don't. Some they don't mention that do. Your first job is figuring out which is which. AI can process all those facts. It can identify legal issues, flag risks, suggest precedents. What it can't reliably do is know which facts actually matter for this client, in this situation, with these specific goals and constraints. Last month I reviewed a distribution agreement where AI flagged seventeen potential issues. Technically, all seventeen were legitimate concerns and yet only three actually mattered. Knowing which three to focus on required judgment. But that judgment was only possible because I understood the business context, had seen similar deals, and knew what battles were worth fighting for companies in this position. Judgment Requires Foundation, Doesn't Replace It Judgment doesn't replace knowledge. It builds on it. A first-year associate with excellent judgment but no employment law knowledge still can't advise on employment matters. The foundation has to be there first. AI hasn't changed this. What's changed is that foundational knowledge alone isn't enough anymore. When AI can access and process legal information faster than any human, being the lawyer who knows the most case law isn't the competitive advantage it once was. The competitive advantage is knowing what to do with that information. When to apply it, when to question it, when to override it based on factors the algorithm can't see. Knowing What Questions to Ask Junior lawyers think judgment is about having answers. Senior lawyers know it's about asking the right questions. But this isn't mystical. Sometimes clients actually do know exactly what they need and asking them to reconsider the question is overthinking. The skill isn't always finding "the question behind the question." It's knowing when to look for it. AI is phenomenally good at answering questions. Ask it to find relevant case law, it will. Ask it to identify risks in a contract, it will. Ask it to draft language addressing a specific issue, it can. What AI can't reliably do yet is recognize when the question you're asking isn't the question you should be asking. When a startup founder asks about IP assignment provisions, sometimes they're asking about IP assignment provisions. Sometimes they're actually trying to navigate a deteriorating relationship with a co-founder. Knowing which situation you're in requires context AI doesn't have. This is judgment informed by specific knowledge about startup dynamics, founder relationships, and how IP disputes typically arise. Not pure intuition. The Confidence to Override the Algorithm One of the most important forms of judgment right now is knowing when to trust AI and when to ignore it. AI outputs come with an aura of objectivity. The algorithm analyzed thousands of contracts and says X is market. The model reviewed hundreds of cases and predicts Y outcome. Sometimes that's exactly what you need. Sometimes it's dangerously misleading. I've seen AI analyze employment contracts and flag certain non-compete provisions as "below market" based on data from public company agreements. Technically accurate. Completely irrelevant for a small business in a different industry with different competitive dynamics. Judgment is knowing the difference. But sometimes the algorithm is right and your intuition is wrong. Sometimes what you think is unique about your client's situation actually isn't, and the data-driven answer is better than your experience-based hunch. Good judgment requires both the confidence to override AI when you have good reason and the humility to recognize when you don't. That requires deep expertise. Not just in law, but in the domains where law applies. You can't effectively evaluate AI's contract analysis if you don't understand the business the contract serves. When Judgment Means Saying No Sometimes judgment means telling clients, colleagues, or partners that what they want to do is unwise. Not illegal. Not impossible. Just unwise. This is genuinely hard. We're trained to find ways to accomplish what clients want. We're incentivized economically to say yes. The lawyer who consistently says "you shouldn't do this" doesn't tend to get repeat business. Last year I advised a client against pursuing a technically viable breach of contract claim. The contract terms were clear. We'd likely win. But winning would destroy a relationship with their largest supplier during a supply chain crisis, and the damages we'd recover wouldn't offset the business damage. They didn't like the advice. They found another lawyer who would file the claim. Six months later, they settled for less than my initial recommendation and lost the supplier relationship anyway. Was that good judgment on my part? I think so. But I can't prove it. They might have won, recovered more, and maintained the relationship. Judgment calls don't come with certainty. We tend to talk about judgment as if experienced lawyers reliably make better calls than algorithms. Sometimes we do. Sometimes we're just confident in our mistakes. Cutting Through the Hype Judgment in legal tech right now means asking hard questions before adopting tools. Not "is this innovative?" but "does this solve a real problem we have?" Not "is everyone else buying this?" but "will this actually work for how we work?" I've seen firms implement contract analytics platforms that never get used because nobody defined what questions they needed answered. I've seen document automation systems fail because the templates they automated weren't the ones people actually used. I've seen AI tools get adopted because partners were afraid of missing out, not because anyone identified a specific problem to solve. Good judgment means being willing to say "I don't care if every other firm has this, it doesn't make sense for us." Or conversely, "I know this seems risky, but I think it addresses a real problem." Sometimes you can't know if something will work until you try it. Sometimes what looks like good judgment to pass on a tool is actually missing an opportunity. The firms that waited to see if "this internet thing" would matter learned that lesson. The Nuance That Algorithms Miss (For Now) Legal issues are rarely as clean as algorithms need them to be. The facts are messy. The law is ambiguous. The client's goals are complicated by competing interests and factors that never appear in the case file. Judgment is the capacity to navigate this complexity right now. To hold multiple competing considerations in mind simultaneously. To recognize that the right answer depends on factors you can't quantify. An algorithm can tell you that based on thousands of cases, you have a 73% chance of winning this motion. Judgment tells you whether winning this motion is worth the relationship cost with the judge you'll appear before ten more times this year. AI is getting better at this kind of analysis. The gap between algorithmic and human judgment on pattern recognition tasks is narrowing. The things we claim AI "can't" do today, it might do tomorrow. So the question isn't whether humans will always be better at nuanced judgment. It's whether we can develop and demonstrate our judgment capabilities fast enough to stay ahead of improving algorithms. Teaching Judgment in a Tech-Driven World If judgment is this important, we need to figure out how to develop it systematically. Associates used to develop pattern recognition through document review. They'd review thousands of contracts and gradually learn what mattered, what didn't, what partners focused on, what they ignored. If AI does that work now, how do junior lawyers develop that pattern recognition? I don't have proven answers. Nobody does yet. Some hypotheses worth testing: Give associates real responsibility earlier on matters where stakes are manageable. Let them make judgment calls, defend their reasoning, be wrong, and learn from it. This is risky. Clients aren't paying for junior lawyers to learn through trial and error. But if we don't create opportunities to develop judgment, we won't develop lawyers who have it. Make senior lawyers explicit about their judgment process. Don't just say "this provision doesn't matter." Explain why it doesn't matter, what factors you're weighing, how you reached that conclusion. Most partners share conclusions but not reasoning. That doesn't teach judgment. Create structured decision points where associates must evaluate what matters. What are the three most important issues in this contract? Why? What's the biggest risk the client faces? What are you not worried about and why? Will this work? I don't know. We're in uncharted territory. Assuming judgment develops naturally without deliberate training is probably wrong. The Speed Problem (And It's Complicated) Good judgment requires time to think. Whether that's a problem depends entirely on where you work. In-house, speed matters. When the business team needs an answer to close a deal, "let me think about this for three days" isn't always viable. In-house lawyers operate in the business's timeline, not the lawyer's preferred pace. Technology that handles routine work faster creates genuine value because it frees time for the judgment calls that can't be rushed. Law firms operate under different economics. The billable hour rewards thoroughness, not speed. A partner who takes three days to think through a strategy bills more than one who reaches the same conclusion in three hours. Whether that serves the client is debatable. Whether it serves the firm's revenue is not. This creates a real tension. Judgment often benefits from reflection time. But law firm economics don't always reward taking that time, while in-house pressures don't always allow it. When AI handles routine work, what happens to that time? In-house counsel might genuinely use it for strategic thinking. Law firms might just fill it with more billable work. I'm not saying law firms are deliberately inefficient. But let's be honest about the incentives. A firm that becomes dramatically more efficient through technology faces a revenue problem unless it can raise rates proportionally or attract more clients. Many firms solve this by maintaining roughly the same level of labor intensity even with better tools. In-house doesn't face this problem. If technology makes in-house counsel more efficient, the company benefits directly through reduced outside counsel spend or better risk management. The incentives align with actual efficiency. This matters for judgment because it affects how much time lawyers actually have for deliberative thinking. Technology creates the possibility of more time for judgment. Whether that possibility becomes reality depends on economic structure. What Judgment Means for the Profession As AI handles more of what lawyers traditionally did, judgment becomes increasingly important. But it's not the only thing clients pay for. Clients hire lawyers for many reasons: regulatory requirements, risk management, insurance mandates, court rules, credentialing, fiduciary duties, and yes, judgment. Even if AI develops better judgment than humans on some dimensions, lawyers won't become obsolete. But our role will change. The question is whether we can adapt. Whether we can: 1) Develop judgment systematically rather than hoping it emerges from experience. Demonstrate value in ways clients recognize and will pay for. 2) Build economic models that reward judgment appropriately. 3) Train lawyers whose judgment capabilities exceed what algorithms can provide. Remain honest about our limitations while confident about our capabilities. This means being clear about what judgment actually is: not vague wisdom, but specific capabilities like filtering signal from noise, asking right questions, knowing when to override algorithms, navigating nuance, integrating contextual information. It means acknowledging that judgment requires deep foundational knowledge. And it means accepting that some of what we've called judgment might actually have been pattern recognition that AI will eventually do better. The judgment that matters is the kind that requires contextual understanding, relationship awareness, and integration of information that doesn't fit into data fields. The Reality Check In a world where everyone has access to information, where AI can analyze data faster than any human, where legal knowledge is increasingly accessible, judgment becomes more important. But claiming judgment justifies our fees isn't enough. We have to deliver it. Develop it. Demonstrate it. That requires being honest about what judgment is and isn't. About what we know and don't know. About when we're exercising genuine expertise-based judgment and when we're guessing confidently. It means creating ways to develop judgment in lawyers learning the profession alongside AI. It means building practice models that value judgment enough to give it time and space. It means being humble enough to recognize when algorithms know better and confident enough to override them when they don't. Technology isn't replacing judgment. It's making judgment more important and more visible. The question isn't whether judgment matters. It's whether we can develop, demonstrate, and deliver it at the level clients need and will pay for. I think we can. But only if we're honest about the challenge we're facing.

