It's 4:10 on a Thursday at a four-lawyer firm. A client who runs a growing HVAC company forwards a 38-page software agreement and wants to sign by Friday noon. The partner who handles commercial work is in a deposition until five, and the associate has two other deadlines tonight.
This is the moment AI contract review is built for. Not to tell the client whether to sign, but to make sure that when a lawyer sits down with the agreement at six, the hunting is done and the thinking can start.
What AI contract review catches
Good tools do five things, and each one points back to the page and section it came from:
- Clause extraction. Indemnification, limitation of liability, term and renewal, termination, governing law and venue, assignment, confidentiality, data security, insurance, payment terms and IP ownership, each pulled out with its location.
- Playbook comparison. Each clause set beside your firm's standard position and acceptable fallback, with the gap described in plain language.
- Missing clauses. No cap on the vendor's liability, no breach-notification duty, no way out before the term ends.
- Defined terms and cross-references. “Fees” defined in Exhibit B, a cap that refers to a section that was renumbered, and documents incorporated by reference but not attached, such as an online acceptable-use policy.
- A summary with page cites, so the lawyer can go straight to the language behind each point.
Flowchart
One agreement through AI-assisted review
- StartClient forwards a 38-page vendor agreement at 4:10 PM Thursday
- StepText extracted; defined terms, exhibits and cross-references mapped
- StepEach key clause compared with the firm's playbook for vendor agreements
- DecisionClause outside the playbook, or missing entirely?YesStepAdded to the issues list with page, quote and playbook position; review continuesNo
- DecisionRefers to terms or policies that aren't attached?YesStepAdded to the list as a request to send the vendor before anyone advisesNo
- A personLawyer reads the agreement with the issues list and decides what matters for this client
- DoneLawyer-approved issues memo and markup go to the client
Before and after
The first pass, before and after
Reading cold
- 01Associate reads all 38 pages from the top
- 02Issues noted in margin comments as they come up
- 03Firm positions checked from memory or an old memo
- 04Exhibits and linked policies reviewed last, if at all
ResultAn evening gone, and a quiet risk of a missed clause
Playbook-first review
- 01Issues list with page cites waiting before the lawyer starts
- 02Each deviation shown beside the firm's standard position
- 03Missing clauses and unattached documents listed up front
- 04Lawyer reads the full agreement with the list as a checklist
- 05Lawyer decides what to push back on and what to let go
ResultLawyer time goes to judgment instead of hunting
Ask for the output in the shape your clients actually read. A one-page issues list sorted into three groups works well for a business owner: must change, worth asking about, and fine as written. Each line carries the page, the quoted language and, for the first group, proposed wording from your own clause bank.
That proposed wording is a draft like any other. It came from your clause bank, but whether it fits this deal, and whether to spend the client's goodwill asking for it, is the lawyer's call.
Build the playbook before you buy anything
A contract review tool is only as useful as the positions you give it. Without a playbook, it compares the agreement with some general idea of “market,” which may not be your market or your client's. Start with one agreement type you see often and write ten to fifteen entries. The rows below show the shape of an entry; the positions themselves are your firm's to set.
| Clause | Our standard position | Acceptable fallback | Always flag |
|---|---|---|---|
| Limitation of liability | Mutual cap tied to fees paid | Cap with carve-outs for confidentiality and data | Cap that protects only the vendor |
| Indemnification | Mutual, for third-party claims | Vendor covers IP infringement at minimum | One-way, in the vendor's favor |
| Term and renewal | Renews only with notice before the date | Auto-renewal with a reminder duty | Renewal with a price increase and no notice |
| Governing law and venue | Client's home state | A neutral state | Distant venue or mandatory arbitration |
| Assignment | Neither party without consent | Assignment allowed in a sale of the business | Vendor may assign freely |
| Data and confidentiality | Vendor protects client data and reports breaches | Breach notice within a set period | No breach-notification duty at all |
Three kinds of tools you'll come across
Most options fall into three groups: playbook add-ins that run inside Word, contract review features built into a practice-management or document system, and general AI assistants used under business terms. Whatever the type, check four things before a trial: can it hold your own playbook, does every point cite a page and section, will the vendor commit in writing not to train on your documents, and how long does it keep them?
What it misses
The limits aren't about reading ability. They're about everything that isn't on the page.
- Context and bargaining power. Whether this client can push back, how much the relationship matters and what was promised on a sales call.
- The business deal. Whether the price, scope and service levels make sense for an HVAC company of this size.
- State-specific nuance. Rules on clauses such as non-competes and automatic renewals differ by state and change over time. A tool may not know your state's current position, or may sound sure when it isn't.
- How clauses interact. An indemnity that sits outside the liability cap, or a definition that quietly widens an obligation three sections later.
- What isn't in the file. Order forms, side letters, earlier emails and handwritten changes on a signed copy.
Who signs off, and what the ethics guidance expects
ABA Formal Opinion 512, issued July 29, 2024, doesn't invent new rules for generative AI. It walks through how existing ones apply: competence, confidentiality, client communication, supervision and fees. Competence includes a reasonable understanding of what a specific tool can and can't do. On confidentiality, the opinion calls for informed client consent before confidential information goes into a self-learning tool, and says a general clause tucked into an engagement letter won't do.
On verification, the opinion says the appropriate level of independent review depends on the tool and the task, and that uncritical reliance on output can fall short of competence. It discusses an example close to this one: a lawyer who has already tested a tool's accuracy on a smaller set of contracts may not need to manually re-check every summary in a large set. That speaks to checking summaries, not to skipping the agreement you're advising on.
Supervising lawyers are expected to set clear policies and train the people using these tools, and hourly bills should reflect time actually spent, not time saved. State bars have published guidance of their own; Florida's Ethics Opinion 24-1, for example, stresses researching a tool's data retention, sharing and self-learning policies and verifying its output. Yours may go further than the ABA. Read this as general information, not legal advice for your firm.
A sign-off routine for a small firm
- 01Pick one agreement type to start, such as vendor agreements, NDAs or commercial leases.
- 02Write the playbook with the partner who owns that work.
- 03Run the tool on ten agreements you've already reviewed and compare its issues list with yours. Note every miss.
- 04Set the rule: a lawyer reads every agreement in full; the issues list is a checklist, not a substitute.
- 05Record in the file which tool and playbook version were used and which lawyer reviewed the output.
- 06Check the vendor's terms on confidentiality, training on your documents, retention and where data is stored, and get client consent where your analysis says it's needed.
If your firm wants playbook review set up inside its own document system, our document processing service covers it, and our page for law firms shows where it fits alongside intake and billing. For the wider question of keeping AI output honest, see how to prevent AI hallucinations in a business.