AI contract review reliably catches what's on the page: missing standard clauses, deviations from your playbook, one-sided indemnities, auto-renewals and notice periods, and inconsistent defined terms or dates. It misses what's off the page: the client's commercial deal, how clauses interact, terms incorporated by reference, schedules it wasn't given, and whether a clause will hold up where it's enforced.
The reason is how these tools work. They compare the text in front of them with patterns learned from vast numbers of contracts and, in legal tools, with a playbook of your preferred positions. They report on that text with the same confident tone whether the finding is solid or a guess. So the useful way to think of AI review is as a fast, thorough first pass that produces an issues list. A lawyer still reads the whole contract with the client's deal in mind.
How AI reads a contract, and why that shapes its blind spots
Three mechanics explain nearly every miss.
- It only knows the text you give it. If the payment terms sit in a schedule you didn't upload, or in the supplier's standard terms "available on request", the review can't see them. Some tools say so; many simply review what's there.
- It judges clauses mostly one at a time. Models are good at "is this indemnity one-sided?" and weaker at "does this indemnity escape the liability cap in clause 14 because of the carve-out in clause 14.3?" Interactions across a long document are where attention thins out.
- It has no view of the deal. The model doesn't know that the client agreed a 60-day payment term on a call last week, or that exclusivity is the thing the client cares about most. It can only compare the draft with what's usual.
Scanned documents add a fourth: the text is extracted first, and a misread "30" as "80" in a notice period flows into the review without any warning.
Word files with tracked changes can cause a quieter version of the same problem. Depending on how a tool converts the file, deleted text may be read as if it were still in force, so a review of the other side's markup can praise a 100% liability cap that was struck out two rounds ago. If the tool doesn't say it reads revisions properly, review a clean copy: accept all changes into a duplicate file, or export the current version to PDF, and run the markup comparison separately.
What it catches well
| Issue | Example | How reliable | Still check |
|---|---|---|---|
| Missing standard clauses | No limitation of liability, no termination for convenience, no data protection clause | High | Whether the client actually needs it in this deal |
| Deviations from your playbook | Liability cap at 50% of fees where your position is 100% | High in legal tools with a playbook; medium in general assistants | That the playbook itself is up to date |
| One-sided terms | Indemnity from your client only; unilateral variation rights | High | Whether it matters commercially |
| Renewal and notice traps | Auto-renews for 12 months unless notice given 90 days before expiry | High | The dates, calculated yourself |
| Drafting inconsistencies | "Services" defined but "Deliverables" used; clause references that point to the wrong number | Medium to high | Cross-references in the final version |
| Unusual clauses | A non-solicitation clause in a simple supply agreement | Medium | Why it's there, which usually means asking the client |
What it misses, with examples
The commercial deal
A letting agency client's new software contract says payment is due in 14 days. The review marks this as "standard". The agency's director had negotiated 45 days on a call and assumed it was in the draft. Only the lawyer who asked "what did you agree?" catches it.
How clauses interact
A typical pair that AI reviews separately and passes:
14.1 Supplier's total liability shall not exceed the fees paid in
the 12 months before the claim.
14.3 Clause 14.1 shall not apply to any liability under clause 9
(Indemnities).
9.2 Customer shall indemnify Supplier against all losses arising
from Customer's use of the Platform.Each clause looks ordinary. Together, they cap the supplier's liability to your client while leaving your client's indemnity to the supplier uncapped. A good legal tool may catch this with the right playbook; a general assistant usually needs to be asked about it directly.
Terms incorporated by reference
"This Agreement incorporates the Supplier's Standard Terms of Business, available at [website]." The review covers the eight pages uploaded. The twelve pages of standard terms that actually govern cancellation charges were never seen.
Schedules and annexes that weren't uploaded
The model reports "payment terms are set out in Schedule 2 and appear standard". There was no Schedule 2 in the upload. The sentence is a guess dressed up as a finding, which is why the review prompt below asks for clause quotations.
Whether a term will hold up
Restrictive covenants, penalty-style charges and limitation clauses are treated differently under different governing laws. A general model tends to apply a blended average of what it has seen, and legal tools are only as good as the jurisdiction coverage of their content. Enforceability is a lawyer's call.
