Legal AI Agents, Explained: What Law Firms Automate First, and What Stays With a Lawyer
A legal AI agent, explained: the three jobs law firms automate first — contract review, intake, drafting — the rules it runs under, and what it costs.
Written by Max Zeshut
Founder at Agentmelt
TL;DR: A legal AI agent is a narrow workflow that does the reading, comparing and drafting a law firm or an in-house legal team repeats every week — and stops at a lawyer before anything leaves the firm. Three jobs come first almost everywhere: first-pass contract review against the firm's own playbook, client intake with the conflict-check facts extracted, and routine documents drafted from templates. The agent marks deviations, extracts parties and fills templates; the lawyer decides, redlines and signs. It is not legal advice, it does not research case law on its own authority, and nothing goes to a client or a counterparty unreviewed. Each job runs from a free template or $247–297 a month per process, a legal package installs all three for $490, and the first one is usually live within two working days of access.
Buildable version: the contract review workflow, the client intake workflow and the document generation workflow — what arrives, what gets checked, who signs, a free template and the price to have it run for you — or all three together in the legal professionals package.
What a legal AI agent is, and what it is not
A legal AI agent is software that takes a repeatable piece of legal work, does the mechanical part, and hands a lawyer a file that is ready for judgement. The mechanical part is larger than most firms assume: finding the liability cap in a 40-page MSA, noticing that the termination notice is 30 days when the firm accepts 90, pulling every party out of an enquiry for the conflict check, copying the right fee schedule into an engagement letter. None of that needs a law degree. All of it is where the hours go.
What it is not matters as much. It is not a robot lawyer, and the workflows here do not give advice, quote fees, decide whether a conflict is real or send a redline on their own. The positions it applies — what the firm accepts, what it falls back to, where it walks away — are written by the firm and owned by the firm. The agent applies them the same way every time, which is the part people are inconsistent at.
| What the legal AI agent does | What stays with a lawyer | |
|---|---|---|
| Contract review | Extracts every clause, compares each to the playbook's standard, fallback and walk-away position, scores the deviation, drafts the redline in the firm's fallback language | Whether a deviation is acceptable for this deal; every redline before it goes to the counterparty |
| Client intake | Classifies the enquiry by matter type, extracts parties and dates, checks them against the client, former-client and adverse-party lists, sends a same-day acknowledgement, drafts the engagement letter or the decline | Whether a conflict hit is real; whether to take the matter; the fee; anything that reads as advice |
| Document drafting | Fills NDAs, engagement letters, SOWs and routine filings from the firm's templates and the matter's details, highlights every variable, lists the open decisions at the top | The final read, the decisions in the list, the signature |
| What moves the result | Whether the playbook, the matter taxonomy and the templates are written down | Nothing — this column is the point |
What law firms automate first, and why
Firms start where volume meets a written standard. That is almost always one of three places.
Contract review, if the firm reviews more than about ten agreements a month. NDAs, DPAs, standard MSAs and vendor terms are high-volume and well understood, and most of them do not deviate from the firm's positions at all. The contract review blueprint routes by risk: no deviations go to the business owner, minor ones to legal operations with the redline, and anything touching liability, indemnity or IP always reaches a lawyer with the full comparison. Lawyers review deviations, not documents. The catch is the one every firm hits: the playbook usually lives in a partner's head, and the setup work is writing it down — standard, fallback, walk-away, clause by clause. That document is valuable on its own.
Client intake, if enquiries wait a day for a reply. Enquirers contact several firms and usually hire the one that answers first. The client intake blueprint acknowledges within minutes, classifies against the firm's own matter types, extracts every party — including the ones mentioned in passing — and checks them with fuzzy matching against every list the firm keeps. A hit blocks the engagement letter until a person clears it. The conflict check stops being a spreadsheet search done when the day allows; it runs on every enquiry, within the hour. The law firm client intake guide walks through the eight steps.
Document drafting, if letters start from the last one. Copying the previous engagement letter carries the previous client's fee, scope and name. The document generation blueprint drafts from the template for the matter type, applies the fee schedule, and puts anything it could not source — a date, a discount, a commitment — in a decision list at the top rather than in the body.
| Job | Start here if… | Runs on | What moves it |
|---|---|---|---|
| Contract review | 10+ agreements a month, mostly NDAs, MSAs and vendor terms | Word or Google Docs, the playbook as a document, email or the contract repository | How complete the playbook is; the first two weeks add the clause types it was missing |
| Client intake | Enquiries wait hours or a day; conflict checks are a manual search | Outlook or Gmail, Clio, PracticePanther or a spreadsheet | Whether the client, former-client and adverse-party lists are in one place and current |
| Document drafting | Engagement letters and routine filings are copied from the last one | Word or Google Docs, the matter list | Whether there is one approved template per matter type and one fee schedule |
Legal research — finding precedents, summarising case law — is the use most people picture first, and it is rarely where a firm should start. The output needs checking line by line, the volume is irregular, and it does not remove a queue. It is covered in AI legal research: finding precedents.
The rules every legal AI agent here runs under
- First pass, never final. Every redline, letter and decline is accepted or rejected by a person. Nothing leaves the firm unreviewed.
- The firm owns the positions. The playbook is written and versioned by the lawyers; the agent cannot add a position. A clause type it has not seen is marked "not in playbook" and read by hand.