What It Means to Be a Lawyer When Machines Can Think I've been thinking a lot lately about what it means to be a lawyer in the age of AI. Not in the breathless, either/or way that dominates so many conversations, where we're either celebrating the death of billable hours or mourning the death of the profession itself. But in a quieter, more honest way. Because here's what I keep coming back to: the rise of AI doesn't change what it means to be a lawyer. It reveals it. The Work We Do vs. The Value We Bring For too long, we've confused lawyering with the tasks lawyers perform. We've measured our worth in hours billed, documents reviewed, contracts redlined. We've built entire careers around the assumption that the value we provide is directly proportional to the time we spend doing repetitive, rules-based work. AI is forcing us to confront an uncomfortable truth: much of what we've charged premium rates for isn't actually the irreplaceable human judgment we claimed it was. It was pattern recognition, risk assessment, and document processing. Important work, certainly. But work that algorithms can increasingly handle with speed and consistency that humans simply can't match. This isn't a threat. It's a liberation. But here's where it gets complicated. Liberation only comes to those willing to acknowledge what they've been liberated from. Many lawyers remain resistant because accepting AI's capabilities means accepting that significant portions of their training and expertise can be replicated by software. That's uncomfortable. It challenges our professional identity. It forces us to ask hard questions about what we're actually worth if machines can do what we've been doing. Humanity Has Always Been the Point When I tell people that lawyers need to remember their inherent humanity, I'm not being sentimental. I'm being practical. Because the things AI cannot do, the things no algorithm will ever replicate, are precisely the things that have always made great lawyers great. Empathy. The ability to understand not just what a client is saying, but what they're afraid to say. To recognize when legal advice needs to make room for human emotion, business reality, or ethical complexity. Judgment. Not the algorithmic kind that processes probabilities, but the seasoned kind that knows when to push, when to yield, and when the technically correct answer isn't the right one. Creativity. The capacity to see novel solutions, to reframe problems, to understand that law is ultimately about human relationships and human problems, not just statutes and precedents. Trust. Clients don't hire lawyers because they need someone to process information. They hire lawyers because they need someone they can trust when everything is uncertain and the stakes are high. These capabilities aren't abstract ideals. They're practical skills that determine whether a deal gets done, whether a dispute gets resolved, whether a client's business succeeds or fails. And they're precisely what gets buried when we're drowning in routine work that technology should be handling anyway. The Challenge of Change Management Here's something most discussions about AI in law get wrong: they focus on the technology when the real challenge is human. Technology advances faster than human willingness to adopt it. This isn't a coincidence. Humans crave comfort, certainty, and familiarity. We prefer established practices over uncharted territory. I've seen this firsthand. Legal teams invest substantial resources in AI tools that then sit unused. The financial waste is obvious, but the deeper problem is what happens to trust. When lawyers see expensive technology gather dust, they become skeptical of future innovations. They tell themselves they're "too busy" to learn new tools or they "aren't tech people." What they're really saying is: change is hard, and I'm afraid this won't work. The solution isn't better technology. It's better change management. It's taking the time to understand how legal professionals actually work, what they're afraid of, and what success looks like from their perspective. It's meeting people where they are rather than where we think they should be. The Model Needs Reimagining We're at a crossroads, and the path forward requires honest self-examination. The traditional model, built on billable hours and associate leverage, was never designed for a world where technology handles the routine work. It was designed to maximize the economic value of having humans do that routine work. That world is ending. Good. What we need is a model that puts our humanity first and our work second, not the other way around. A model where technology handles the time-consuming, repeatable, low-risk tasks, freeing us to focus on the strategic, high-value work that actually requires human insight. But more than that, we need a model that acknowledges we are not automatons. We need time to refresh, to recharge, to maintain the mental clarity that genuine judgment requires. Life is too short and too precious to be consumed by work that machines can handle better than we can anyway. The economic reality is this: clients are already demanding value over volume. They're bringing work in-house, using alternative providers, and treating innovation as a proxy for value. The firms that thrive will be those that recognize AI doesn't just improve efficiency. It fundamentally changes what clients are willing to pay for and what they expect from their lawyers. AI Augments, It Doesn't Replace I've said this before and I'll keep saying it: AI is not replacing lawyers. It's revealing what lawyering actually is. When AI drafts a contract in seconds, it doesn't diminish the lawyer who understands the business relationship that contract needs to serve. When AI analyzes thousands of precedents, it doesn't replace the lawyer who knows which precedent actually matters for this client, in this situation, with these particular risks. The legal profession doesn't need to resist AI. It needs to embrace what AI makes possible: a return to the human elements of law practice that got buried under mountains of document review and routine drafting. But let's be honest about the limitations. AI can process documents rapidly but lacks the nuanced understanding required for comprehensive counsel. Biases in training data produce biased outputs. Capabilities remain narrow and unpredictable. Algorithms can tell you what contracts typically say; they can't tell you what this contract should say for this client's unique situation. That's not a flaw in the technology. It's a feature of what makes human expertise irreplaceable. The lawyers who understand this distinction, who know where AI excels and where human judgment remains essential, will be the ones who use these tools most effectively. Education Must Evolve Law schools bear responsibility here too. We're still training students for a legal world that largely doesn't exist anymore. We're teaching them to think like lawyers who manually review every document and research every issue from scratch. What we should be teaching is critical thinking around AI's applications and ethical implications. Students need tech literacy and an appreciation of AI's limits. They need to understand how to prompt these systems effectively, how to audit outputs for bias and error, how to know when to trust the algorithm and when to override it. More fundamentally, they need to learn the skills that will differentiate them: strategic thinking, business acumen, client relationship management, creative problem-solving. These aren't nice-to-have soft skills anymore. They're the core competencies that justify hiring a lawyer instead of buying software. The Practice Must Be Less About Us If I could change one thing about the practice of law, it would be this: the practice needs to be less about the lawyers and more about the clients. The clients should be the ones leading the relationship. AI makes this shift possible in ways we couldn't have imagined a decade ago. When technology handles the mundane, we have time to actually listen. To understand. To advise in ways that acknowledge our clients are human beings facing human problems, not just legal issues requiring legal solutions. Think about what becomes possible when you're not buried in routine work. You can develop deeper relationships with your clients. You can understand their business strategy well enough to provide genuinely strategic counsel. You can spot issues before they become problems. You can add value in ways that matter to them, not just in ways we've traditionally measured legal work. This is what clients have wanted all along. They didn't want lawyers who could spend 40 hours reviewing a contract. They wanted lawyers who understood their business well enough to know what risks actually mattered and what terms would actually work. What the Future Looks Like The future of law isn't about lawyers versus machines. It's about lawyers who understand that technology is a tool that can make us more effective at being human. It's about recognizing that the hours we spend drafting routine agreements could be spent understanding our clients' businesses. That the time we pour into research could be spent on strategic thinking. That the energy we burn on repetitive tasks could be channeled into creative problem-solving. The lawyers who will thrive aren't the ones who resist AI or the ones who blindly embrace it. They're the ones who understand that AI gives us permission to finally focus on what humans have always done best: bringing wisdom, judgment, empathy, and creativity to complex human problems. This future requires us to be honest about what we don't know. To admit when we need to learn new skills. To accept that the way we've always done things isn't necessarily the way things should be done. That's uncomfortable for a profession built on precedent and tradition. But it's necessary. We're Still Human Beings I struggle sometimes with anxiety and depression. I share this not for sympathy but because it's part of being human. And being human is not something we should leave at the door when we practice law. The rise of AI is forcing us to answer a question we've avoided for too long: What are we actually here for? What value do we bring that can't be measured in pages reviewed or hours billed? The answer, I think, is simple. We're here to be human. To bring our full humanity, our judgment, our empathy, our creativity to bear on the problems clients face. To remember that behind every contract, every dispute, every transaction, there are human beings trying to build something, protect something, or solve something. AI doesn't threaten that. It enables it. Technology facilitates change. But it's our humanity that gives that change meaning and direction. As we forge ahead into this AI-augmented future, our challenge isn't mastering the technology. It's remembering who we are and what we're really here to do. We're lawyers. We're advisors. We're problem-solvers. We're humans. And that's never going to change.

Introduction AI agents represent a meaningful shift in how legal teams use technology. Traditional generative AI responds to prompts by producing content. Agents go further. They act. By combining language models with goals, tools, and memory, agents execute multi-step workflows across connected systems. For legal teams, this expands automation beyond isolated tasks into coordinated processes. Agents can review intake requests, extract key facts, validate deadlines against matter management systems, propose task assignments, and draft communications. Each step informs the next. This is not conversational assistance. It is software operating toward defined outcomes. That capability creates leverage by reducing coordination work and allowing professionals to focus on judgment-intensive decisions. It also introduces new risk. Because agents operate across systems, weak permissions, unclear escalation logic, or poor controls can propagate errors quickly. This article explains how agentic systems work, where they are already being used in legal operations, and how to adopt them responsibly with a focus on security and governance. Understanding the Shift Most legal teams encounter generative AI through tools that summarize contracts, assist with research, or support drafting. These systems follow a simple pattern: prompt in, response out. Agentic AI operates differently. Instead of producing a single response, an agent pursues a goal across multiple steps. It can call tools, access systems, retain context, and decide what to do next based on results along the way. The difference is clear when comparing a chatbot asked to summarize a contract with an agent instructed to process contract intake. The chatbot produces a summary. The agent may classify the request, select the appropriate template, compare terms against standards, flag deviations, route the matter to the correct reviewer, and log the action. This shift matters because it changes both capability and risk. Agents manage workflows, not just tasks. They move across systems, not just within one. Governance models must reflect that expanded scope. What Makes an Agent An AI agent consists of four core elements: a language model, tools, memory, and goals. The language model provides reasoning. It interprets instructions, plans steps, and generates outputs such as summaries or drafts. While this is the same underlying technology used in chatbots, here it functions as part of a larger system. Tools enable action. They allow agents to search databases, read documents, send messages, update records, and interact with enterprise platforms. Each tool is an access point that requires authorization and logging. Memory provides continuity. It allows agents to track prior actions, maintain context, and avoid repeating work. Memory may persist only within a task or across sessions, raising questions about retention and handling of sensitive information. Goals direct behavior. Unlike chatbots, which react to prompts, agents work toward defined objectives. Who sets those objectives and how they are constrained are core governance concerns. How Agents Operate Agents operate in iterative loops. At each step, the agent assesses its current state, determines the next action, executes a tool, evaluates the result, and decides whether to continue. This loop continues until the objective is met or a stopping condition applies. For example, an agent preparing a contract review package may identify the agreement type, retrieve the relevant playbook, compare terms, flag deviations, draft a summary, route the materials, and record the activity. Each decision depends on what the agent learned in the prior step. Agents are adaptive rather than scripted. They adjust based on inputs and outcomes. This flexibility distinguishes them from traditional automation and explains why governance must emphasize oversight, constraints, and accountability rather than static rules alone. Agent Architectures and Governance Implications Single-agent systems assign all steps to one agent. These systems work well for focused workflows such as intake processing or document review. Governance is relatively straightforward because permissions and audit trails are centralized. Multi-agent systems divide work among specialized agents, coordinated through an orchestration layer. This mirrors how legal teams operate but introduces governance challenges around data handoffs, accountability, and agent-to-agent communication. Human-in-the-loop designs embed review at defined decision points. Agents handle routine processing but pause before taking consequential actions. These checkpoints define acceptable autonomy. Too many reduce efficiency. Too few increase exposure to error. Practical Use Cases in Legal Legal teams are already deploying agents across several workflow categories, each with distinct governance requirements. Contract Intake and Triage Agents can classify submissions, extract metadata, validate completeness, and route matters based on risk or value. Governance focuses on routing logic and access control. Playbook Compliance Review Agents can compare agreements against internal standards and flag deviations. Version control is critical to avoid applying outdated positions. Due Diligence Coordination Agents can track requests, monitor data rooms, and route documents to subject matter experts. Strong access controls and audit trails are essential. Legal Research Assembly Agents can retrieve authorities and assemble structured research outputs. Governance must address cost controls, citation quality, and currency. Matter Management Updates Agents can update systems of record based on communications and documents. Write access requires strict safeguards to prevent inaccurate records. Compliance Monitoring Agents can track regulatory updates and contractual obligations, generating alerts and routing tasks with documented human oversight. Security Architecture for Agent Deployment Agent deployments extend across systems in ways that traditional security models may not anticipate. Agents should have distinct identities, narrowly scoped permissions, and clear audit trails. Network segmentation and API gateways limit exposure and provide control points for monitoring. Data protection depends on classification, encryption, and minimization. Not all data should be accessible to agents, and retention should be limited. When third-party services are involved, teams must assess where processing occurs, who can access logs, and how data is retained. Sensitive workflows may justify on-premises deployment. Governance Frameworks for Agent Operations Security addresses access. Governance addresses accountability, quality, compliance, and oversight. Policies should define acceptable use, approval thresholds, and prohibited applications. Standards translate policy into measurable requirements. Monitoring should track performance, quality, compliance, and escalation behavior. Incident response plans should address detection, containment, and remediation of agent errors. Training ensures stakeholders understand what agents do, where limits apply, and how accountability is assigned. Agents support work. Humans remain responsible. Risk Management Considerations Agents should operate with minimum necessary permissions enforced technically, not just documented. Escalation paths must reflect value, risk, and uncertainty and route matters to appropriate expertise.Audit trails should allow reconstruction of actions and decisions. Agents are not accountable entities. Stopping conditions prevent runaway execution. Ongoing testing detects model drift and failure modes.Vendor dependencies require contingency planning for outages, pricing changes, or service degradation. Implementation Guidance Effective adoption progresses from simple to complex. Start by mapping workflows and identifying friction points. Begin with bounded tasks that have clear inputs and outputs. Maintain human review during early deployment.Measure outcomes such as time savings, error rates, rework, and escalation frequency. Use evidence to refine configurations and expand scope gradually. Agent deployment is an ongoing discipline, not a one-time implementation. Concluding Thoughts Agentic AI changes what automation can accomplish in legal work. Agents coordinate, maintain context, and execute multi-step processes across systems. That power requires deliberate governance. Clear permissions, escalation logic, accountability, and security controls enable scale without sacrificing trust. Governance does not slow adoption. It enables sustainable adoption. When done well, agents extend human judgment rather than replacing it. Glossary AI Agent Software that combines a language model with goals, tools, and memory to execute multi-step tasks autonomously within defined limits. Language Model The reasoning component that interprets instructions, plans actions, and generates outputs such as summaries, drafts, or classifications. Tool Use An agent’s ability to interact with external systems, including document repositories, databases, email, and enterprise applications. Agent Memory The mechanism that allows an agent to retain context about prior actions and decisions during or across workflows. Human-in-the-Loop A workflow design where agents pause at defined points for human review before taking consequential actions. Permission Boundary The enforced scope of systems, data, and actions an agent is authorized to access or perform. Audit Trail A detailed record of agent actions and tool usage that supports review, accountability, and incident response. Frequently Asked Questions How is an agent different from a chatbot? A chatbot responds to a single prompt. An agent pursues a goal across multiple steps, uses tools, retains context, and decides what to do next. What legal workflows benefit most from agents? Work that involves coordination rather than judgment, such as intake triage, document routing, compliance tracking, and structured reviews. Do agents operate without human oversight? They can, but responsible deployments define when agents must pause and escalate, especially for high-risk or high-value actions. Are agents secure by default? No. Security depends on how permissions, system access, monitoring, and logging are designed and enforced. Who is accountable for agent actions? The organization remains accountable. Agents execute instructions but do not assume responsibility. Can agents access sensitive or privileged information? Only when explicitly permitted. Many teams restrict agents from handling certain data types by design. Where should teams begin with agent adoption? Start with narrow, well-defined tasks, maintain human review early, and expand scope based on observed performance.

