AI Agents for Lawyers
- Colin S. Levy
- 1 day ago
- 2 min read
Most of the attention on AI in legal practice has focused on generative tools: drafting, summarizing, answering research questions. This guide is about what comes next. Agentic artificial intelligence doesn't respond to prompts. It pursues goals. You hand an agent a contract to process and it reads the document, selects the playbook, drafts redlines, routes the file, and updates the matter record without being re-prompted between steps. The lawyer's job shifts to designing the workflow, setting the boundaries, and reviewing what comes out.
That shift has real consequences. When an agent acts across connected systems, errors don't stay contained. A misclassification at step one produces the wrong playbook at step two, wrong redlines at step three, and the file going to the wrong attorney at step four. I include a worked example of exactly this in Part Three: a contract intake that looks fine at every individual step and produces a serious problem at closing because no one caught a DPA buried in what the agent read as a standard MSA. Each step was defensible. The compounding is what caused the harm.
The professional responsibility framework is more developed than many lawyers expect. ABA Formal Opinion 512, along with state-level guidance from California, Florida, New York, and others, treats AI tools as non-lawyer assistance governed by Rules 5.1 and 5.3. The supervising lawyer is accountable for what the agent does, the same way they're accountable for what a paralegal does. Rule 1.5 bars billing for time AI saves. Some jurisdictions require affirmative disclosure of material AI use; others don't. Lawyers with EU-facing work face additional obligations under the AI Act and GDPR that U.S. ethics rules don't address. The guide maps all of this, including where state positions diverge from the ABA framework in ways that actually matter to practitioners.
The governance section is where I focused most of the practical work. I give a framework for scoping what an agent can access and do, a decision tool for figuring out which steps require a human in the loop before execution, guidance on building audit logs that would satisfy a court or regulator rather than just a debugging session, and a vendor questionnaire that covers the questions standard SaaS diligence misses entirely. There is also a pilot readiness checklist for teams getting started. The goal throughout was to produce something short enough to actually use.