Contract Templates From AI: Where It Is Safe

OpenAI announced Astra for Law on September 17, a legal configuration built on an index of more than 230 million URLs of United States law. On its own benchmark testing it passed the overall correctness check on 54% of questions, against 38.7% for the general model using web search. That is real progress and it is the state of the art, which is a useful thing to hold in mind before asking a general chatbot to write your client agreement.

Three tiers, sorted by consequence

The middle tier is where the money is. A service agreement you sign forty times a year is worth paying a lawyer to get right once, and AI can cut that engagement substantially by arriving with the commercial terms already drafted in plain language. The lawyer then spends their time on the parts that need judgment rather than on transcribing what you want.

The bottom tier is the one where people talk themselves into trouble, usually because the document reads well. Fluency and enforceability are unrelated properties.

What AI does genuinely well here

Explaining a contract you have been sent. Pasting a supplier agreement in and asking what each clause means, what is unusual, and what to negotiate is one of the highest-value uses available to a small business. You are not relying on it to be right about the law, only to help you read.

Checking your draft against a list. Ask what a service agreement in your industry normally covers, then check your existing template against that list. The gaps it finds are worth taking to a lawyer.

Turning legalese into something your team can follow. A one-page internal summary of what your own terms commit you to, so the people quoting jobs know what they are promising.

Preparing the brief. What you want the agreement to achieve, what has gone wrong before, and what questions a lawyer would ask. Arriving with that shortens the meeting and the bill.

The jurisdiction problem

Most legal training material, and most legal AI tooling, is oriented to United States law. OpenAI's new index is explicitly US sources. A Canadian business asking a general assistant for a non-compete clause can get something fluent, confident and shaped by a body of law that does not apply.

Canadian contract law differs by province. Employment standards are provincial. Consumer protection rules differ. Limitation periods vary. A clause that is standard in one place can be unenforceable in another, and an unenforceable restrictive covenant can leave you in a worse position than having written nothing, because you relied on protection you did not have.

State your province in the prompt every time. Treat that as a partial mitigation rather than a fix, because the model will answer with confidence either way and has no reliable way to tell you when it is out of its depth.

The missing clause is the dangerous one

Ask a model to review a contract and it assesses what is in front of it. Wrong text is visible, and a careful reader will often catch it.

An absent protection is invisible. A seven-page agreement that never addresses what happens if the client becomes insolvent mid-project, or who owns the work product, or how either side exits, reads exactly like a complete one. You find out in the week you needed it.

The partial defence is to ask the negative question explicitly: what would a lawyer say is missing from this, and what scenarios does it fail to address. That surfaces more than asking whether the document is any good, and it is the same technique that makes AI contract review useful rather than reassuring.

What 54% should calibrate

The number deserves a fair reading. It is OpenAI's own figure, on a benchmark measuring legal research rather than drafting, at the highest reasoning setting, on US law. It represents a purpose-built system with a dedicated index beating a general model by a wide relative margin, and publishing it rather than burying it is to the company's credit.

It also means the best available configuration, built by the largest lab, aimed at large firms, gets roughly half of a professional research task right. Whatever a general assistant does with your employment agreement sits well below that line. We looked at the wider implications in the best legal research AI gets 54% right.

None of this is legal advice, and nothing here substitutes for a lawyer who knows your province and your business. The realistic position is that AI has made preparing for legal work much cheaper and has not made the legal work optional. Accountability for what you sign stays with you either way, which is the point of who is accountable when AI is wrong.

Frequently Asked Questions