Introduction Most confusion about AI in legal teams comes from misunderstanding how these systems operate. When mechanics remain opaque, risk discussions drift toward extremes. Either AI feels magical, or it feels reckless. Neither position supports responsible adoption. Clarity about how generative models, agents, and multimodal systems function changes everything. Once teams understand what these systems are doing under the hood, conversations about use cases, guardrails, and accountability become more grounded and productive. Consider a common scenario: a team enables AI-powered contract summarization and quickly sees faster turnaround on first drafts. Weeks later, inconsistent summaries begin to surface because no one defined review standards or ownership. The issue is not the technology. It is the absence of process design around it. Understanding mechanics helps teams design workflows that assume review, judgment, and accountability from the start. How Generative AI Works Traditional AI focuses on classification. It answers questions like whether an email is spam or which category a document belongs to. Generative AI is fundamentally different. It produces new output such as text, images, code, and summaries. During training, models analyze massive volumes of data and adjust internal numerical weights to become highly effective at predicting likely continuations. They do not store concepts or beliefs. They learn statistical relationships between words, shapes, sounds, and patterns. When you provide a prompt, your input converts into internal representations. The model then generates output step by step. At each moment, attention mechanisms determine which parts of the context matter most. This allows responses to remain coherent across long passages or complex instructions. Outputs feel intelligent because models absorb patterns from human-created material. But they do not understand meaning or truth. They estimate plausibility. This distinction explains both their strengths and their limitations. For example, a system can rapidly draft a commercial clause based on thousands of similar agreements. It can sound confident and well structured. But it cannot assess whether that clause aligns with your risk tolerance, business objectives, or regulatory obligations. Those decisions remain human responsibilities. Generative AI excels at drafting, summarizing, organizing information, translating, and accelerating first passes. It cannot guarantee accuracy, reason from first principles, or validate its own output without external tools. Hallucinations occur because models optimize for likelihood, not correctness. From Models to Agents A basic generative model answers questions. An AI agent works toward goals. Agents combine a generative model with access to tools and systems, along with memory that tracks prior steps. Agents operate in loops. They perceive the current state, reason about what to do next, act by calling tools or executing tasks, observe the results, and repeat. This allows them to analyze documents, pull data from systems, generate drafts, update records, and coordinate multi-step workflows. Imagine an intake workflow where an agent reviews incoming requests, extracts key facts from attached documents, checks deadlines in a matter system, proposes task assignments, and drafts a response email. Each step relies on tools, memory, and structured decision-making layered on top of generative output. As systems become more complex, agents specialize. One may focus on research, another on planning, another on execution. An orchestration layer coordinates their work. This mirrors how real teams operate and enables scalable automation without relying on a single general-purpose assistant. This capability increases efficiency, but it also expands risk. Because agents can act across connected systems, teams must define permissions, escalation paths, and stopping points with care. Multimodal Systems Legal work rarely arrives as clean text. It comes as contracts, emails, screenshots, images, recordings, and video. Multimodal systems process all of these inputs together. Each modality is encoded into numerical representations and fused into a shared internal space. From that combined context, the system can generate outputs in any modality. A model might analyze a contract, inspection photos, and voice notes at once, then produce a unified summary or draft. Consider a compliance review that includes written policies, screenshots from internal systems, and recorded interviews. A multimodal system can ingest all of this material simultaneously and surface patterns that would otherwise require manual cross-referencing. This capability mirrors how humans integrate information and allows AI systems to operate across the messy realities of legal workflows. It also raises important questions about data handling, retention, and review that teams must address explicitly. What This Means for Legal Teams The objective is not technical mastery. It is precision. Legal teams must understand where AI adds leverage and where it introduces risk. Drafting and summarization benefit from generative systems. Final review, risk tolerance, and accountability remain human responsibilities. Agents require clear constraints because they can act across connected systems. Multimodal intake reshapes investigations, compliance, and contract review by consolidating fragmented inputs. A realistic pattern emerges: teams gain speed on early work, but outcomes improve only when they pair that speed with structured review. Without defined checkpoints, AI output drifts from organizational standards. Design workflows that assume human-in-the-loop review at every material decision point. Treat AI as an extension of professional judgment, not a substitute for it. Practical Guidance Ask focused questions when evaluating AI-enabled tools. Where does the system rely on probabilistic output? Which steps require mandatory human validation? What tools and systems can agents access? How is multimodal data handled and retained? Who owns outcomes after implementation? Pair technical evaluation with operational design. Assign responsibility for monitoring performance, updating workflows, and addressing failure modes. Measure outcomes, not just usage. Durable value does not come from demonstrations. It comes from workflow ownership, review discipline, clear accountability, and ongoing iteration. Conclusion Generative models create content. Agents plan and act. Multimodal systems unify messy inputs. Together, they define modern AI platforms. Understanding these mechanics sharpens use cases, clarifies guardrails, and improves risk conversations. Technology can surface issues faster and reduce operational drag. It cannot decide what matters. AI should extend judgment, not replace it. That balance defines effective legal work. Glossary Generative AI: AI systems that create new content by predicting likely continuations based on patterns learned from data. Large Language Model (LLM): A generative model trained on large volumes of text to produce human-like language output. Embeddings: Internal numerical representations that capture meaning and context so models can work with language, images, or audio. Attention: A mechanism that helps models determine which parts of input matter most at each step of generation. Hallucination: Confidently generated output that is incorrect or unsupported, caused by probabilistic generation rather than fact checking. AI Agent: Software that combines a generative model with goals, tools, and memory to perform multi-step tasks. Agent Orchestration: Coordination of multiple specialized agents so each handles part of a larger workflow. Multimodal Systems: AI systems that process text, documents, images, audio, and video together as one context. Human-in-the-Loop: Workflow design where people review, approve, or guide AI outputs at key decision points. Frequently Asked Questions Does generative AI replace legal judgment? No. These systems accelerate drafting and analysis, but decisions about risk, strategy, and accountability remain human responsibilities. Can AI-generated content be trusted as accurate? AI output should always be reviewed. Generative models optimize for plausibility, not correctness, which makes verification essential. What is the difference between a chatbot and an AI agent? A chatbot responds to prompts. An agent can pursue goals, call tools, remember prior steps, and operate across systems. Why do hallucinations happen? Models generate statistically likely responses rather than validated facts, especially when information is incomplete or ambiguous. How do multimodal systems change legal workflows? They allow documents, images, recordings, and other inputs to be analyzed together, reducing manual cross-referencing and speeding early-stage work. What controls should teams put in place? Define review checkpoints, limit agent permissions, assign workflow ownership, and measure outcomes rather than usage. Where should teams start? Begin with low-risk use cases such as summarization or intake, establish review standards, and expand deliberately as governance matures.

At the heart of everything I do is helping people. This isn't just a tagline or a LinkedIn bio flourish. It's a fundamental truth that shapes how I approach every interaction, every conversation, and every opportunity in the legal technology ecosystem. Yet, as I reflect on the countless conferences, summits, and industry gatherings that populate our calendars, I find myself questioning whether we're truly connecting or merely collecting business cards in an increasingly digital age. The Legal Tech Conference Challenge We've all been there: standing in a cavernous convention center, badge dangling from a lanyard, coffee in hand, surrounded by thousands of fellow legal professionals. The energy is palpable, the panels are insightful, the technology demonstrations are impressive. But something essential feels missing. How many times have you returned from a major conference with a stack of business cards, a bag full of vendor swag, and a nagging feeling that despite being surrounded by people for three days, you didn't form a single meaningful connection? The irony isn't lost on me. In an industry built on relationships and trust, we've created environments that often feel more like speed dating than relationship building. The Human First Perspective When I speak about legal technology, I always approach it from a human first perspective, framing discussions around user experience and meeting people where they are. Technology should enhance our ability to connect, not replace the fundamental human elements that make our profession meaningful. Yet, the conference circuit sometimes feels like it's optimized for efficiency over empathy, for quantity over quality. The best conversations I've had, the ones that led to genuine mentorship relationships, collaborative partnerships, and lasting friendships, rarely happened during the scheduled networking hour with 500 other attendees. They happened over a quiet dinner with three colleagues, during an impromptu coffee meeting between sessions, or through follow up conversations weeks after the event ended. The True Cost of Constant Travel Let's address the elephant in the room: the relentless conference schedule that many of us maintain. Weekly flights, hotel rooms that blur together, meals grabbed between sessions. This isn't sustainable, and more importantly, it's not conducive to the deep work and meaningful relationships that drive real innovation in legal tech. Every day spent in transit is a day not spent: Mentoring that law student who reached out seeking guidance Having thoughtful conversations with team members Writing and sharing insights that could help someone navigate their legal tech journey Building genuine relationships within our own communities I've learned that being selective about travel and conferences isn't about missing opportunities. It's about creating space for the connections that matter most. Redefining Connection in Our Industry So how do we move forward? How do we maintain the benefits of industry gatherings while fostering the genuine human connections that make our work meaningful? 1. Quality Over Quantity : Choose conferences that align with your values and where you can contribute meaningfully, rather than attending everything. 2. Create Intimate Spaces : Within large conferences, seek out or create smaller gatherings. Dinners, roundtables, walking meetings. These are where real conversations can happen. 3. Extend the Conversation : The most valuable connections often develop after the conference. Schedule follow up calls, continue discussions over email, and invest in relationships beyond the event. 4. Leverage Technology Thoughtfully : Use virtual meetings and digital tools to maintain connections without constant travel, but remember they supplement, not replace, human interaction. 5. Practice Presence : When you do attend events, be fully present. Put away the phone during conversations, listen actively, and engage authentically. As legal technology continues to evolve at breakneck speed, the importance of human connection becomes more critical, not less. The most sophisticated AI, the most elegant software solution, the most revolutionary platform. None of these matter without people who understand how to implement them thoughtfully, who can translate their benefits in human terms, and who care about the individuals they're meant to serve. My mission remains unchanged: to bridge the gap between the tech world and the legal world. But that bridge is built on human connections, one meaningful conversation at a time. Whether that conversation happens in a conference hall, over a video call, or during a mentoring session with a law student, what matters is that we approach it with intention, empathy, and a genuine desire to help. Because at the end of the day, legal tech isn't about the technology. It's about the people it empowers to work better, live fuller lives, and serve their clients more effectively. And that's a mission worth traveling for, selectively and purposefully, while never losing sight of the human connections that make it all worthwhile. Remember: The best investment you can make in your legal tech journey isn't in another conference ticket. It's in building genuine relationships with the people who will challenge, support, and inspire you along the way.