Numbers
Fee schedules, cancellation percentages and minimum spend calculations get read, not calculated. If a cancellation table says 25% at 60 days, 50% at 30 days and 100% at 14 days, the tool will describe it; it won't notice that a clause elsewhere treats "30 days" as calendar days and another as business days.
Dates built on definitions go wrong the same way. Take an illustrative software agreement signed on 3 March, with a 12-month initial term running from the Effective Date (defined as the signature date) and auto-renewal unless notice is given at least 90 days before the term ends. Schedule 3 mentions that services commence on 1 May. A model that anchors on the commencement date reports the term as ending on 30 April, with notice due by about 30 January. The term actually ends on 2 March, so notice is due by 2 December. A diary entry built from the AI's answer would arrive two months late and the contract would roll over for another year. Work every notice date from the definitions clause yourself, and write the calculation in the file note.
A worked review: a tour operator's hotel allocation agreement
Here is how this plays out in a two-solicitor commercial firm (an illustration). Their client, a small tour operator, is signing an allocation agreement with a boutique hotel: a block of rooms held for the operator's tours each season, with release dates, cancellation charges and a minimum take-up. It runs to 22 pages plus two schedules. Before AI, a first review took the associate about 110 minutes.
The associate uploads the agreement (but, by mistake, not Schedule 1) to the firm's review tool and asks for an issues list against the firm's standard positions for supply agreements. Sample output (illustrative):
1. Clause 4.2 - Release: unsold rooms released to hotel 21 days
before arrival. [Quote] Flag: shorter than typical 30 days.
2. Clause 7 - Cancellation: 100% charge within 14 days of arrival.
[Quote] Flag: no cap on total cancellation charges.
3. Clause 11.1 - Auto-renewal for further season unless notice given
by 31 October. [Quote] Flag: notice date falls in peak sales period.
4. Clause 13 - Liability cap: hotel's liability limited to 10% of
annual contract value. [Quote] Flag: low.
5. Minimum take-up: "as set out in Schedule 1". Appears standard.Items 1 to 4 are genuinely useful and accurate, and took five minutes. Item 5 is the model guessing about a schedule it never saw. When the associate adds Schedule 1, it reveals a minimum take-up of 85% of the allocation, with shortfall charged at the full room rate. Read together with the 21-day release in clause 4.2, the client could be charged for rooms it has already had to release. That interaction was the most important issue in the contract, and the first pass missed it entirely.
A quick sum shows what it's worth. Say the allocation is 40 rooms for 20 nights, or 800 room-nights, at $140 a night. The minimum take-up is 85%, so 680 room-nights. A slow season in which the operator sells 600 and releases the rest at 21 days, as clause 4.2 requires, leaves an 80 room-night shortfall charged at full rate: $11,200 for rooms the hotel had already taken back and could resell. The fix the associate proposed was simple once the issue was visible: rooms released under clause 4.2 count towards the minimum take-up.
Total associate time with AI: about 70 minutes, including a full read. The saving came from not having to build the issues list from scratch; the value came from the lawyer reading the schedule against clause 4.2.
General assistant or contract-review tool?
A general assistant such as ChatGPT or Claude on a business plan will do a reasonable first pass if you give it a clear prompt and your standard positions. Contract-review tools add structure. Two examples small firms look at:
- Spellbook works as a Microsoft Word add-in, with review and redlining, drafting, playbooks and benchmark comparisons. Pricing is set by the number of team members and quoted, rather than published, and there is a 7-day free trial.
- LegalOn also works in Word, flags issues by severity, and comes with a library of pre-built playbooks for common contract types, plus custom playbooks written in plain English. Pricing is through its sales team.
| If your firm... | Start with |
|---|---|
| Reviews a few contracts a month, of many different types | A general assistant on a business plan, with a good prompt |
| Reviews the same few contract types repeatedly | A legal tool with a playbook for those types |
| Redlines all day in Word | A Word add-in, trialled on real past files |
| Hasn't settled confidentiality rules for AI yet | Neither, until it has |
The broader comparison is in ChatGPT or a legal AI tool for a small firm, and the pricing models in per-seat or pay-as-you-go legal AI pricing. For clients who want to check their own contracts before instructing you, can AI review a contract? explains the limits from the business owner's side.
