AI Contract Review: What It Can Safely Automate (UK)

AI contract review can safely do the reading and sorting. It can pull out parties, dates and values, compare a contract with your written playbook, flag missing or unusual clauses, and summarise the result for a person to check. It cannot decide whether a risk is acceptable, advise a client or sign off a contract. Those jobs stay with a person, and where a regulated lawyer is involved, they stay with that lawyer.

For standard NDAs and supplier agreements, buying a dedicated tool is usually the right answer, and we show published prices below. Building makes sense when your own precedents, data limits or connected systems are the point. Our intelligent document processing pilot starts from £5,000 for one document type such as contracts. This article covers what to automate, what to keep human, whether ChatGPT will do, buy versus build, the UK guidance, and how to pilot on one contract type.

Key takeaways

• AI is safe for first-pass reading: extracting terms and dates, comparing against a playbook, flagging gaps and summarising. A person checks every output.
• Judgement, advice and sign-off stay with a person. The SRA says a firm stays accountable for AI outputs, and a UK court has said lawyers must check AI-assisted research.
• Before you paste a contract into any AI tool, check where the text goes, who can see it and whether it trains a model.
• For standard NDAs and supplier agreements, a dedicated tool is usually the better buy. Building pays off when your own precedents, data limits or connected systems decide the job.
• Pilot on one contract type, count what the AI misses, and send anything it is unsure about to a named person.

Can AI review a contract for me?

Yes, as a first reader. AI does the reading and sorting, and a person does the deciding. A good tool finds the termination clause, lists the notice period, notes that the liability cap differs from your usual position and points at the text it relied on. If a tool cannot show you where in the document it found something, do not trust the answer.

That split matters because the output of a contract review is a decision, and the decision belongs to someone accountable for it. SpotDraft’s own page says legal professionals remain responsible for legal judgement, negotiation strategy and final decisions (SpotDraft). Adobe’s guide says human oversight remains indispensable for interpreting nuanced legal risks (Adobe). Two vendors that publish guides on AI contract review both say so.

If you have a single contract to check, build nothing. Ask a solicitor, or use a tool under your work account with the checks later in this article. Automation starts to pay when the same kind of contract arrives often. Our guide to where automation pays off explains how to count that.

What AI contract review can safely automate

Five jobs suit AI because the answer sits in the text and a person can check it quickly.

• Extracting key terms and dates. Parties, start and end dates, renewal and notice periods, payment terms and governing law. The output is a row in a register and carries no opinion.
• Comparing against a playbook. A playbook is your written list of positions on each clause, covering what you accept, what you push back on and what you never agree. SpotDraft describes AI review as checking contract language against playbooks, clause libraries and company policies (SpotDraft).
• Flagging missing or unusual clauses. Examples are a missing limitation of liability, an auto-renewal nobody expected or a governing law you do not use. Adobe describes systems that flag clauses that deviate from standard positions, are missing or carry high-risk language (Adobe).
• Summarising. A one-page summary for the person who needs to know what they are signing. Treat it as a draft that points at the source text.
• Building a clause register. Across a backlog of signed contracts, the register shows which renew next quarter, which need consent to assign and which carry a price rise. This suits teams with a pile of old agreements nobody has time to re-read.

Task	What the AI does	Who decides
Extract terms and dates	Reads and lists	A person spot-checks
Compare with playbook	Flags differences	A person
Flag missing clauses	Highlights gaps	A person
Summarise	Drafts	The author
Accept a risk	Nothing	Owner or lawyer
Sign	Nothing	An authorised person

The first four rows are preparation. The last two are decisions, and no tool should make them for you.

The same read, check and route pattern works on other paperwork. We describe it for invoices in accounts payable automation, where anything below a confidence threshold goes to a person rather than into the ledger.

What must stay with a person

Three things stay human: judgement, advice and sign-off. Judgement, because whether a risk is acceptable depends on your commercial position, which the tool does not know. An uncapped liability clause may be fine with a long-standing partner and unacceptable with a new supplier. Advice, because a solicitor who tells a client what a contract means stays accountable for what is said. Sign-off, because signing commits the business, and the authority to sign belongs to named people.

The SRA says the same about accountability. The SRA’s Risk Outlook report on AI, dated 20 November 2023, says a firm remains responsible and accountable for the outputs from AI it uses. It tells firms to supervise AI systems and staff use of them. It also says not to trust an AI system to judge its own accuracy, since current AI “does not have a concept of truth” (SRA). The SRA regulates solicitors and law firms, so its view carries weight for them, and the logic carries over to any business that signs the contract.

The courts have addressed the same failure. In R (Ayinde) v London Borough of Haringey ([2025] EWHC 1383 (Admin), 6 June 2025), the Divisional Court said freely available generative AI tools such as ChatGPT are not capable of reliable legal research. It added that lawyers who use AI for legal research have a “professional duty” to check the output against authoritative sources (National Archives). That case concerned legal research rather than contract review, but the lesson carries over. Check what a tool produces against the source before you rely on it.

Signing is also its own workflow. When we built Gemdarr, an online will platform, the signing and witnessing flow used DocuSign. That is a separate step from reviewing a contract.

Can ChatGPT review contracts?

It can read a contract you paste in and comment on it, and many people use it that way. Whether that is safe depends on three checks. Check where the text goes, whether it trains a model and whether you can verify the answer.

Where the text goes. The same assistant can carry different data terms on different plans, and providers change them. Microsoft’s documentation for its work version of Copilot (formerly Microsoft 365 Copilot), for example, states that prompts, responses and data accessed through Microsoft Graph aren’t used to train foundation LLMs (Microsoft Learn). Other providers publish their own terms. Read the current terms for the exact plan you use before you paste in a client’s contract.

