A law firm running on AI is not a firm that handed its judgment to a machine. It is a firm that pointed a reviewed AI workflow at the high-volume, low-discretion work, the first research memo, the chronology, the intake summary, the routine letter, and kept a licensed lawyer on every result that leaves the building. The drafting moved. The duty did not. The firms doing this well treat AI like a fast, tireless, occasionally wrong junior associate: useful for a first pass, never trusted without a check, and never shown a client confidence until the privilege question is settled in writing.
If you practice law, you have spent two years inside the noise. One headline says AI will replace associates by next quarter. The next says a lawyer got sanctioned because the tool invented a case. Both are describing the same tool from opposite ends of the same mistake: treating a drafting assistant as a source of truth. This page does something more useful. It walks the firm, function by function, and shows where AI genuinely earns its keep, where it gets lawyers in trouble, and what every careful firm keeps human.
Read it as a working map, not a sales pitch. Every section links to a deeper briefing where we show the actual workflow. This is the hub; the briefings are the rooms.
Key takeaways
- AI moves the first draft across the whole firm: research memos, motions, discovery review, medical chronologies, demand letters, intake summaries, and routine correspondence. A licensed lawyer keeps the judgment and the signature.
- The two non-negotiable guardrails are confidentiality and verification. Decide what client data may enter a tool before anyone uses one, and check every citation and fact the AI produces, every time. Lawyers have been sanctioned for skipping the second step.
- The biggest, safest wins are in research and drafting where a human reviews the output anyway, not in autonomous decisions or advice that goes out unread.
- Your duty of competence now includes understanding the tool. The ABA and several state bars treat basic AI literacy as part of the job, not an optional extra.
What AI actually does across the firm
Start with the map. The table below is the whole pillar in one view: the function, what AI does well there, the discipline that keeps it safe, and the briefing where we walk the real workflow. Nothing here is autonomous. Every row assumes a lawyer reviews the output before it counts.
| Function | What AI does well | The guardrail | Go deeper |
|---|---|---|---|
| Legal research | Issue spotting, summarizing a body of law, finding the shape of an argument fast | Every cite verified in a real database. AI suggests, it does not source | Research vs. drafting |
| Drafting | First drafts of motions, briefs, letters, and clauses from your own materials | Lawyer rewrites and owns the argument; AI never has the final word | Draft motions and briefs |
| Discovery and review | First-pass document review, summarizing depositions, building timelines | Your prompts and outputs may be discoverable; assume they will be read | Are AI prompts discoverable |
| Medical chronologies | Turning a stack of records into a dated, sourced timeline in a fraction of the time | Spot-check entries against the underlying records before relying on them | Medical chronology workflow |
| Demand letters | Structured first drafts of PI and other demand letters from the case file | Numbers and liability theory reviewed by the lawyer who signs it | Demand letters with AI |
| Client intake | Summarizing intake notes, drafting follow-up questions, organizing facts | Confidentiality decided before any client detail enters a tool | Client notes to billable drafts |
| Billing and ethics | Drafting bill narratives, flagging time, summarizing matters | Bill for the work done, not the time AI saved; disclose where required | Billing for AI-assisted work |
AI in a law firm is a fast junior associate who never sleeps, never bills, and occasionally makes things up. You would not file that associate's first draft unread. The rule does not change because the associate is software.
Research: the fast first read, never the citation
This is where most firms start, and where the most public mistakes happen. AI is genuinely good at the early read: spotting issues, summarizing a doctrine, telling you the shape of an argument before you have spent three hours in the library. It is genuinely bad, and dangerous, at sourcing. A general assistant will produce a confident citation to a case that does not exist, because it is built to predict plausible text, not to retrieve verified law.
The discipline is simple and absolute: AI helps you find the argument, a real legal database confirms the law. Lawyers in multiple jurisdictions have been sanctioned for filing AI-invented citations they never checked. The fix is not to avoid AI. It is to separate the two jobs and never let the drafting tool be your source of truth. We walk the exact split, what to ask AI for and what to confirm yourself, in AI legal research versus drafting, and the sanction pattern in AI citation hallucinations in legal filings.
Drafting: where the hours actually come back
The largest, safest time savings in most firms are in drafting, because the lawyer reviews the output anyway. Pointed at your own templates and case materials, AI produces a usable first draft of a motion, a brief section, a routine letter, or a contract clause in minutes instead of hours. The blank page disappears. The lawyer's job becomes editing and owning the argument, which is the part they are actually paid for.