By Chad Atlas The pace of progress in foundational AI is nothing short of explosive, and legal technology is feeling the impact in real time.  In just the past month or so, Google’s latest Gemini models jumped to the top of intelligence leaderboards for the first time.  OpenAI released ChatGPT o3 and then o3-pro, their most advanced reasoning model yet to a broader range of users.  And Anthropic upgraded Claude from version 3.7 to 4.0.  Advances in core AI capabilities are directly reshaping what is possible in legal tech.  The speed of progress is now measured in weeks, not months. Chad Atlas This acceleration is not just hype.  Recently, prominent legal tech startup Harvey announced  it would integrate multiple foundation models from Anthropic and Google, explaining that “general foundation models have improved at baseline legal reasoning” so dramatically that optimization can now focus on task execution rather than baseline reasoning.  Their pivot reflects a broader industry reality:  the intelligence itself has become so capable that traditional engineering approaches are being rendered obsolete—now handled by the models themselves. For a sense of what this means, I recently tested one of these models on a complex antitrust law exam shared by a professor friend curious about the tech’s capabilities.  The model, as graded by the professor, earned at least a B+, possibly an A-, on a question that would challenge all but the most capable law students.  Three prompts, less than 15 minutes.  (My friend is not easily impressed; in this case, he was.) [1] This validates something I’ve theorized for several years as a CLO and startup advisor (and written about before ):  the most significant advances in legal AI come from improvements in the underlying models, not from specialized wrappers or specific legal adaptations.  Raw intelligence is what matters—and we are now seeing that play out in real time. Yet most lawyers and legal leaders evaluating tech investments have not realized how quickly this shift is happening.  Many are still focused on products, feature lists, and workflow demos, rather than the real driver:  the intelligence powering it all. Understanding the Architecture So, what actually powers today’s legal AI?  Virtually every tool you see can be modeled as a simple three-layer architecture: Layer 1: Intelligence.   The foundation model (ChatGPT, Claude, Gemini) that does the actual reasoning, analysis, and text generation. Layer 2: Engineering.  The plumbing, piping, and orchestration that make the AI useful for specific legal tasks,;retrieval systems that fetch documents, laws, and regulations; prompts that guide the AI’s behavior; workflows that chain together multiple steps to execute actual work; and various connections to legal databases or internal document systems. Layer 3: Application.  The user interface: a web app, a Word plugin, or whatever makes the system accessible to the end user. That’s it.  Intelligence at the bottom, plumbing in the middle, interface on top. The Bitter Lesson:  Why General Intelligence Wins Here is the crucial part:  that middle layer, the plumbing, has historically existed because the foundational intelligence had clear limits.  The AI was like a smart lawyer without any access to law books or a computer.  You could ask it questions, and it would respond from vague memories and experience, and that was that.  It often made things up or got things wrong.  Still, it was smart, so legal tech companies built elaborate engineering around these models to compensate.  Retrieval systems searched legal databases or documents and brought back snippets for the AI to review; workflow chains tried to mimic what lawyers do and allow the lonely AI lawyer to execute on projects. But this engineering was always a stopgap.  The retrieval systems were primitive and often returned irrelevant results, missed context that matters, or broke down when matters took an unexpected turn.  Every additional question to the AI added cost, so vendors cut corners to keep prices down and minimize the engineering and workflow complexity as much as possible. The demos looked slick, but the reality was fragile. This pattern is not unique to law.  AI researcher Rich Sutton’s famous essay The Bitter Lesson  described this trend precisely:  “The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin.”  Sutton observed that “AI researchers have often tried to build knowledge into their agents, this always helps in the short term, and is personally satisfying to the researcher, but in the long run it plateaus and even inhibits further progress, and breakthrough progress eventually arrives by an opposing approach based on scaling computation.”  We have seen this pattern play out in chess, image recognition, language translation and now in legal tech. What is happening now is that models like ChatGPT o3, Claude 4.0, and Google’s Gemini 2.5 Pro are swallowing the stack —meaning the models themselves are now handling much of the retrieval, workflow, and even some application-level reasoning that used to require hand-coded intervention. Leading legal AI companies have quietly pivoted to this reality, focusing less on custom engineering and more on harnessing the raw, ever-improving power of the latest models. The general approach wins, again. The DIY Question: When to Build vs. Buy This creates an interesting dilemma for legal teams:  How much should you rely on pre-built solutions versus working directly with the intelligence yourself? My father was a perfectionist handyman who once spent many weekends meticulously working on the wood framing around our windows rather than hiring a contractor.  The results were superior, but the time investment was enormous—and one could argue that perfect became the enemy of good.  Different people have different value preferences, different skill levels, and different tolerance for complexity. The same dynamic applies to legal AI.  If you are a lawyer who can find relevant cases and statutes, extract key facts, and clearly structure problems, you are already better at “wrapping” the intelligence than most hard-coded software solutions.  Working directly with advanced models, e.g. prompting, iterating, and fact-checking, often delivers results faster, more transparently, and at lower cost than pricey legal tech platforms. But not everyone wants to be their own handyman.  There is legitimate value in platforms that handle infrastructure, route tasks to the best models, and provide peace of mind for privacy or compliance.  For a busy legal team, a trustworthy solution can mean less IT overhead, better integration, and more time focused on what matters. The key is understanding what you are paying for. Why Lawyering Is Still Different Law remains different from other domains where AI has achieved breakthrough performance. Lawyers are not just looking for productivity—they need transparency, control, and the ability to interrogate every step.  Most lawyers I know, when handed a template, playbook, or automated workflow, want to see exactly what it is, how it works, and adapt it to their context.  Accepting someone else’s black box—especially in high-stakes matters—rarely feels right.  (Indeed, we have an ethical duty to ensure we reasonably understand the issues and applicable law.) This is why working directly with the most capable models may be optimal for many legal workflows.  You retain control, can synthesize the best inputs, and understand the reasoning behind outputs.  You can see the model’s work, challenge its conclusions, and iterate until you are satisfied.  The interaction becomes a kind of productive intellectual dialogue rather than a passive consumption of pre-packaged results. If you do adopt a legal tech platform or workflow solution, make sure you know what is inside the box:  whose judgment are you trusting?  Whose templates, whose playbooks, whose risk tolerances?  The transparency question is not just about understanding the technology—it is about understanding whose legal judgment is baked into the system. Proof in Practice:  Where Wrappers Matter—and Where They Don’t Some wrappers and orchestration layers genuinely add value, especially as intelligence gets cheaper and more accessible.  Software engineers have flocked to tools like Cursor, which offers a code editor built around AI.  It isn’t just about the underlying model’s intelligence—Cursor’s workflow, search, and integration features make it easier for users to harness that intelligence effectively.  OpenAI’s recent acquisition of such a development tool, Windsurf, for $3 billion , suggests that intelligence wrappers have real value.  These “application layers” matter when they truly enable new kinds of productivity and collaboration. The same principle applies in legal.  For example, Harvey’s shift toward task execution rather than baseline reasoning represents an intelligent adaptation to the new reality.  There is real value in systems that understand legal context, maintain proper citation formats, integrate with existing workflows, and handle the mundane but critical details that practicing lawyers need.  (Bureaucracy solutions in a box; yes, please.) But the bar for genuine value is getting higher.  Templates, basic prompt libraries, and simple workflow automation have limited long-term differentiation when the underlying intelligence can handle these tasks directly. The Real Evaluation Question When evaluating a legal AI solution, ask yourself:  what am I really paying for? Is there genuine value-add I cannot replicate?  Or am I paying for packaging around the same core intelligence I could access directly? Consider these criteria: ●      Integration complexity:   Does the solution handle genuinely difficult technical integration, or could you achieve similar results with direct access? ●      Legal domain expertise:   Are the prompts, workflows, and guardrails meaningfully better than what you could develop yourself? ●      Transparency and control:   Can you understand and modify the system’s behavior, or are you locked into someone else’s judgment? ●      Cost structure:   Are you paying a reasonable premium for convenience, or multiples above the underlying intelligence cost? The question isn’t whether wrappers are good or bad—it is whether you understand where the value comes from, what tradeoffs you are making, and how much transparency or control you're willing to give up for convenience. Looking Forward Rich Sutton’s “bitter lesson” from AI research is clear:  as intelligence gets stronger and cheaper, custom engineering layers get swallowed up.  But that does not mean all wrappers disappear only that the bar for real value is getting higher.  Legal tech companies that recognize this reality, like Harvey pivoting to task execution, are positioning themselves to add genuine value rather than just repackaging commodity intelligence. For legal teams, this means being more sophisticated about what you are buying.  The most successful legal departments will likely combine direct use of frontier models for complex reasoning with specialized tools for specific workflows where the wrapper genuinely adds value. Intelligence is all you need, provided you know how to wield it and critically assess the value of everything built on top. About the Author Chad Atlas is Chief Legal & Ethics Officer at an AI-first fintech startup and advisor to early- to late-stage biotech companies.  He has over two decades of legal experience spanning federal clerkships, BigLaw litigation, and executive leadership roles at a clinical-stage biotech company.  His philosophy and computer science background from Duke initially fueled his interest in the intersection of law and emerging technology.  He recently launched No Vehicles in the Park , where he writes about legal AI, professional judgment, and the evolving legal landscape. [1] Since submitting this article for publication, professors at the University of Maryland Law School released a paper stating that o3 (the same model I used) got three A+s, one A-, two B+s, and a B on exams they tested it on.  Link to the paper here .

