How to check an AI-drafted contract before it goes out
To check an AI-drafted contract, treat it like a junior's first draft that arrived without a note: confirm which precedent or template it started from, read every operative clause against the client's instructions and your agreed positions, check defined terms and cross-references, verify any law it relies on, and have a qualified lawyer sign it off before it goes out. The checking is quicker when the AI drafts from your firm's own precedents, because you are reviewing language you already trust.
Most law firms and in-house legal teams now have someone producing first drafts with AI, whether through a dedicated tool such as Harvey, Legora, Spellbook or Scriptum AI, or through Microsoft Copilot and general chatbots. The draft arrives faster. The question is what a careful lawyer should do with it before it reaches the client or the other side.
How do you check an AI-drafted contract?
Work through it in a fixed order, starting with where the text came from and ending with sign-off. A consistent routine matters more than any single check, because it stops a polished-looking draft from being waved through.
- Confirm the source. Which precedent, template or prior deal did the draft start from, and is it the current approved version?
- Check the instructions. Does the draft reflect the term sheet, heads of terms or the business's request, including the parties, price, dates and conditions?
- Read the operative clauses. Warranties, indemnities, limitation of liability, termination and payment terms should match your agreed positions, not a generic market view.
- Check defined terms. Every capitalised term should be defined once, used consistently and actually needed.
- Check cross-references and schedules. Clause numbers, schedule references and annexes should all point to the right place.
- Verify the law. Check any legislation, regulator or case the draft names, and make sure governing law and jurisdiction wording suits the deal (England and Wales, Northern Ireland, Ireland or elsewhere).
- Look for what is missing. AI is better at writing what is there than noticing what should be. Compare against your checklist for the document type.
- Check boilerplate against the deal. Notices, assignment, confidentiality and data protection clauses are easy to accept without reading.
- Compare with the precedent. Run a comparison against the starting document so every AI change is visible as a tracked change.
- Sign off and record it. A qualified lawyer approves the draft, and any new agreed position is captured for next time.
What mistakes do AI-drafted contracts usually contain?
The errors tend to be confident rather than obvious. The table below sets out the common ones and where to look.
| Typical issue | Why it happens | Where to check |
|---|---|---|
| Positions that do not match instructions | The tool fills gaps with a generic market position | Liability, indemnities, warranties, payment |
| Inconsistent defined terms | Text is assembled from more than one source | Definitions clause and every capitalised term |
| Broken cross-references | Clauses are added, moved or renumbered | Clause and schedule references |
| Wrong law or jurisdiction wording | Training data leans towards other jurisdictions | Governing law, statutory references, regulator names |
| Invented or outdated authority | General models can produce plausible but false references | Any named case, statute or guidance |
A general picture of current tools, not a claim about any one product.
The last row is not theoretical. In Ayinde v London Borough of Haringey [2025] EWHC 1383 (Admin), the Divisional Court dealt with legal documents that cited cases which did not exist, and made clear that lawyers remain responsible for checking material they put forward, however it was produced. That was litigation, but the lesson carries over to transactional drafting. Our post on managing hallucination risk in legal AI goes further.
Who is responsible for an AI-drafted contract?
The lawyer who sends it, and the solicitor supervising them. The SRA Code of Conduct for Solicitors says that where you supervise others, you remain accountable for the work carried out through them (paragraph 3.5).
The Law Society's guidance on generative AI makes the same practical point for solicitors and firms, particularly small and medium-sized ones: define what the tool is for, follow your professional obligations, and keep a human check on what it produces. We cover the regulatory side in the SRA professional duties and using AI.
A clean, well-formatted draft is not the same as a checked one.
Does the type of AI tool change how much checking you need?
Yes. A draft built from your own approved precedents needs less checking than one written from a general model's idea of a typical contract, because most of the language is already yours.
With a general chatbot, every clause is new text and has to be read from first principles. With a tool that starts from the firm's precedent bank and shows its changes as tracked changes in Word, the reviewer can focus on what changed and why. That is the difference we describe in legal AI vs ChatGPT for contract drafting.
How should in-house legal teams check AI drafts?
The checks are the same, but in-house teams should decide in advance where the lawyer's review effort goes. A small team cannot read every AI-generated NDA in full while a backlog of supply agreements waits.
A sensible approach is to agree which contract types the business may generate from approved templates with light-touch review, and which always need a lawyer to work through the full checklist. Recording the positions agreed on each contract then makes the next review faster.
Where does Scriptum AI fit?
Scriptum AI is designed to make the checking step easier. It is a Microsoft Word add-in that drafts from your firm's own precedents, with partner approval before anything is released, and every clause your lawyers agree is captured into a clause library the firm owns, so the next draft starts from positions that have already been checked. Scriptum is built for transactional work at mid-sized firms with between 5 and 500 fee earners in the UK, Ireland, the USA and Canada, and for in-house teams doing the same kind of contract work. The Draft product page shows how it works in Word, and current prices are on our pricing page.
Want to see how much quicker review is when the draft starts from your own precedents? Book a demo and we will draft from one of your templates in Word.
Sources
- SRA: Code of Conduct for Solicitors, RELs and RFLs (paragraph 3.5)
- The Law Society: Generative AI – the essentials
- Courts and Tribunals Judiciary: Ayinde v London Borough of Haringey [2025] EWHC 1383 (Admin)
Common questions
How do you check an AI-drafted contract?
Check it as you would a junior's draft, but with extra attention to where the text came from. Confirm the starting precedent, read every operative clause against the client's instructions, check defined terms and cross-references, verify any law or authority the draft relies on, and have a qualified lawyer sign it off before it leaves the firm.
Who is responsible if an AI-drafted contract contains a mistake?
The lawyer and the firm, not the software. Under the SRA Code of Conduct, solicitors remain accountable for work carried out through people they supervise, and using an AI tool does not move that responsibility to the vendor.
Is it quicker to check an AI draft than to draft from scratch?
Usually, if the AI drafts from the firm's own approved precedents, because the reviewer is checking known language rather than unfamiliar text. A draft from a general-purpose chatbot can take longer to check, because every clause has to be verified from first principles.
What mistakes do AI-drafted contracts typically contain?
The common ones are positions that do not match the client's instructions, defined terms used inconsistently, broken cross-references, wrong governing law or jurisdiction wording, and references to legislation or cases that are out of date or do not exist.
Should in-house legal teams check AI-drafted contracts differently?
The checks are the same, but in-house teams should agree in advance which contract types a business user may generate from approved templates and which always need a lawyer's review, so that review effort goes on the higher-risk agreements.