Test any tool on contracts you've already reviewed
Vendor demos use contracts chosen to look good. Your own past files tell you more. A seeded test takes about two hours:
- Pick five contracts the firm has already reviewed, of the types you see most.
- List the issues your lawyers found in each; that's the answer key.
- Plant three extra issues across the set: a clause reference pointing to the wrong number, a notice period changed in one place only, and an indemnity that escapes the cap.
- Run each tool on the same five contracts with the same instructions.
- Score what it found, what it missed, and what it invented.
A filled-in score sheet from a test like this might look as follows (illustrative):
| Tool | Known issues found (of 31) | Planted issues found (of 3) | Invented or wrong findings | Minutes per contract |
|---|---|---|---|---|
| General assistant, basic prompt | 19 | 1 | 4 | 3 |
| General assistant, evidence prompt below | 24 | 2 | 1 | 4 |
| Legal tool with firm playbook | 27 | 2 | 1 | 2 |
Notice what the numbers say: no tool found everything, and the prompt mattered almost as much as the product. The invented findings count is the one to watch, because each one costs a lawyer time to disprove.
A review prompt that asks for evidence, not opinions
Review the attached contract for our client, the [customer/supplier].
Our standard positions are: [paste 5-10 positions, e.g. liability
cap at 100% of annual fees; mutual indemnities only; 30-day payment].
For each issue:
- quote the clause number and the exact words
- say which standard position it departs from, or why it is unusual
- rate it High / Medium / Low for our client
Rules:
- If something is referred to but not in the document (a schedule,
standard terms, a policy), list it under MISSING DOCUMENTS and do
not guess its contents.
- Check whether any clause limits or overrides another; list these
under INTERACTIONS.
- If a standard clause is absent, write "NOT FOUND", not "standard".
- Do not comment on enforceability.The standard positions are the part most firms leave vague, and a vague list gets vague findings. A filled-in set for supply agreements, where the firm usually acts for the customer, might read:
- Supplier's liability cap: at least 100% of fees paid in the previous 12 months; data-protection breaches and confidentiality outside the cap.
- Indemnities: mutual, limited to third-party IP claims and breach of data-protection obligations.
- Payment: 30 days from receipt of a valid invoice; disputed amounts can be withheld while the dispute is resolved.
- Term and renewal: no auto-renewal longer than 12 months; notice of 60 days or less.
- Price increases: no more than once a year, with at least 60 days' notice and a right to terminate.
- Termination: customer may end for convenience on 90 days' notice after the first year.
- Data: supplier returns or deletes customer data within 30 days of termination, on request.
Run against a supplier's standard terms, that list and the prompt produce something like this illustrative excerpt:
ISSUES
1. Cl 8.1 "Supplier's aggregate liability shall not exceed the fees
paid in the 3 months preceding the claim." Departs from: cap at
12 months' fees. HIGH.
2. Cl 5.4 "Supplier may increase the Fees on 30 days' written notice."
Departs from: annual increase, 60 days, right to terminate. HIGH.
INTERACTIONS
- Cl 8.3 excludes clause 8.1 for "any breach of clause 12". Clause 12
is Confidentiality. No issue: consistent with our position.
MISSING DOCUMENTS
- Cl 2.2 refers to the "Service Level Schedule". Not provided.
NOT FOUND
- Termination for convenience.Two things to fix before this goes near a client letter. Open clause 12 yourself: if it's headed Confidentiality but also contains the data-protection obligations, the INTERACTIONS note is right; if data protection sits in clause 13, the cap still covers it and the "no issue" note is wrong. And ask the other side for the Service Level Schedule before finishing, because service credits there often act as the customer's only remedy for poor service. The quotation rule makes invented findings easy to spot, and the missing-documents section would have caught Schedule 1 in the worked example above. Keep the output alongside the file as a working paper, not as advice, and check any case or statute it mentions against the source, as set out in how to stop AI inventing case law.
Where it fits in a small firm's review routine
- Confirm you have the complete document set, including schedules and anything incorporated by reference.
- Run the AI first pass with the evidence prompt or the tool's playbook.