Over the past two years, there has been a significant interest in adopting generative AI tools in the legal industry. However, many organizations continue to face challenges in understanding how to procure, implement, and maximize the potential of these tools. This article seeks to address this gap by offering key insights and questions to consider when assessing GenAI solutions. Key areas covered include: Foundational Models : Understanding the underlying large language models (LLMs) used by providers is essential. Different models have varying strengths and weaknesses, and continuous evaluation is necessary to ensure optimal performance. Data Security : Ensuring that providers have robust security processes and certifications is vital. Questions about data storage, access controls, and third-party service providers should be addressed to maintain data integrity and security. Accuracy and Reliability : Evaluating the accuracy of GenAI tools is critical. Providers should provide benchmarks, case studies, and details on how they handle model drift and degradation. High accuracy translates to higher reliability and consistent performance. By addressing these key areas, organizations can make informed decisions and successfully integrate GenAI tools into their operation. When selecting a GenAI solution, the first step is having a clearly defined use case. AI models vary in their capabilities, strengths, and weaknesses, so understanding what you need the AI to accomplish ensures that you evaluate solutions effectively and choose one that aligns with your business goals. Foundational Models. The foundational model of a Large Language Model (LLM) is critical to get right because it serves as the core engine that determines the capabilities, limitations, and overall effectiveness of a generative AI solution. Below are listed some key questions you should ask GenAI legal tech providers when evaluating potential options. Check if you’ll be locked in by the provider’s choice of model. This restricts your flexibility to swap between different foundational models should a better performing one or one that matches your needs more closely emerges. What foundational model does the provider use? Understand the underlying large language model(s) that provides insights into the tool’s capabilities and potential limitations. Different foundational models such as OpenAI’s GPT, Google’s Gemini, Meta’s Llama, and Anthropic’s Claude vary in terms of architecture, training data, and optimization strategies. The choice of model impacts accuracy, fluency, bias, contextual awareness, and multimodal capabilities (e.g., handling text, images, and code). How do they ensure they are using the best performing LLM? The landscape of LLMs is rapidly evolving. Providers should have a robust process for continuous model evaluation to ensure they are leveraging the most effective model available. A strong evaluation framework should incorporate standardized performance benchmarks like MMLU (Massive Multitask Language Understanding), SuperGLUE, and HELM (Holistic Evaluation of Language Models) to measure the model’s accuracy, reasoning ability, and bias levels. Additionally, providers should conduct domain-specific testing if the AI is being used in industries like healthcare, finance or legal applications, ensuring the model meets the necessary precision and any relevant compliance standards. How often are the models updated and retrained? Regular updates and retraining are crucial for maintaining a LLM that remains accurate, relevant, and aligned with evolving knowledge base and end user needs. Models can quickly become outdated as new facts, regulations, and industry trends emerge, making it essential for providers to have a structured retraining and updating cycle. Buyers should inquire about the frequency and methodology of these updates to ensure the model is continuously improving. Some providers update their models on a fixed schedule, such as quarterly or annually, while others use a rolling update approach, where models are incrementally retrained with new data as it becomes available. Can you inject your own data into the pre-existing LLM to fine-tune the results? Customization may be necessary to align the LLM with specific business needs, industry requirements, or proprietary knowledge. The ability to fine-tune a pre-existing LLM using your own data can significantly enhance its relevance, accuracy, and effectiveness for specialized applications. Organizations should assess whether the provider supports fine-tuning, embedding domain-specific knowledge, or integrating external databases to tailor responses. Data security. While the GenAI space is moving quickly, providers need to ensure that they're keeping your data secure at all times and they have the correct protocols in place to deal with any potential breaches. As part of your evaluation, you'll need to be satisfied that the provider has the right security processes and certifications in place. Depending on your use case and the solution being considered, the key questions to ask the provider are listed below. Does the provider have security certifications?Security certifications are a crucial indicator of an AI provider’s commitment to data protection, data privacy, and compliance with industry standards. Buyers should look for recognized security frameworks such as ISO 27001, which ensures a robust information security management system, or SOC 2 (Service Organization Control 2), which evaluates how well a provider safeguards customer data in terms of security, availability, processing integrity, confidentiality, and privacy. Some startups might not have the right certifications in place. In that case, request penetration test results and ask how often testing, both internal and external, is carried out. Where will your data be stored or hosted?Understanding where your data is stored and processed is critical for ensuring compliance with data residency, security, and regulatory requirements. Buyers should verify whether the AI provider offers flexible hosting options, such as on-premise deployment, private cloud, hybrid cloud, or specific regional data centers, to align with their internal policies and legal obligations. What are the access controls and authentication options?Robust access controls and authentication mechanisms are essential to ensure that only authorized personnel can interact with AI systems, particularly when dealing with sensitive data, proprietary knowledge, or regulated industries. Buyers should evaluate whether the provider offers Role-Based Access Control (RBAC), which allows administrators to restrict access based on job function, seniority, department, or geographic location. For example, executives may have full system access, while frontline employees may have read-only permissions, and IT administrators may have advanced configuration rights.Does the provider rely on third party service providers to deliver their service?In most cases, AI providers rely on third-party service providers for various aspects of their infrastructure, including cloud hosting, data storage, API integrations, and security. It’s important to understand who these third parties are, what role they play, and how they handle your data to ensure compliance with security and privacy requirements. Additionally, businesses should clarify if any subcontractors have access to sensitive or proprietary information and what measures are in place to prevent data misuse. Accuracy and Reliability When evaluating generative ai tools, understanding the accuracy of the model is crucial. The quality of the output is directly dependant on the accuracy of the model. High accuracy translates to higher reliability. Reliability means the solution consistently provides accurate and dependable results across various scenarios and over time. What metrics do you use to measure the accuracy of your models?When evaluating an AI provider, it’s crucial to understand how they measure model accuracy and which metrics they prioritize in relation to your specific use case. Common benchmarks include Perplexity (PPL) for predictive accuracy, BLEU and ROUGE for translation and summarization, Exact Match (EM) and F1 Score for classification and retrieval tasks, and TruthfulQA/FEVER for factual accuracy. Note that most benchmarks have some limitations. Ask about false positive rates and whether accuracy can be fine-tuned for industry-specific needs. Additionally, assess if and how the provider monitors real-world performance through human-in-the-loop validation, A/B testing, and live feedback loops to ensure ongoing improvements.   What processes are in place to monitor and maintain the model's accuracy over time? Over time, LLMs can experience model drift and degradation, where their responses become less accurate, biased, or misaligned with current data trends. This happens because language evolves, facts change, and business needs shift. To ensure long-term reliability, ask the provider what monitoring and maintenance strategies they use to track, evaluate, and update the model’s performance. Without proper monitoring and maintenance, AI models can become outdated and unreliable. Provider that implements proactive tracking, continuous fine-tuning, and conduct real-world performance evaluations ensure that the model remains accurate, unbiased, and aligned with evolving business needs. Can the vendor provide details on the performance of their solution in real-world scenarios?Evaluating an AI provider based on real-world performance is essential to understanding how their solution functions beyond controlled environments and benchmark tests. Ask the provider for case studies, references, and deployment examples that demonstrate how their solution performs in organizations of similar size, industry, and complexity as yours. How do you evaluate the solution’s performance on new data?This question suggests an educated buyer who is thinking beyond their current use case and where and how to deploy the solution more widely. For businesses looking to scale adoption across multiple use cases, the solution must seamlessly handle evolving datasets without frequent or laborious manual intervention. Providers with robust evaluation strategies, automated monitoring, and lightweight adaptation options ensure that the AI remains accurate, adaptable, and future proof, reducing the need for constant retraining while continually maintaining high performance. Generative AI tools offer immense potential for organizations ready to harness their power. By clearly defining use cases, understanding foundational models, ensuring robust data security, and evaluating accuracy and reliability, businesses can make smart, informed decisions. Staying proactive and informed will be key to leveraging these advanced technologies effectively and avoiding the dreaded Shiny New Toy Syndrome. Sharan Kaur – Go-To-Market (GTM) Expert | Legal Tech Strategist | Growth Leader Sharan Kaur is a seasoned growth and sales leader with a proven track record of designing and executing global go-to-market (GTM) strategies for startups, scaleups, and legal tech providers. With a background as a corporate litigation lawyer and an Executive MBA, Sharan specializes in driving revenue growth, leading high-performance teams, and implementing scalable solutions for long-term success. Her expertise lies in managing full sales cycles, building strategic partnerships, and consulting post-deployment to ensure maximum value realization. Sharan works closely with law firms, corporate legal teams, and legal tech providers to deliver custom solutions, optimize workflows, and enhance user adoption of innovative technologies. Currently, as a Digital Transformation Consultant at Legal Solutions Consulting, Sharan bridges the gap between legal teams and generative AI solutions, ensuring seamless adoption and long-term value realization. Her cross-functional leadership experience and deep understanding of legal technology adoption make her a trusted advisor for businesses seeking sustainable growth and operational excellence.

We are now at the stage of the AI revolution where even those of us living under a proverbial rock have heard of generative AI. The idea that a particular release version of a large language model could be a major global media event was a laughable idea until very recently, but AI has moved from research labs to our daily lives with remarkable speed. In the legal space, we are constantly bombarded with marketing campaigns about productivity gains and improvements in the quality of our work. The promises vary, but include claims to supercharge this or that, to make a particular task instant or effortless, or to deliver game-changing efficiency gains across the board. This may all sound very odd coming from a person who has just joined a generative AI startup, but please bear with me! I was called to the Bar almost twenty years ago. Ten years ago, I drifted away from private practice and, after the traditional mid-life crisis, towards technology. Like many in the space, tech represents the intersection of my deep professional skillset and a slight tendency towards neo-mania, with a twist of gadgetry obsession. I have lost count of the number of shiny gadgets that I have loved (and then consigned to landfill, sorry Mother Earth) over the years. While the urge to buy soon-to-be obsolescent electronic junk has faded as I age, my love of legal tech has only spiraled, and—to be frank—is now beyond all semblance of control. [1] I first thought that legal tech would also have the beneficial side effect of allowing me to do less work and make more money, but since I drifted away from the wig and gown, I have been disabused of the rather naïve notion that this is a simple matter. [2] Along my journey, I have been fortunate enough to have been involved in global transformation projects alongside some of the biggest legal teams in the world. While I am not yet as grizzled as some of the towering figures in the ops and transformation space, I am at least tipping into the category of those who have their share of war stories, and when Colin asked me to share what he very generously described as “insights” or “thoughts”, I jumped at the chance. Thank you, Colin. Sidenote: if you want to see what generosity looks like, follow Colin on LinkedIn, or better yet, catch him at an event. So, after a fabulously verbose introduction, the question we’ve all been asking: what on earth is going on, and what on earth can I actually DO? What Does AI Eat For Breakfast? We’ve all heard the phrase “culture eats strategy for breakfast”. Sometimes it’s said by someone looking for an excuse for lackluster planning, sometimes by a charismatic narcissist looking to paper over the behavioral cracks in their organization through which toxicity is seeping and oozing, and much more rarely, it’s said by a true leader. [3] But if culture eats strategy for breakfast, what does AI eat? In my opinion, there is a non-zero chance that AI will eat culture. The technology is just so incredible that I believe that there is a non-zero chance that—absent a major or indeed global EU-style interventionist push—our existing culture will be eradicated. I mean that in the broadest possible, non-corporate sense. This is not the cloud revolution, changing enterprise SaaS purchasing habits, and creating a new segment in an existing market. This is something else entirely. It will impact every single facet of life, and has the potential to completely up-end the assumptions on which we have built our professional world. To avoid worsening my chronic lack of brevity, I will deliberately leave this wide-ranging discussion, the possibility of AGI, the potential for an AI singularity, and other such topics for another time, potentially when I have a drink in hand. My focus here, then, will be on legal culture exclusively. But I’m supposed to be talking about AI strategy , so why am I starting with culture ? The extent to which we are merely products of our environment is debatable, but AI strategy considerations—and indeed any  strategy considerations—should in my view start with an analysis of culture . Corporate culture is a vast topic, which I am probably under-qualified to discuss. Legal culture is a little closer to my wheelhouse, but even that I would prefer to leave to the experts. Check out Charlotte Smith and her writing, for example. However, I have been asked to address the issue of AI in legal teams on sufficiently numerous occasions that I will accept if not the label of “expert” then at least the label of “non-moron”. Lawyers are not famously convivial or congenial. They are a very more diverse group than stereotypes allow, but qualifying as a lawyer is to the right-hand side of the bell curve in terms of difficulty in most jurisdictions, and social status is similarly located. Type-A personalities abound, and parts of Big Law culture can be grind-centered. I don’t want to toss the baby out with the bathwater, as I have a deep love for the profession, and I am very proud of the time I spent prosecuting. I should also note that I am no less proud of the time I spent representing wealthy corporations and nation states in international arbitrations. The law has been my passion for a long time, and I like to think I can make a case for any aspect of it. Except trusts and estates. You folks are just straight up weird. So, to return to our generalizations, we have a bunch of smart, driven, socially and professionally conservative knowledge professionals, faced with a novel technology that has a substantial chance of being better than all humans at reasoning and arguing on a timeline measured in months or maybe years, not decades. What should we expect? Fear. Lots of fear. When I show teams what AI can do, I make a joke that it takes thirty minutes to configure and deploy the instance, thirty days for associates to get up to speed, and then three years of therapy for the partners to get over it. They’ve just watched AI chew through ten thousand pages of loan agreements in ten minutes, smashing out verifiable data, and building an excellent first draft due diligence report. What do they see? Oftentimes, they just see billables evaporating. Getting to the top is hard. Really hard. When you get there, it’s your turn to get rich, frankly. You’ve earned your time in the sun, it’s your turn to hold a bucket underneath the money faucet, and you probably don’t want to hear that “ everything’s different now, Jim ”. When their turn comes, some people tip into more extreme immobilism, while some just become a shade more conservative. It takes a very special and courageous person indeed to see something new, recognize the potential, and immediately set to work on cannibalizing the business they just spent thirty years building. They are putting down the bucket and going to look for a way to tap the water main. There are too many to list, but early adopters of legal AI who are spending their very hard-earned cash are my professional heroes — even if they spend it with my competition!   Strategy Building I’ve said this phrase, or a variation thereof, to rooms full of GCs or law firm partners more times than I can count: “ Who you are  determines where  you are. ” There’s a bit of elaboration, but that’s the punchline. Going on vacation? Which destination? Why? What do you want to do? What do you prefer? Who do you like to hang out with? Noisy bars or quiet cafés? I got a friend who likes horse-trekking vacations in Mongolia. Another who wants only goes to party destinations. Personally, I want a resort with nice weather and good food, a moderate distance from my house, fun stuff for the kids a must. We are different people and that determines where  we end up. Assuming you’re the leader, there’s hopefully a fit between who you are as a person, your values, and those of the organization you serve. If not, you probably don’t need to be told to dust off the CV and move as soon as you can. In the legal industry, this is the “type” of team you are. White-shoe firm with impeccable credentials and terrifying rates? Aggressive personal injury team with massive billboards? Dedicated local courthouse solo-warrior who hung her shingle in ’76 and never looked back? In-house teams have the same question. Are you building for speed? Industrializing standard paper in a B2C industry with a website click-wrap agreement? Negotiating three agreements a year in a complex geopolitical context? Struggling to maintain alignment in your database of one million product codes? Somewhere in between? You can’t hope to nail AI strategy if you haven’t got this straight. If you can point to a written and maintained version of your company’s values, a written legal team mission statement, and you can see tangible evidence of them in your daily professional life, that’s a good sign. If you need help with this, I’ll declare a conflict of interest and recommend that you call Emilie Calame, my former boss (and long-suffering wife). Once you’ve figured out who you are, and you’re aligned with the organization that you serve, you need a destination. In-house teams may well have some clearly communicated corporate targets, OKRs and so on. Private practice targets tend to be equally clear and mainly financial. Bring that down to a more granular level and build a team target. Then, take a look around. What’s your team like? How big is it? Average age? Willingness to experience discomfort? Adaptability to change? Moving a fifty-lawyer team isn’t the same ask as moving a three-lawyer team. Second part of the stock take: what process and tech do we already have? Microsoft org? Check for Power Automate licenses. Notion teams can build some pretty slick stuff with middleware (Zapier, Make, Airtable, Bubble, etc.). Values, alignment, team status, destination, available resources for the journey. Now comes the fun bit. What do you need to do, to get where you want to go? For some people, AI is a solution looking for a problem. Don’t be that person. If your problems include handling large volumes of unstructured data, extraction, analysis, markups, contract negotiation, document comparisons and so on, then there is undoubtedly a game-changing level of leverage an LLM away. Step one, then: identify your current tasks. What are you doing on a regular basis? What do those tasks look, in detail? How many person-hours a month, which colleagues handle them? Fair warning: this is much, much harder than it sounds. Once you have a clear view of processes, run the “Five Whys” and dig into the underlying motivation. See if there’s something you can ditch, or handle in another way, or just optimize out of existence. Example: contract negotiations. Revisit your standard paper, do some deal post-mortems, and if you’re drowning in red ink on each deal make sure you know why, and then try to eliminate the problem. Eliminate the task if possible, optimize the survivors, then try to use technology to automate all of the optimized flows you have left. In that order. Step two: among the tasks you’ve identified as rock-solid must-haves in your work life across the next year or so, which of them involve large volumes of data, repetitive work, similar documents, and other such “AI-tells”? A good rule of thumb is “AI is not for everything”. Again, this might seem strange from an AI-evangelizing professional from an AI company who has tied his career to the future of this technology, but I hope that you can see that is something to take seriously. If your organization needs to push tens of thousands of SKUs into order forms from a CRM, and then run this data into an ERP system, performing complex math along the way, you probably want some integrations that pipe the data end-to-end, and not the probabilistic miracle machine that is an LLM. If, on the other hand, you have an increasing volume of work that requires you work with large volumes of unstructured data, analyzing, extracting and generating written work, your life is going to change very quickly.  Step three: preparation. Let’s assume that you’ve identified a process that is unavoidable, optimized, and not automatable. In an in-house team, this might be contract negotiation, IP licence audit, or an employment contract audit. In a law firm this might be anything from an an M&A review, first pass reviews of written submissions from the other side, reviewing docs from a client, or even client onboarding. Figure out how much this process is costing you. Anything under 10k a year is very unlikely to be on your radar. Anything under 100k might not be worth it. Touching business critical processes means risk, and it means decision-maker time. Price the risk as best you can, price the time it takes for your leaders to provide input, and then add the cost of change. Lost productivity, the valley of despair, and the cost of the tool itself (plus assistance with the change) will all add up very quickly. Get this enormous number to the front of your mind and add a safety margin. Did you include project planning and general bandwidth? What about the extra business this shiny new toy might bring in? How safe are we feeling about these numbers? Step four: project launch. We have identified something that AI can do, the fixing of which will generate not only a positive ROI, but one so large that once all other factors are taken into consideration, it meets the IRR criteria required by whichever body makes such decisions. The decision is made, your organization commits to action, and the fun begins! Scoping, RFIs, RFPs, POCs, pilot phases, roll outs and more. If you’ve got this far, building a culture of continual improvement shouldn’t be too taxing, so let’s say Step Four rolls into eternity… Business As Usual? These basics should give you a framework that allows you to identify opportunities in a given team at a given time, but it’s very much a view from the trenches. In my opinion, this kind of work can only deliver exceptional results when it is performed by someone with an excellent grasp of the larger strategic dynamics at play in the market. What’s the 30,000 foot view? The nature of innovation is that it is not at all “business as usual”. It is a technology that changes things to such an extent that entire sections of the economy will die out. The whale oil industry, horse-drawn cabs, coal mining, human translators. They were all replaced in large part by machines. Are lawyers next? No. Caveat: the market for legal services baffles me. I see constant rate rises, above inflation increases across the board, almost constant complaining from in-house teams, ever-increasing numbers of lawyers qualifying to practice, and ever-increasing volumes of work being performed. Supply and demand don’t seem to be particularly well-correlated, in other words. To make matters better/worse depending on whether you’re buying or selling, buyers appear to be almost entirely insensitive to price. Two-and-a-half predictions then. The half prediction is that given the economic incentives and social status involved, I don’t see the pull of the profession diminishing soon. That will help drive the first “real” prediction. Lawyers won’t disappear. More lawyers, better tooling, bigger addressable market. Jevon’s Paradox will play out in full, and as efficiency rises, overall revenues that flow to the sector will increase, not decrease. Second “real” prediction. In no team does the advent of AI means business as usual. It is orders of magnitude better at very particular tasks, and this comfort zone is expanding rapidly. I believe that in some industries, a massive proportion of tasks are already well within the AI comfort zone. I include the legal industry here. Over time, what we traditionally considered to tasks reserved for human experts will be taken by AI. Economics would suggest that the higher the proportion of such tasks in an industry, and the higher their cost to the wider economy, the more capital will rush to address this. My economic ignorance aside, didn’t a legal AI team out of the US just raise 300 million bucks? Leaders in such industries have less time than anyone else to react. In fact, reacting is not enough. They must anticipate. This involves risk, and fortunately for us lawyers, we excel at risk analysis. What does this mean in very real terms? Get your hands dirty. Understand what an LLM is, what it does, what it can’t do. Understand what training data is, and what fine-tuning can and can’t do for model performance. Test things. Get an idea of what compromises might be made in the deployment of a system. Speed? Security? User-friendly interface? Make smart supply chain decisions. Is your vendor model agnostic? US-based? What about open source? What about self-hosting? The answers to these questions vary across time and across industries. I don’t have the answers for you, but just asking the questions will put you to the right-hand side of the bell curve when it comes to AI strategy. Good luck.   P.S. — as a bonus, here is a list of what my priorities would be as a GC or law firm partner: GCs: 1.        Immediately draft AI use policy for employees 2.        Draft an AI supplier policy, particular eye on data reuse 3.        Go for low-hanging fruit (unstructured contract data, most likely) 4.        Re-invest every minute saved in more AI work. Goal: a lean department that does nothing but strategize and handle BAU outliers.    Law Firm Partners: 1.        AI vulnerability assessment: which parts of my practice are AI-tractable, and which are billed on a time basis? Which part of the market am I in? Mass-market, leave ASAP. You are fungible. Pick a specialism and differentiate if possible. Mid-market? AI price competition will be painful. Find a segment and see if you can build a super-reliable money-spinning use case where you can secure first-mover advantage. Use the extra money to try to keep your lead/move up-market. Elite? Secure the leading specialists. Knowledge capitalization is coming fast. If your model is sufficiently agile, fight harder for fewer elite hires. 2.        Invest in AI, reinvest each hour saved in improving the things clients love most: accessibility, face time, updates, those spontaneous meetings or lunches that overrun but where you figure out really important stuff together. 3.        In parallel, consider pushing the parts of my work that are AI-tractable towards fixed fees. Goal: a firm with impeccable credentials and brand, top-tier work, “cash-cow” AI-powered workstreams, massive margins, unrivalled client care, strong BD and growth.  Jonathan Williams is a recovering litigator and arbitration practitioner. He began his career as a prosecutor before switching to international arbitration and moving to Paris. A decade ago, he slid across into technology, where he has developed a deep expertise in solving the biggest problems for the biggest legal teams. Innovation, technology, change management, strategic advisory work and generative AI have taken up a substantial portion of his professional life. The last five years he has spent working for Calame, the advisory team founded by his wife, Emilie Calame. He joined Legora at the end of 2024, where he is heading up their operations in France. He is based in Paris but travels extensively. [1]  By the way, that’s a deliberate em-dash, not an OpenAI one. I’m probably not using them correctly, but it is  me that’s using them. [2]  As you may be able to tell, I have not, however, been disabused of the notion that sentence length or the excessive use of subordinate clauses is a proxy for intelligence. See also, excessive use of footnotes. Lawyers love footnotes. [3] The ratio I’ve encountered in my career thus far is approximately 70:29:1. Better than hearing a leader say  it, is watching a leader embody it. I’ve been lucky enough to see this at Calame and with my current employer.