The boundary is that AI drafts, the lawyer decides. The argument, the theory of the case, the words that go in front of a judge or a client, those are the lawyer's, reviewed line by line. See the workflow in drafting motions and briefs with Claude, and the firm-wide version in the senior lawyer AI operating model.
Discovery, chronologies, and the discoverable prompt
On the litigation side, AI does its quietest, most valuable work: first-pass document review, deposition summaries, and the dated medical chronology that used to eat a paralegal's week. A reviewed tool can turn a banker's box of records into a sourced timeline you spot-check, rather than build from scratch. The medical chronology workflow is one of the highest-leverage uses in the whole firm.
Treat your AI prompts and outputs as potentially discoverable. How you instructed the tool, and what it returned, can become a record. Do not type anything into an AI prompt you would not want read aloud in a deposition, and settle the confidentiality question before any client fact goes in. The full analysis is in are your AI prompts discoverable.
Demand letters and client correspondence
For plaintiff-side and PI work, the demand letter is a perfect fit: high-volume, structured, and built from a case file the lawyer already knows. AI assembles the facts, the timeline, and the liability narrative into a strong first draft. The lawyer reviews the numbers, the legal theory, and the tone, then signs. The result is more letters, written better, in less time, without the lawyer ever surrendering the judgment that makes the letter land.
We cover the full method, including how it compares to specialist tools, in demand letters with AI, how to write a demand letter with Claude, the PI demand letter template, and the head-to-head in EvenUp versus Claude.
Intake and billing: the back office that runs itself
At the front door, AI summarizes intake notes, drafts the follow-up questions, and organizes the facts of a new matter so the lawyer walks into the first call prepared. At the back, it drafts bill narratives and matter summaries. Both are real wins, and both carry a specific ethics rule.
On intake, confidentiality is decided in writing before any client detail touches a tool. On billing, you bill for the work and value delivered, not for the time the AI saved you, and you disclose your AI use where your engagement letter or jurisdiction requires it. The lines are drawn in client notes to billable drafts and billing for AI-assisted legal work.
The four guardrails every careful firm keeps
Across every function above, the same four disciplines separate the firms that benefit from the firms that get sanctioned. None of them is technical. All of them are the practice of law.
- Confidentiality first. Before anyone uses a tool, decide in writing what client data may enter it and which tools are approved. Privilege and the duty of confidentiality do not pause for convenience. See the confidentiality guide for attorneys.
- Verify everything. Every citation, every fact, every number the AI produces gets checked against a real source before it leaves your hands. This is the single rule that keeps lawyers off the sanctions list. See AI citation hallucinations.
- Competence includes the tool. Your duty of competence now extends to understanding the technology you use. The checklist is in the AI competence duty for lawyers, and court-level rules are tracked in AI court rules in New York and Florida.
- The signature stays human. A licensed lawyer reviews and owns everything that goes to a client, a court, or the other side. The AI never signs. The lawyer always does.
What does not change
It is worth saying plainly, because the marketing rarely does. AI does not give legal advice. It does not exercise judgment. It does not carry your bar license, your duty to the client, or your liability when something goes wrong. It cannot read a room, manage a difficult client, or decide whether to settle. The relationship, the strategy, the courtroom, and the trust are still entirely yours.
What changed is narrower and more valuable than the hype: the blank page, the first read, the receipt-pile equivalents of legal work. A firm running on AI is a firm whose lawyers spend less time assembling and more time deciding. That is the whole story, and it is enough.
The drafting moved. The duty did not. Any firm that confuses the two is one unverified citation away from a sanctions hearing.
How to think about your first step
Do not buy a strategy. Pick one painful, high-volume task you already review by hand, the research memo, the chronology, the demand letter, and point one approved, confidentiality-cleared tool at it for a month. Verify the output every time. Measure the hours. When it works, stack the next task. The firms that expect a revolution in week one are the ones that quit in week three. The ones that compound small, verified wins are the ones that, a year later, quietly run on AI.
For where the broader rules are heading, follow our AI regulation news hub, and for reusable patterns across professions see the AI workflows library. If you are still choosing a tool or a course, start with Claude for lawyers and what to look for in a Claude course for lawyers.