- Read the whole contract yourself, with the issues list beside it and the client's deal in mind.
- Check every date and number yourself, and every interaction the list raises.
- Write the client's issues list in order of what matters to them, not in clause order.
- Diary renewal and notice dates, which is where tracking contract renewals with AI helps once the contract is signed.
Step 5 is where the AI's output changes shape most. For the tour operator, the first-pass list ran in clause order: release period, cancellation charges, auto-renewal, liability cap, minimum take-up. The associate's letter to the client ran in order of money at risk, in plain words:
"1. The biggest risk: you could pay for rooms you've already handed back. If you release unsold rooms 21 days out, as the contract requires, they still count against your 85% minimum. In a slow season that could cost several thousand dollars. We've proposed that released rooms count towards the minimum. 2. Cancellation charges have no ceiling. 3. The contract renews unless you give notice by 31 October, when you're busiest selling next season; we've diarised a reminder for 1 October."
The liability cap didn't make the letter's top three, because the client's exposure to the hotel's failures was small next to its own commitments. That reordering is judgement about the client's business, which is exactly the part the first pass can't supply.
Run this way, AI review makes the first hour of a contract review shorter and more complete. It doesn't make the lawyer's read optional, and the worked example shows why: the most expensive issue was one only a person reading two parts of the document together would find.
Contract review with AI: questions small firms ask
Can a small firm charge less for contract review done with AI?
It can, but it doesn't have to. What clients pay for is the judgement on what matters in their deal, and that still takes a lawyer's time. Many firms move routine reviews, such as standard supplier terms or NDAs, to a fixed fee and keep the saving as margin. Record the time actually spent if you bill hourly; don't bill for hours AI removed.
Is it safe to upload a client's contract to an AI tool?
On a business plan or a legal tool with a data processing agreement and no training on your content, it's comparable to other outsourced processing. Check the contract itself first: some include confidentiality clauses that restrict sharing with service providers. Don't use free or personal accounts, and avoid free chat-with-PDF websites for any client document.
Does AI contract review work on scanned contracts?
Less reliably. A scanned PDF has to be converted to text first, and errors in that conversion, such as a misread number or a dropped line, pass straight into the review without warning. Where you can, get the editable version from the other side. If you must use a scan, spot-check the converted text of key clauses, especially figures and dates, against the image.
Further reads
- Which Legal Tasks Should a Small Firm Never Hand to AI? — Where AI help should stop in legal work, and why.
- AI for Small Law Firms: What to Automate First — Why admin usually comes before contract review in a small firm.
- How Small Law Firms Use AI to Draft Letters and Routine Documents — Turn the review's issues list into client letters faster.
- Can Solicitors Use ChatGPT Without Breaching Confidentiality? — Confidentiality rules before any client contract goes in.
- AI Software Contracts: Auto-Renewals, Price Rises, Notice Periods — The renewal and price-rise clauses AI flags most often.
- AI Consulting Contracts: 9 Clauses to Check Before You Sign — A clause-by-clause checklist for another contract type.
- How Brokers Use AI to Compare Policy Wordings at Renewal — A two-pass method for renewal wordings: a text compare that can't skip a page, then AI to explain each change, graded on a shared rubric and checked at source.
- How Conveyancers Use AI to Cut Admin on Each Transaction — Stage-by-stage admin savings for a conveyancing file, a title-summary prompt with sample output, chaser wording and the tasks AI must never touch.
- Employment Contracts With AI: Drafting Steps and Legal Checks — Draft a clear employment contract with AI from a job facts sheet, then send the right clauses to a lawyer instead of paying for the whole thing.
- How to Write Terms of Business With AI and What Needs a Lawyer — Map your trading terms clause by clause, draft the operational ones with AI, and brief a lawyer on the risky ones so the review stays short.
- How to Summarise Long Documents With AI Without Missing Details — A six-step method for summarising long documents with AI that keeps the numbers, dates and conditions, shown on an 84-page insurance policy.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Spellbook pricing page (custom pricing by team size, 7-day free trial, features); LegalOn contract review product page (Word add-in, pre-built and custom playbooks, pricing via sales). Worked example and test results are illustrative.