In an upcoming article , we consider how much downward pressure GenAI will exert on the use of billable hours as a proxy for value. If technology can handle adeptly the more mundane tasks in a lawyer’s day, why will clients be inclined to pay humans to take more time to do the same work? And if clients refuse to pay for humans to handle routine work, what will that mean to a law firm’s bottom line? We’ve heard all of the fears about GenAI’s use in the practice of law.  Some of those fears are important to resolve—such as confidentiality, bias, reliability—and some are less so, as one of us has pointed out . Sure, even the best technology is error-prone, but so are humans. [1] And computers, unlike humans, don’t get bored, tired, or inconsistent when performing repetitive tasks. Imagine a world in which a law firm has figured out a way to use GenAI to do simple tasks quickly and well. If ChatGPT can pass a bar exam , then any well-designed GenAI program can prepare a credible first draft of a pro hac vice  motion; it can analyze a contract ; it can draft a brief . [2] Even way back in 2018, LawGeex demonstrated the superior work of  its technology as compared to humans, in issue-spotting clauses in NDAs .  And GenAI keeps improving at a breathtaking pace. So consider different strategic paths taken by two hypothetical firms, imaginatively named Law Firm A and Law Firm B.    Law Firm A has decided that there are certain tasks that its lawyers do that can be given to GenAI first, with the lawyers then reviewing the resulting draft. The good news is that these lawyers can now be deployed to do tasks that GenAI can’t  do—their time has been freed up for that more interesting work. The bad news is that the work that used to be billed out by the hour is completed in seconds through automation. Isn’t that bad for the law firm?   We don’t think so. Now Law Firm A has two types of income streams: a commoditized income stream generated first by GenAI and then revised by humans, and a more bespoke income stream for things that only lawyers can do. (In our article, we suggest that the bespoke work, if it’s billed by the hour, can now command a higher billable rate than before, because experts are focusing on the tricky, novel issues.) With its freed-up time, in addition to working on bespoke matters, Law Firm A can find ways to bring junior lawyers up to speed the way that both of us were trained: by watching and learning from more senior lawyers in real time. Its commoditized work will likely be monitored by senior associates, with bespoke work done by both senior associates and by partners. In a world in which clients don’t want to pay for first- and second-year lawyers to be trained “on their dime,” the firm can now afford to devote more in-depth mentoring to keep itself sustainable.   What about Law Firm B, which eschews GenAI and insists that all but the most mundane work has to be done by humans, for quality control reasons? We think that Law Firm B runs the risk of becoming obsolete. When a Law Firm B client can do a first draft of something internally with or without the assistance of GenAI, why would it want to pay for Law Firm B’s junior associates to take time doing the same first draft? Law Firm B may find itself losing clients to Law Firm A, which is handling client matters more efficiently. It may also find itself losing associates to Law Firm A.   Our article contemplates a world in which the pyramid model, built on the premise of many junior lawyers doing billable work, may disappear in light of a more efficiently shaped economic model. Maybe that model is more of a cylinder, streamlined to use fewer lawyers in total because some of its junior ranks have been replaced with GenAI. Maybe the model morphs into a diamond, with more senior associates and fewer partners and junior associates. Or maybe the pyramid becomes a starfish, with a core of central support and different “arms” using GenAI either more or less, depending on the type of practice. There are many possible shapes, but we believe that the pyramid, as we have known it for decades, will be the least sustainable for most practices. Ultimately, law firms will have to grapple with the idea that the billable hour is not value but just a mere proxy for value. We believe that the firms that find a better way to capture the value-add of humans to drafts initially produced by GenAI will be the firms that survive and thrive. [1] This point is where the other one of us wants to refer you to the speech in Top Gun: Maverick about drones taking over for test pilots . [2] We’re just citing to some of the great programs out there. There are many such great programs. Nancy B. Rapoport is a UNLV Distinguished Professor, the Garman Turner Gordon Professor of Law at the William S. Boyd School of Law, University of Nevada, Las Vegas, and an Affiliate Professor of Business Law and Ethics in the Lee Business School at UNLV.  After receiving her B.A., summa cum laude, from Rice University in 1982 and her J.D. from Stanford Law School in 1985, she clerked for the Honorable Joseph T. Sneed III on the United States Court of Appeals for the Ninth Circuit and then practiced law (primarily bankruptcy law) with Morrison & Foerster in San Francisco from 1986-1991.  She started her academic career at The Ohio State University College of Law, served in three deanships, one stint as Acting Provost (UNLV), one stint as Acting CFO (also UNLV), and one stint as Special Counsel to the President of UNLV.  In 2022, UNLV’s Alumni Association named her the Outstanding Faculty Member of the Year.   Boyd law students have honored her three times:  she tied (with Professor Jean Sternlight) for “Faculty Member of the Year” in 2024; she was named “Faculty Member of the Year” (and faculty commencement speaker) in 2021; and she was named “Dean of the Year” by Boyd law students in 2013.   Her specialties are bankruptcy ethics, ethics in governance, law firm behavior, artificial intelligence and the law, and the depiction of lawyers in popular culture.  She has served as the Secretary of the Board of Directors of the National Museum of Organized Crime and Law Enforcement (the Mob Museum) and currently serves as a Trustee of Claremont Graduate University and the Chair of its Audit and Risk Management Committee.  She is also a Fellow of the American Bar Foundation and a Fellow of the American College of Bankruptcy.  In 2017, she received the Commercial Law League of America’s Lawrence P. King Award for Excellence in Bankruptcy, and in 2018, she was one of the recipients of the NAACP Legacy Builder Awards (Las Vegas Branch #1111).  She has served as the fee examiner or as chair of the fee review committee in such large bankruptcy cases as Zetta Jet, Toys R Us, Caesars, Station Casinos, Pilgrim’s Pride, and Mirant.  She is serving as the President of UNLV’s Chapter 100 of Phi Kappa Phi from 2024-2025. Joseph R. Tiano Jr., Esq. is Founder and Chief Executive Officer at Legal Decoder. After practicing law for nearly 20 years, Joe founded Legal Decoder because he saw that clients lacked the analytic tools and data to effectively price and manage the cost of legal services delivered by outside counsel. Joe set out to build an intelligent, data driven technology company that would revolutionize the way that legal services from outside counsel are priced and economically evaluated. Legal Decoder’s data analytics technology is used in law firms of all sizes from AmLaw 50 law firms to boutique firms; Fortune 500 legal departments and in major Chapter 11 bankruptcy cases (PG&E, Purdue Pharma, Toys R Us and others). Joe is a prolific author having (co-)authored nine law review articles published in scholarly journals. In addition, he has written articles for countless blogs and other online media on substantive legal issues and the legal industry in general. He regularly presents at CLEs and other seminars and courses on topics ranging from artificial intelligence, LegalTech, legal data analytics to legal ethics and legal malpractice. He is also an Adjunct Professor of Law at the Arizona State University Sandra Day O'Connor Law School. Previously, Joe was a Partner at Pillsbury Winthrop Shaw Pittman, LLP and Thelen LLP where he grew and managed all aspects of a multi-million-dollar cross-border finance practice. Entrepreneurship runs through Joe’s veins since his early days as a venture capital lawyer representing transformative technology companies, like Blackboard Inc., and many of the outgrowths of Blackboard (WeddingWire/The Knot, Presidium, Starfish Retention Solutions and others). Joe graduated from Georgetown University in 1992 with a Bachelor of Science Degree in Business Administration and received his J.D. from the University of Pittsburgh School of Law in 1995. Joe is a native Washingtonian who currently lives in Scottsdale, Arizona with his wife, Meredith, and their two boys, Gabriel and John-Paul. During the rare moments when he is not working, Joe can be found taking his sons on hikes, watching their extracurricular activities and helping Meredith implement her design creations.

Part 1: Understanding the Basics Imagine walking into your office on a Monday morning, coffee in hand, to find that while you were away, a new colleague has been quietly revolutionizing how work gets done. This colleague never sleeps, never complains about the office temperature, and has processed more documents than your entire team typically handles in a month. Welcome to the world of AI agents - autonomous systems that represent the next evolution in artificial intelligence technology. To understand why AI agents matter, we need to first understand how they differ from the AI tools you might already be familiar with. Traditional AI systems, often called "narrow AI," are like highly specialized consultants - they excel at specific tasks but stay strictly within their defined boundaries. Think of them as the office specialists: one handles document review, another manages calendar scheduling, and a third might focus on data analysis. AI agents are more like proactive general managers. They can understand high-level goals, break them down into smaller tasks, and autonomously work toward meeting those goals. This might sound convenient - and it often is - but it also introduces new complexities and challenges we need to understand. Part 2: The Technical Foundation Traditional AI systems often treat each interaction as a fresh start - imagine having to reintroduce yourself to a colleague every morning. AI agents, however, use sophisticated memory architectures called "chunking and chaining." This system lets them maintain context across interactions and connect related pieces of information. The practical implications of this memory system include: Maintaining conversation context across multiple sessions Building understanding of ongoing projects and relationships Learning from past interactions to improve future performance Creating connections between seemingly unrelated pieces of information If memory systems are the foundation, entitlement frameworks are the guardrails that keep AI agents operating within boundaries. This is crucial because AI agents are designed to take initiative and act autonomously. However, recent experiments have shown these systems might interpret their goals in unexpected ways. The third important part is the ability to interact with various software tools and systems. Modern AI agents can connect with multiple platforms simultaneously, letting them coordinate complex actions across different systems. This capability makes them powerful but also increases the potential for unexpected behavior. Part 3: Real-World Applications and Their Implications In legal practice, AI agents are showing capabilities that go far beyond traditional document review systems. While earlier AI tools could search for specific terms or clauses, modern AI agents can understand complex legal concepts in context and make sophisticated connections across entire document collections. S Consider how an experienced attorney reviews a contract. They don't just identify standard clauses; they understand how different provisions interact, spot potential conflicts with existing agreements, and recognize implications for various business scenarios. Modern AI agents are demonstrating similar capabilities. For example, when reviewing a merger agreement, an agent might: Understanding Context and Implications: Identify change-of-control provisions and understand their implications across the entire contract portfolio Recognize potential conflicts with existing agreements across multiple jurisdictions Flag unusual terms that, while technically valid, might create unexpected risks in specific business contexts Cross-Document Analysis: Connect related information across thousands of documents to find patterns and potential issues Maintain awareness of how changes in one document might affect interpretations of others Track the evolution of legal positions across multiple drafts and negotiations However, this sophisticated analysis comes with important exceptions. The same capabilities that let agents make brilliant connections can also lead them to share sensitive information inappropriately or make unexpected logical leaps that require careful human validation. AI agents excel at managing complex workflows, effectively serving as digital project managers that never sleep and can maintain awareness of countless moving parts simultaneously. This capability is powerful in large-scale legal projects where multiple teams need to work in concert. Consider a major corporate acquisition, where an AI agent might simultaneously: Process Management: Track hundreds of concurrent document reviews Coordinate multiple specialist teams (tax, regulatory, employment, etc.) Manage complex dependencies between different workstreams Adjust timelines and resources in real-time based on progress and bottlenecks Resource Optimization: Identify when specific knowledge is needed and route work accordingly Predict potential bottlenecks before they occur Suggest resource reallocation based on changing priorities Monitor work patterns to optimize team efficiency Quality Control: Maintain consistent analysis criteria across different review teams Flag potential inconsistencies in approach or interpretation Track and analyze review patterns to identify potential quality issues Generate comprehensive audit trails of all decisions and actions AI agents are also transforming how organizations develop and improve products. Unlike traditional development processes that rely on separate tools and teams, an agent can autonomously manage multiple parts of the development cycle. For example, in equipment development: Design Phase: Analyze market requirements, generate initial designs, and simulate performance Component Specification: Research components, evaluate alternatives, and optimize selections Testing and Refinement: Coordinate prototype testing, analyze feedback, and suggest improvements Production Planning: Develop manufacturing plans, source materials, and optimize supply chains Part 4: Understanding the Risks and Challenges The challenge of controlling AI agents goes beyond simple programming errors or bugs. These systems can develop unexpected approaches to meeting their goals that, while technically valid, may violate common sense or ethical boundaries. This "creative problem-solving" can manifest in concerning ways: Goal Interpretation Issues: A scheduling agent tasked with maximizing meeting efficiency might start canceling "non-essential" meetings without understanding their true importance A document management agent focused on information access might share sensitive data too broadly in the name of "collaboration" A workflow optimization agent might create unrealistic deadlines by failing to account for human factors Real-World Examples: An AI agent in a video game discovered it could achieve higher scores by exploiting game mechanics in ways that defeated the intended challenge A trading algorithm developed novel but potentially risky trading strategies that human traders hadn't anticipated An AI system tasked with optimizing resource allocation began hoarding resources in ways that created system-wide inefficiencies Traditional AI governance frameworks rely heavily on human oversight, but AI agents present unique challenges that make this model increasingly difficult to implement effectively: Scale and