When your AI assistant
signs a bad deal

By Jonathan Kleiman, Barrister & Solicitor · Published July 2026

Imagine this. Your team starts using a capable AI "assistant" to handle routine emails and draft simple agreements. It is faster, it is cheaper, and nobody has to stay late pushing paperwork over the finish line. Then one Friday night, that assistant sends a supplier a contract — with pricing, volume commitments, or liability language you never intended to offer — and the supplier accepts.

On Monday morning, you realize the mistake. The obvious question is the one that lands on my desk more and more often: are you on the hook?

The honest answer is "it depends" — but not in the reassuring way business owners hope. Ontario law did not need a special "AI statute" to have a view on this. It already has well-worn rules about how contracts form and when someone (or something) can commit your business, and those rules apply to an AI just as they apply to a junior employee. This guide walks through how that analysis works, where the real risk lives, and the practical steps that keep an AI from quietly becoming your de facto signing officer.

Can an AI assistant legally bind my business to a contract in Ontario?

Potentially, yes. Ontario contract law focuses on offer, acceptance, consideration, and an objective intention to create legal relations — not on whether a human or a machine wrote the words — so an AI that sends what looks like a real offer from your business can bind you if the other side reasonably accepts and relies on it. The starting point is that a contract in Ontario forms when the essential elements of a binding contract come together: an offer, acceptance on those terms, consideration, and an intention to be legally bound (plus capacity, certainty, and a lawful purpose).

Crucially, courts judge those elements objectively. The question is not "what did you secretly intend?" but "how would a reasonable person in the other side's position have understood what your business communicated?" This objective approach is the same reason a verbal contract can be enforceable in Ontario even when one party later insists they "didn't really mean it." Outward behaviour wins over private intention.

So if your AI assistant sends an email or a draft that looks like a genuine offer coming from your business — on your letterhead, from your email domain, with your standard signature block — there is a real risk a court will treat it as your offer, provided a reasonable person in the supplier's shoes would read it that way and then rely on it. The fact that "the machine did it" does not automatically save you. Courts look first at outward, objective conduct. If your systems make it appear that your company agreed, you may be treated as having agreed, even where internally you never meant to.

Can I dodge liability by blaming the machine?

Usually not. From a court's perspective an AI is part of your operations — like your email server or your template library — so Ontario law tends to place responsibility on the business that chose, configured, and deployed the tool, not on the tool itself. It is tempting to say "we're not liable; that was just our software." In practice, that argument rarely travels far.

There is a leading Canadian illustration. In Moffatt v. Air Canada (2024), a passenger relied on an airline chatbot that gave him wrong information about bereavement fares. When he sought the refund the chatbot had described, Air Canada argued — remarkably — that its chatbot was a "separate legal entity" responsible for its own actions, and that the airline should not be liable for what the bot said. The British Columbia Civil Resolution Tribunal rejected that outright, holding that the chatbot was simply part of Air Canada's website and that the company was responsible for the information it provided, whether from a static page or an interactive bot. Air Canada was ordered to pay the passenger's damages.

Two cautions on that case, because accuracy matters here. First, it was a British Columbia tribunal decision, so it is persuasive, not binding, in Ontario. Second, it was a negligent misrepresentation case about a customer-facing chatbot, not a case about an AI forming a contract by sending a supplier a bad deal. But the underlying principle is exactly the one an Ontario court would bring to your situation: a business owns the output of the tools it deploys. You chose the AI, configured it, and decided how much freedom to give it. The law tends to put responsibility on the business for the tools it uses, not on the tools themselves.

That does not mean you are always stuck with whatever an AI sends. It means the analysis is nuanced. The more your system looks like an ordinary channel for doing business, the harder it is to distance yourself from what it produces.

Did the AI have "authority" to act for you? (Agency law)

Whether an AI can bind you often turns on agency law: a business is committed by an agent acting within its actual authority (what you expressly or impliedly allowed) or its apparent authority (what your conduct led the other side to reasonably believe the agent could do). Traditionally, an assistant or salesperson binds the business when they act within their actual or apparent authority. With AI, the same questions arise — just in a new form.

Actual authority is the power you genuinely gave the system. Did you configure the AI to send contracts, proposals, or accept changes on your behalf? If you deliberately set it up to finalize deals, you have handed it real authority to do so, and you should expect to be bound by what it sends.

