The thing to understand about AI is that it wants to make you happy. Not in a sinister way — it was trained to. It is an engine tuned to give you a satisfying answer to whatever you type, and it will do that even when the honest answer is “you’re asking the wrong question.” It never crosses its arms. It never says “wait, why do you want that?” It just produces — confidently, in clean paragraphs, in about four seconds.
For most tasks, an eager assistant who never pushes back is a gift. For a legal agreement, it is precisely the wrong instrument, because the value of a good agreement lives entirely in the pushing back.
AI is the ultimate pleaser
Roughly 63% of mid-sized law firms have now formally adopted generative AI, and business owners are using the same tools directly to draft their own contracts. So this is not a hypothetical. People are asking AI to “write me an employment agreement with a non-compete” or “draft a shareholder agreement for a 50/50 company,” and AI is cheerfully obliging.
The problem isn’t that the output looks bad. It looks great. The problem is that it looks equally great whether it’s right or catastrophically wrong, because fluency and accuracy are two different things and AI has mastered only the first. Stanford researchers testing the purpose-built legal-research tools — the good ones, made for lawyers — found they still hallucinate on more than one in six queries. General-purpose chatbots are far worse. In a British Columbia case, Zhang v. Chen, 2024 BCSC 285, a lawyer filed court materials citing two cases that ChatGPT had simply invented; she was ordered to pay the other side’s costs, and the judge noted studies putting the hallucination rate of some models at 69 to 88 percent on legal questions. She is not alone — one running database has catalogued more than 1,600 filings worldwide caught containing fabricated AI content.
Read that again: the tool will invent a court case, present it in the correct citation format, and defend it if you ask. It does not know it is wrong, and it will never volunteer that it might be. An anxious junior lawyer at least knows the feeling of being unsure. AI feels nothing, and reports certainty either way.
Three times the answer wasn’t the question
The confident-wrong problem is the dramatic one. The quieter, more common failure is worse for a business owner, because it produces an agreement that is technically fine and strategically useless. Here is what that looks like.
You ask for a non-compete. AI writes you a strong one — broad territory, long duration, the works. What it won’t tell you is that in Ontario, non-competes in ordinary employment contracts have been largely banned since 2021 under the Working for Workers Act, with only narrow exceptions. So the clause you asked for is, in most cases, worth nothing. What you needed was a carefully drafted non-solicitation and confidentiality clause that a court will actually enforce. AI gave you the X you named. A lawyer gives you the Y that works.
You ask for a shareholder agreement, 50/50. AI produces a clean, balanced, symmetrical document. Beautiful. It also has no deadlock mechanism, because you didn’t ask for one, and a 50/50 company with no tie-breaker is the single most reliable way to build a partnership that ends in a stalemate. What you needed — and didn’t know to request — was a shotgun buy-sell clause, or a casting vote, or a defined path out. The agreement AI wrote is the trap, drafted flawlessly.
You ask for “a quick NDA” before a meeting. AI gives you a mutual NDA in thirty seconds. Fine. But you were about to hand a competitor your customer list and your pricing, and what you actually needed was a one-way agreement with a real survival period, a non-solicit, and honestly a reason to slow the whole thing down until there was a term sheet on the table. The NDA wasn’t the deliverable. The judgment about whether to be in the room at all was the deliverable.
What a business lawyer actually sells
It is not typing. It never was. When you hire an experienced business lawyer, the drafting is the last five percent; the value is everything that happens before a word gets written. It’s the questions — what are you actually trying to do, who’s on the other side, what happens if this goes well, what happens if it goes badly, what did the last three clients in your position wish they’d done differently. It’s discernment: the ability to look at the X you requested and see the Y underneath it. AI has no clients, no scar tissue, and no stake in whether you’re still standing in three years. A lawyer who has watched a hundred deals has all three.
| What you ask for | What AI hands you | What a lawyer helps you see (the Y) |
|---|---|---|
| “A non-compete for my employee” | A broad, confident non-compete clause | It’s largely unenforceable in Ontario employment — you need a non-solicit and confidentiality clause that holds up |
| “A 50/50 shareholder agreement” | A clean, symmetrical 50/50 document | With no deadlock mechanism, 50/50 is how partnerships die — you need a tie-breaker or exit path |
| “A quick NDA” | A generic mutual NDA | You’re the one disclosing — you need a one-way NDA, a non-solicit, and maybe to pause the deal |
| “Add an indemnity so I’m protected” | A mutual indemnity clause | Mutual cuts both ways — a liability cap may protect you far more than an indemnity does |
| “Make it enforceable” | Formal, legalese-heavy prose | Enforceable is about substance, not tone — and it depends on which country’s law is bleeding into your template |
Notice the pattern in the last column. In every row, the useful move is a question AI was never going to ask, because you didn’t prompt it and it has no reason of its own to wonder.