What the vendor keeps. goHeather’s homepage says it does not use customer data to train its models. It also says the AI APIs it uses keep request logs for up to 30 days solely for abuse monitoring (goHeather). The page does not state where data is stored, which a UK buyer would ask about. Statements like these are what you want in writing, including the ones you have to request.

Whether you can verify the answer. Microsoft’s page also says that the responses generative AI produces aren’t guaranteed to be 100% factual. So ask the tool to quote the clause it relies on, then read that clause yourself.

Before you paste a contract into any AI tool, ask four questions.

• Is our text used to train any model, and can we switch that off?
• Where is it processed and stored, and for how long?
• Who else sees it, including AI providers the vendor subcontracts to?
• Can we delete it, and is there a written contract covering how it is handled?

Our practical rule is simple. Anything confidential goes only into a tool your organisation has approved in writing. The last question links to UK GDPR, which we cover below.

Buy a legal AI tool or build your own?

For standard NDAs, MSAs and supplier terms, buy. Dedicated tools already compare contracts against playbooks and explain their findings in plain English. A licence will often cost less than a build for a team with one or two contract types. We build only in the situations below, and we would rather tell you that now.

Route	Published cost	Best when
Dedicated review tool	goHeather: $400 USD a month	Standard contracts, small team
Platform sold by demo	No price on SpotDraft’s page	In-house team, demo first
Our document pilot	From £5,000	One contract type, your rules
Our document processing build	From £10,000	Several types, into your systems

At the time of writing (October 2026), goHeather’s pricing page lists a Starter plan at $400 USD a month, or $200 USD a month billed annually, and a Team plan on custom pricing. Over twelve months that is $4,800 USD billed monthly or $2,400 USD billed annually. That is simple arithmetic on the listed prices, and the page carries no tax note. SpotDraft’s page shows no price and asks you to request a demo. Adobe’s page is a guide to the task, with a free trial and a buy link but no price. These are examples only.

When building is the better answer

Four situations tip the balance.

• Your own precedents are the product. A firm’s clause bank, fallback positions and house style should drive the output in place of a vendor’s defaults.
• The text cannot leave your environment. Client terms or confidentiality rules rule out sending contracts to a third-party service, so processing has to happen where you control it.
• The result has to land somewhere. Terms need to reach your CRM, matter system or finance tools without re-typing. For an international events group we connected Pipedrive, SmartDocs and Xero without replacing them. Signature detection updates the CRM, multi-event agreements split into separate deals and an invoice is raised in Xero on signature. That is sales-contract plumbing rather than legal review, but it shows the pattern.
• The documents are unusual. Scanned or oddly formatted contracts, or several document types at once, are where a purpose-built pipeline can earn its cost.

If none of that describes you, buy a licence.

Our intelligent document processing pilot automates one document type, and the service page names invoices, contracts and forms as examples. Data is extracted, checked and filed where it belongs, accuracy is measured on your real documents, and anything the AI is unsure about goes to a person. It starts from £5,000. A build covering several document types starts from £10,000.

If the job is an agent that collects contracts from an inbox, runs the review and routes exceptions, that is AI agent work, with a first agent from £5,000.

UK rules and guidance that apply

This section is general information, not legal advice. Ask a qualified solicitor about your own contracts and obligations. Three sources are worth reading, and one of them is in flux.

SRA, for solicitors’ firms. Its Risk Outlook report on AI (20 November 2023) says firms adopting AI need to protect confidentiality and legal privilege. It also says firms will still need to check outputs for accuracy, however well controlled the system is (SRA). It dates from 2023, so check the SRA site for anything newer.

ICO, for personal data. Contracts routinely hold names and signatures, and sometimes pay details, so AI review is processing personal data. The ICO publishes guidance on AI and data protection. Its page says the guidance is under review because of the Data (Use and Access) Act and may change, and it shows an update date of 15 March 2023 (ICO). Treat it as a baseline and check the ICO site before you rely on any detail.

UK GDPR Article 28, for vendors. If a supplier handles that personal data for you, you are using a processor. Article 28 says to use only processors providing sufficient guarantees, and that processing by a processor must be governed by a contract (legislation.gov.uk). In practice, get the written terms before you upload anything.

How to pilot AI contract review on one contract type

Pilot on one contract type, with known answers to test against. The aim is evidence about misses. A demo that looks good tells you little.

1. Pick one contract type and one owner. NDAs, supplier agreements or customer terms. One type keeps the playbook short.
2. Write the playbook. For each clause that matters, record the position you want, the fallback you accept and the one you refuse. If the team cannot agree on the playbook, the pilot has already found its first problem.
3. Collect a test set. Use past contracts that a lawyer or senior colleague has already reviewed, so the right answers are known. Include some awkward ones.
4. Run the AI and compare. Count the clauses it missed separately from its false alarms. Misses are silent, so they matter most.
5. Route doubt to a person. Anything the tool is unsure about, and any clause outside the playbook, goes to a named reviewer. Our pilots work this way, with human review wherever the AI is unsure.
6. Run alongside the normal process. Keep reviewing as usual for a few weeks while the tool shadows, then compare the time taken and the misses before you buy, build or stop.

We will not quote an accuracy figure before testing on your documents, because it depends on the contract type and the playbook.

Your next step

Choose one contract type and run the pilot above, on a licence trial or with us. If a licence covers it, buy it and stop there. If your precedents, data limits or connected systems rule that out, a document pilot from £5,000 tests whether a build is worth it on your own contracts. Scope and price are fixed and agreed before work starts, and all code and IP are yours once the work is delivered and paid for.

If you are not sure contract review is worth automating at all, our AI readiness assessment is a one-week programme at £3,000. For law firms and professional services teams, our legal and professional services page shows what else we build for that work.