- Major risk always reaches a lawyer. Liability, indemnity and IP clauses go to a lawyer regardless of the deviation score.
- Confidence is shown. Low-confidence extractions are flagged for manual reading rather than compared.
- Client data stays in the firm's systems. The automation runs inside the firm's own Word, Google Docs or Clio accounts; model calls go through a provider with a zero-retention agreement and no copies are kept. The install includes a one-page data-flow note for the file.
- Everything is logged. Each intake keeps an audit line — received, classified, checked, approved by whom, when — and each review outcome is recorded on the matter.
Most bars ask for competence, confidentiality and supervision when lawyers use AI. A first pass with mandatory review, running in the firm's own accounts with zero retention, is built to fit that, but the firm's own bar guidance governs.
What to expect in numbers
The figures come from two places in our own material: the legal package's worked example — an illustrative two-partner commercial practice on Google Docs, about 14 contracts a month, not a client report — and the corporate law firm case study, a 25-attorney M&A practice that ran first-pass review on specialist legal tools rather than on these workflows.
| Metric | Before | After | What moves it |
|---|---|---|---|
| First-pass review per contract | 1 h 45 min | 22–25 min by weeks 2–4 (worked example) | Playbook coverage: "not assessed" clauses fell from three per contract to one as the clause types were written down |
| Hours a month on first passes | about 25 | about 6 (worked example) | Volume and contract mix; NDAs with no deviations need almost no lawyer time |
| Paralegal time on a standard contract | 20–30 min | 10–15 min (case study) | Paralegals validating flagged risks instead of reading every page |
| Flag accuracy against manual review | 89% at the start of the parallel run | 97% after the playbook was refined (case study) | A parallel run on live matters before switching over; the firm used two deals |
| Due-diligence review cycle per deal | 4 weeks | under 2 weeks (case study) | Throughput per paralegal (3x in the case study), not fewer people |
| Documents leaving the firm unreviewed | 0 | 0 | This one should never move |
Partner review time does not fall, and it should not: the saving is in the first pass, not in the judgement. To measure it on your own matters, time the first pass on the next ten contracts of each type before you start, run the workflow beside a person for two weeks, and compare minutes per contract and the deviations each side caught. For intake, the number to watch is enquiry-to-acknowledgement time and the share of enquiries conflict-checked on the day they arrive.
What it costs
As workflows on the systems the firm already runs: a free template to build it yourself, or a monthly subscription to have it run for you — contract review at $297 a month for three contract types and up to 150 contracts, client intake at $247 a month for one intake inbox and up to 300 enquiries, document drafting at $247 a month for three document types and up to 50 documents. The legal professionals package puts all three together: the kit with the templates and role prompts at $49 to set up yourself, the first workflow installed in your own accounts at $249 (live within two working days of access), or all three adapted to the firm's positions and house style at $490, live in about a week. Building the playbook from past agreements, several practice areas with their own rules, integration with a contract lifecycle system or a client portal are a custom build from $2,000; the blueprints price the larger ones at $3,500–9,000. The free audit names the job to start with from your answers, and the plan arrives by email.
Questions, answered
What is a legal AI agent?
A legal AI agent is a workflow that does the mechanical part of repeatable legal work — extracting clauses and comparing them to the firm's playbook, pulling parties out of an enquiry for the conflict check, drafting routine documents from templates — and hands a lawyer a file ready for judgement. It applies positions the firm wrote; it does not give advice, and nothing it produces leaves the firm until a person approves it.
What do law firms automate first with legal AI agents?
First-pass contract review if the firm reviews more than about ten agreements a month, client intake if enquiries wait hours for a reply, and routine drafting if engagement letters are copied from the last one. All three have a written standard to apply and enough volume to matter. Legal research usually comes later, because its output needs line-by-line checking and it does not clear a queue.
Can an AI legal agent replace a lawyer?
No, and the workflows here are built so it cannot act like one. The agent marks deviations and proposes the firm's own fallback language; a lawyer accepts or rejects every redline. It extracts conflict-check facts; a person decides whether a hit is real. It never quotes a fee or says anything that reads as advice. What changes is where lawyers' time goes: to deviations and decisions rather than reading pages to find them.
Is it safe to put client documents through a legal AI agent?
It can be, if the documents never leave the firm's systems. The workflows run inside the firm's own Word, Google Docs or Clio accounts, model calls go through a provider with a zero-retention agreement, and no copies are kept outside the firm. The install includes a one-page data-flow note you can put in the file. Check it against your own bar's guidance on confidentiality and supervision.
How much does a legal AI agent cost?
From a free template if you build it yourself, or $247–297 a month per process to have it run for you. The legal professionals package adapts and installs contract review, intake and drafting together for $490; $249 installs the first one. Building the playbook from past agreements, several practice areas or a contract lifecycle integration are custom builds from $2,000.
Sources and further reading
- AI legal agents: the pillar — what they do, the tools, the cost
- Contract review: the blueprint, template and price
- Client intake, from enquiry to engagement letter: the blueprint
- Document and proposal generation: the blueprint
- The legal professionals package
- AI agents for legal professionals
- A corporate law firm's first-pass contract review: the case study
- AI contract review: how legal agents save hours per deal
- Law firm client intake automation