So where does AI actually earn its keep?
Plenty of places — and pretending otherwise would be its own kind of dishonesty. AI is genuinely good at first drafts you intend to tear apart, at summarizing a forty-page agreement into something you can read over coffee, at translating legalese into plain English, at building a checklist, at flagging that a clause you’d expect is missing. Used as a fast, tireless junior whose work you always check, it makes a good lawyer faster and a diligent owner better informed. That is a real gain, and I use these tools too.
The mistake is treating the draft as the destination rather than the starting point. If you’ve already generated an agreement with AI, the smart next step isn’t to sign it — it’s to have a lawyer review the AI-drafted document for the things it can’t see: the enforceability, the missing protections, and above all whether it’s solving the problem you actually have. Just be careful what you feed it in the first place, because pasting a live contract into a public AI tool can quietly waive confidentiality. And if you want a working lawyer’s longer view on all of this, I’ve written separately about what AI actually changes for law firms.
The pleaser will always give you an answer. Whether it’s the answer to your actual problem is a judgment call — and judgment, so far, is the one thing nobody has figured out how to automate. When AI finally learns to look up from the request and ask “what are you really trying to do here?”, I’ll worry. Until then, that question is the job.
Common questions
Can I use AI to draft a business contract?
You can, and the draft will look professional. The risk isn’t typos — it’s that AI produces exactly what you ask for without judging whether it’s what you need. It won’t tell you a clause is unenforceable, that you’re missing a protection, or that you’re solving the wrong problem. Treat an AI draft as a starting point to be reviewed, never a finished agreement to sign.
What is AI actually good at in legal work?
First drafts you plan to revise, summarizing long documents, explaining legalese in plain language, building checklists, and flagging an obviously missing clause. Used as a fast junior whose work you always verify, it genuinely helps. It fails when it’s trusted to exercise judgment — which is the part that matters most in an agreement.
What can a business lawyer do that AI can’t?
Ask the right questions before drafting, understand your business objectives, and spot when what you asked for isn’t what you need. That’s discernment, interpretation, and judgment built from years of watching deals succeed and fail. AI has no clients, no experience, and no stake in whether you’re protected in three years. A lawyer has all three.
Is an AI-drafted agreement legally binding in Canada?
It can be, if the essentials of a contract are present. But “binding” and “good for you” are different questions. An AI draft can be fully enforceable and still fail to protect you, or quietly import U.S. law that doesn’t apply here. See our fuller answer on whether an AI-drafted contract is legally binding in Canada.
Is it safe to paste my contract into ChatGPT?
Be careful. Pasting a live or confidential contract into a public AI tool can waive confidentiality, breach an NDA, or expose privileged information. Before you paste anything, understand what you may be giving away — we cover this in detail in our piece on what you give away pasting contracts into ChatGPT.
Does AI ever get the law wrong?
Frequently, and confidently. Stanford found purpose-built legal AI tools hallucinate on more than one in six queries; general chatbots are worse. In Zhang v. Chen, 2024 BCSC 285, a lawyer was ordered to pay costs after filing cases ChatGPT had invented. AI states fabrications in the same confident tone as facts, and never signals which is which.
Won’t AI just make lawyers unnecessary?
It makes routine drafting faster, which is good for everyone. But the core of the work — judgment about what you actually need, and responsibility for getting it right — is exactly what AI can’t do and can’t be accountable for. The likely future is lawyers who use AI well, not the absence of lawyers.
What’s an example of asking for X when you needed Y?
You ask AI for an employee non-compete; it writes a strong one. But in Ontario, non-competes in ordinary employment have been largely banned since 2021, so the clause is usually worthless — what you needed was an enforceable non-solicit and confidentiality clause. AI gave you the X you named. A lawyer gives you the Y that works.