AI agents help lawyers across separable jobs: legal research, first-draft contracts and briefs, document review, e-discovery, and client intake. Tools like Harvey, Spellbook, CoCounsel, and Lexis+ AI produce review-ready work. What none of them should own alone is the last step — verifying every citation and fact — because that is exactly where legal AI still fails.
The legal workflow, and which parts an agent can take
Legal work splits into a research-and-drafting front end and a judgment-and-attestation back end. An AI agent is genuinely useful for the repetitive front end — pull the relevant authority, draft the clause, review the data room, summarize the deposition — because a lawyer checks the output before it reaches a client or a court. Harvey's own framing is that legal teams use agents to "draft, analyze, research, and deliver review-ready work while keeping lawyers focused on strategy and judgment." The part to keep human is the one that carries professional liability: the sign-off, where a lawyer attests the work is correct. So the honest way to read any "AI agent for law firms" pitch is to ask, at each stage, does the agent do the work, or make the call? It's the same automate-the-work-not-the-decision line we drew for accountants and recruiters, and — before any of them — for inbox triage in Issue #001.
Research and draft — let the agent produce, not conclude
Research and drafting are where agents remove the most hours. Purpose-built legal tools generate first drafts of contracts, briefs, and memos and surface the authority behind them: Spellbook markets agents that draft and redline contracts inside the document, and Harvey positions its agents around research and drafting for litigation and transactional teams. The discipline is the same one we hold everywhere on this site: the agent produces a draft, a person makes the call. That matters more in law than almost anywhere, because the failure mode is not a typo — it is a confidently fabricated case.
Verify — the step you cannot delegate
Legal AI still invents authority, and the numbers are documented. A Stanford RegLab study — the first preregistered evaluation of legal research tools, hand-scored across 202 queries — found Lexis+ AI produced incorrect information more than 17% of the time and Westlaw's AI-Assisted Research hallucinated more than 34% of the time, despite vendor "hallucination-free" claims. These are the purpose-built legal tools, not a general chatbot.
The consequences are not hypothetical. Courts are sanctioning the lawyers who skip the check: Norton Rose Fulbright's 2026 litigation update records a March 2026 Sixth Circuit order in Whiting v. City of Athens fining two attorneys $15,000 each after their briefs contained more than two dozen fabricated citations. This is why the profession's own guidance draws the line at verification: ABA Formal Opinion 512 (July 2024) holds that a generative-AI tool "cannot solely substitute for a lawyer's competent legal work," and that the required verification depends on the task — a document review demands more independent review than idea generation. The citation check is the one step you cannot delegate.
Which agent should a lawyer start with?
Start where your practice actually stalls. If contract turnaround is the bottleneck, a drafting-and-redline tool like Spellbook works inside the documents you already edit. If research and litigation drafting eat your week, Harvey is built for that end. New to agents entirely? Begin at Agent 101, or see the no-code AI agent tools you can wire a single legal step into. Whatever you start with, hold the same line: automate the drafting, own the citation check.
FAQ
What AI agents can help lawyers? Match the agent to the task. For contract drafting and redlining, Spellbook works inside the document; for research and litigation drafting, Harvey delivers review-ready work; for legal research inside the big databases, Lexis+ AI and Westlaw AI-Assisted Research surface authority — though a Stanford RegLab study found both still hallucinate. Start with the task that costs you the most billable hours, and verify every output.
Can an AI agent do my legal work on its own? No. It can draft, research, and review, but the sign-off must stay human — and here that is an ethics rule, not just caution. ABA Formal Opinion 512 states that a generative-AI tool cannot substitute for a lawyer's competent legal work, and that verification scales with the task. The attestation carries professional liability, which is exactly where a person belongs.
Do legal AI tools really make up cases? Yes, and it is measured. The Stanford RegLab study found purpose-built tools hallucinated between roughly 17% and 34% of the time. Courts have responded with sanctions — Norton Rose Fulbright documents a $15,000-per-attorney order in Whiting v. City of Athens over fabricated citations. Never file an AI-drafted citation you have not read in the reporter yourself.
Do I need code to use legal AI agents? No. Harvey and Spellbook ship as software built for lawyers, not developers. If you want to wire a single legal step — intake triage, document sorting — into a custom flow, the no-code AI agent tools guide covers the builders that can do it.
Where should a solo or small firm begin? With the one task that eats the most hours and is safe to check. For most that is drafting — a first-draft engagement letter or contract you review line by line, the same capture-don't-commit setup we show for inbox triage in Issue #001. Layer research and review on top only once your verification habit is solid.
One real AI agent a week, straight to your inbox. Free, no upsell.