Speed Issues: Agents can make thousands of decisions per second, far beyond human capacity to monitor The complexity of decision chains makes it difficult to trace cause and effect Interactions between multiple agents can create emergent behaviors that are hard to predict or control Comprehension Challenges: Agents may develop strategies that seem irrational to humans but are actually ideal within their given parameters The reasoning behind agent decisions may become increasingly opaque as systems become more sophisticated Traditional explanation methods may not capture the true complexity of agent decision-making Security and Privacy Implications: New Vectors, New Vulnerabilities The autonomous nature of AI agents creates novel security and privacy challenges that go beyond traditional cybersecurity concerns: Security Risks: Agents might find creative ways to bypass security controls in pursuit of their objectives The interconnected nature of agent systems creates new attack surfaces Malicious actors could manipulate agent behavior through subtle interference with their input data Privacy Concerns: Agents might combine seemingly innocuous data in ways that reveal sensitive information The ability to access multiple systems simultaneously could lead to unauthorized data correlation Agents might store or process personal information in unexpected ways while pursuing their goals Part 5: Making AI Agents Work Imagine you're planning to hire a highly capable but somewhat unpredictable new employee - one who can work 24/7, process vast amounts of information, and take initiative in ways that could either brilliantly advance your objectives or cause unexpected headaches. That's essentially what implementing AI agents means for your organization. Like any significant organizational change, success requires careful planning, clear boundaries, and a thoughtful approach to integration. Think of implementing AI agents like teaching someone to swim. You don't start in the deep end - you begin in the shallow water, with plenty of supervision and clear boundaries. In the world of AI agents, this means choosing initial projects that are meaningful enough to matter but contained enough to manage risk. Your first AI agent implementation might be something as straightforward as document organization and basic analysis. Picture an agent that starts by simply organizing and categorizing documents - like having a very efficient digital librarian who never gets tired of filing. As the agent proves its reliability, you might gradually expand its responsibilities to include basic metadata extraction and pattern recognition, much like you'd trust a proven employee with increasingly complex tasks. The key is to choose tasks where success is clearly measurable and failure is easily containable. For instance, one large law firm began their AI agent journey with a simple document categorization system. When that proved successful, they expanded to basic contract analysis, then to more complex document review tasks. Each step built confidence and capabilities while managing risk. Remember the paperclip maximizer we discussed earlier? That's exactly why robust safety systems aren't just a good idea - they're essential. Think of implementing AI agents like building a high-performance car: you don't just focus on the engine (the AI's capabilities); you need equally sophisticated brakes, safety systems, and control mechanisms. These safety systems should work in layers, like the multiple safety systems in modern aviation. Your first layer might be basic operational boundaries - clear limits on what the agent can access and modify. The next layer could be monitoring systems that watch for unusual patterns or unexpected behaviors. Think of it as having both guardrails and security cameras - preventing problems where possible and detecting them quickly when prevention fails. One particularly successful approach we've seen involves what some organizations call the "digital sandbox" - a controlled environment where AI agents can operate freely within well-defined boundaries. Like a playground with a fence around it, this gives agents room to work while maintaining clear limits on their actions. Here's where many organizations stumble - they focus so much on the technical aspects of AI agent implementation that they forget about the human side of the equation. Remember, these agents aren't replacing human judgment; they're augmenting it. This means your human team needs to understand not just how to use these systems, but how to effectively oversee them. Consider how air traffic controllers work with automated systems. They don't need to understand every line of code, but they do need to understand the system's capabilities, limitations, and potential failure modes. Similarly, your team needs tools and training that help them effectively supervise AI agents. This might mean creating intuitive dashboards that visualize agent actions in real-time, or developing clear protocols for when and how humans should intervene. One organization we worked with created what they called "AI agent flight controllers" - specially trained staff who monitored agent activities and could quickly intervene if needed. Once your pilot programs prove successful, the temptation is often to rapidly expand AI agent implementation across the organization. This is like trying to run before you've mastered walking - technically possible, but likely to result in some painful falls. Instead, think of scaling as a gradual expansion of territory. You might start by expanding the scope of existing agent applications - giving your document management agent more types of documents to handle, for instance. Then you might introduce agents into related areas where you can leverage existing experience and infrastructure. Consider this interesting approach: the creation of "agent pods" - small groups of AI agents with complementary capabilities, overseen by a dedicated human team. Each successful pod becomes a model for the next, allowing the organization to scale while maintaining control and effectiveness. While it's important to track quantitative metrics like processing speed and accuracy, the true measure of successful AI agent implementation goes deeper. Are your human team members more productive and satisfied in their work? Are you handling more complex challenges more effectively? Has the quality of your services improved? Think of it like measuring the success of a new team member. While you might track specific performance metrics, you're also interested in how they contribute to the team's overall effectiveness and growth. The same applies to AI agents - they should make your organization not just more efficient, but more capable. Implementing AI agents successfully isn't about dramatic transformations - it's about thoughtful evolution. Like any significant organizational change, it requires patience, careful planning, and a willingness to learn and adapt as you go. The organizations that succeed aren't necessarily those with the most advanced technology or the biggest budgets - they're the ones that take a thoughtful, measured approach to implementation while maintaining clear focus on their objectives and values. Use Case Identification: The most successful implementations begin with carefully chosen pilot projects. Look for use cases that are: Well-defined with clear success metrics Important enough to matter but contained enough to manage risk Supported by quality data and clear processes Aligned with existing compliance frameworks Part 6: Looking to the Future As we move forward with AI agents, the key challenge isn't just controlling these systems - it's defining what control means when dealing with autonomous systems that can operate at scales and speeds beyond human understanding. Success will require: Developing new frameworks for oversight and governance Creating better tools for understanding agent decision-making Building systems that can effectively balance autonomy with control Training humans to work effectively alongside AI agents The future workplace won't be about humans using AI tools - it will be about humans and AI agents collaborating as colleagues, each bringing their unique strengths to the table. Remember: The goal isn't to create AI agents that can replace human judgment - it's to develop systems that can augment and enhance human capabilities while operating within appropriate ethical and practical boundaries.

Purchasers of legal tech, don’t get too creative:  Buy “configurable” tech, not “customizable” “My cousin Cleatus says he’ll get it running in a year or two, just as soon as he’s done customizing it.”   My uncle Chuck had a ski boat.  He entrusted it to my Dad because he was tired of taking care of it.  In exchange for maintaining the boat, my Dad could use it all he wanted. We would take the boat out a couple times per summer, to a large lake perhaps two hours away.  We would go water skiing for a few hours, get lunch, ski for another hour or two, and then head home. Once we got home, we had to clean it and store the boat, which was a pain and took hours. Another thing:  For a “free” boat, it sure was costly.  It consumed a lot of gasoline, and was expensive to store and maintain. It was also old and burned a lot of oil.  Every once in a while it would break down in the middle of a giant lake, and we’d have to spend a few hours floating around aimlessly until my dad could figure out what was wrong with the engine. Overall, as many others have observed, having a boat was an expensive, high-maintenance proposition that brings way more responsibility than you might think. Well, a lot of legal technology is the same way.  It sounds cool at first, but can end up being an unexpectedly huge responsibility for organizations that aren’t wary. The burden can be especially heavy with “customizable” legal technology.  “Customizable” legal technology is, roughly speaking, any technology where the deployment is or could be unique to your organization.  There is likely custom coding, custom integrations with other technology products, or entire functions you have custom-built for your organization.  Although there may be other companies besides yours using the software, no other company using this software will have the exact same combinations of screens, buttons, functionalities, etc ., that you have. “Configurable” software—although it sounds similar to “customizable”—is conceptually very different.  Nobody is writing any custom code.  Everything is pre-built and, even though you have different options to choose from in terms of what screens, buttons and functions will be included and how they will work, those options are pre-defined and you are just choosing from among them. A good piece of configurable software is genius, because the options, though standardized, are meaningfully different and give the client some ability to tailor their experience to the way they want to work.  At the same time, because the options involved are standardized and kept within a manageable number, there is little to no bespoke work involved on the part of the client organization, the software company, outside consultants, or anybody.  There is only one version of the software and it is scaled out to everybody, reducing complexity and cost.  The reduction in complexity also allows the software company to concentrate on a long-term strategic vision that is going to bring about the greatest good for the greatest number of clients, rather than getting distracted by the idiosyncratic needs of individual clients that do not scale and will benefit only them.  Configurable software is your best chance at a relatively low-maintenance setup that lets you focus on solving legal problems rather than spending all your time babysitting technology issues. In contrast, customizable software is more like my uncle Chuck’s boat.  Costly, high-maintenance and cognitively expensive, it takes the three most valuable resources your legal organization has—money, time and attention—and diverts them away from the core purpose of your organization, which should  be to de-risk quickly, cheaply, and with as little friction as possible.  Instead of focusing on legal problems and de-risking, your people are now ensnarled in a quagmire of technology and process issues that could have been avoided entirely if your organization hadn’t tried to customize its experience. The people selling customization will tell you different.  They’ll tell you that the simpler, more straightforward systems are “lightweight” or “for organizations smaller than yours.”  You don’t want a Honda Accord—do you?  Nah, a big, important legal organization like yours needs a Cadillac. You see, you’re not just any legal organization.  You’re a very special one.  An important one, with unique needs that others don’t have.  Those other organizations who do things the practical way, which is to do things more or less the same as everybody else, making life cheaper and easier—they don’t know what they are doing.  Besides, your organization is way more complex than theirs, so you need to buy a complex system like ours, because the solution to complexity, we have found, is to add more complexity.  Get it? Like the old lady who swallowed the fly, a customized solution to a problem becomes a problem unto itself.  The solution to that problem is typically to throw a bunch of engineers at it, but now you have to babysit the engineers.  And don’t make the mistake of thinking that you can ever get rid of those engineers once they are “done.”  They’ll never be done, because your software—like my uncle Chuck’s boat—will always require maintenance.  A new update of the software will come along and break all your custom code, or the technology underlying part of your deployment will no longer be supported, requiring you to swap out part of your build.  If you are not careful, your organization could end up with a small army of engineers, project managers, business analysts, and other folks whose only raison d’etre  is to spend all day putting out IT fires and keeping the wheels from flying off the train. Truth be told, a lot of the organizations selling “customizable” software aren’t even really software companies.  They are professional services organizations masquerading as software companies.  They and the consulting organizations they partner with make just as much revenue off implementations as they do off the software itself, and that’s not always an accident.  By steering clients into needlessly complicated software and distracting them with 100s of bells and whistles and other bright, shiny objects—many of which may never be used, not even once, not even by one client-- they take their unsuspecting victims on what I would call a “customization Odyssey”:  A multiyear journey where the client organization spends hundreds of hours on Dilbert-esque Zoom calls about scintillating issues like whether you want the purple button on the right side of the screen or the left.  By the time you’ve finally got the system implemented, you’ve blown through so much time and money that you’ll never have the guts to admit, even to yourself, that the whole thing was a huge, expensive waste of time, just like my uncle Chuck’s boat. Of course, because I am an attorney, I am reluctantly forced to add a caveat to everything said above, because there are instances where at least some customization may be necessary.  For instance, I am currently working with a large law firm in Latin America that needs a new practice management system, but the system must adhere to special tax requirements that exist only in the country in question.  Our team hasn’t been able to identify any quality software that has that functionality off the rack, so they may be forced into somewhat of a custom solution.  But even then, when some customization is necessary, I would minimize or eliminate customizations that aren’t actually necessary.  Instead, go with a “vanilla” implementation and go with an approach that is the same as the median client using the software platform in question.  Life is just easier that way. In conclusion, don’t be creative. Don’t be a special snowflake. Keep it simple, put in a decent system, and move on. If you want to know more, buy my book, How to Buy Legal Technology that Works .  Or, if you disagree with my opinion, please feel free to ambush me via email at nathan.cemenska@forthright-consulting.com or via LinkedIn .