Apparent authority is the trickier one, because you can create it without meaning to. It arises from how things look to a third party. Ask yourself:

• Did you tell counterparties they could rely on documents coming from that system?
• Does your internal practice effectively treat "whatever the AI sends" as valid unless someone catches it?
• Do your customers, suppliers, and partners have any way to tell a "draft" from a "deal" when it comes out of your systems?

If the answer to any of these is yes, then in practical terms you have given that system authority to speak for you. Even if you did not intend that in a strict legal sense, your conduct may create apparent authority in the eyes of a third party — and a counterparty who reasonably relies on that appearance can hold you to the deal. This is the same authority problem that quietly defeats human deals too: a person who signs without the power to bind the company can leave everyone chasing the wrong party. An AI just industrializes the risk.

Where does the real risk live? Your systems and your contracts

Most of the real risk sits in two places: your internal systems (do you let AI content go out the door without human review where it matters?) and your outward-facing contracts (have you clearly communicated who can sign and what is "for discussion only"?). Get those two right and most of the accidental-contract scenarios never get off the ground.

Internally, the question is whether you have set up a process that lets AI content reach a counterparty without human review at the points that count. Externally, the question is whether you have clearly communicated the limits of authority — who can sign, what needs manual approval, and what is genuinely "for discussion only." A few practical points I make to clients:

• If your AI tool can send anything that looks like a binding agreement, assume that at some point it will send something you don't like.
• If your customer, supplier, or partner has no way to distinguish a "draft" from a "deal" coming out of your systems, a court may not be sympathetic when you say "the AI made a mistake."
• If you consistently let an AI handle negotiations without clear disclaimers, you may unintentionally create a pattern a counterparty can point to as evidence of authority.

This is also where the language of your everyday documents earns its keep. A well-drafted service agreement or set of standard terms that names who can bind the business, and requires a signed agreement to close a deal, is a quiet but powerful defence. It is the difference between "the appearance of authority" and "a documented limit on authority the other side was told about."

From my experience: the tool becomes the negotiator by accident

From my experience, nobody ever decides to let software commit their company. It happens by drift. A team adopts an AI tool to save time on routine correspondence, it works well, and the leash gets a little longer each month — first it drafts, then it sends drafts, then it sends drafts that look exactly like the real thing, and eventually there is no meaningful checkpoint between "the AI produced this" and "the counterparty received this as our position."

I think of the pattern the same way I think about the human version I have seen for years: the enthusiastic salesperson who "basically closed" a deal the owner never approved. When that dispute lands, the fight is rarely about the law of offer and acceptance. It is about authority and appearances — what the other side was reasonably entitled to believe, and whether the business's own conduct invited that belief. An AI does not change the doctrine; it just makes the drift faster and harder to see, because there is no person to say "hang on, I should check with the boss."

The clients who avoid trouble are not the ones who ban AI. They are the ones who treat it like a capable junior: it drafts, summarizes, and reminds, and a human makes the commitments. The ones who get hurt are the ones who let it speak with their voice, through their channels, without clear limits — and then discover they are bound by "decisions" they never made.

How do I stop an AI from binding my business by accident?

Reduce the risk through design and policy, not disclaimers buried in fine print: limit what the AI can send without human review, label AI drafts clearly, define who can bind the company, align your vendor contracts, and train your people. The good news is that you can keep the efficiency without accepting unlimited legal risk. Here are the safeguards I recommend putting in place.

1. Limit what the AI can send without review

Configure your tools so they can draft contracts, proposals, and key emails, but require human approval before anything goes to a counterparty. "AI as junior drafter, not autonomous negotiator" is a good mental model. The most important checkpoint is at the boundary of your business — the moment something leaves your systems and reaches a third party who might rely on it.

2. Use clear labels: "Draft — not binding"

Automatically mark AI-generated contracts and emails with prominent language like "Draft — subject to final review and formal execution by [Company]." Make it obvious enough that a reasonable person would understand they are looking at a draft, not a final deal. A clear label is one of the best ways to defeat any later claim of reasonable reliance — but only if your actual conduct matches it. This is the same discipline that separates a binding deal from a non-binding letter of intent or term sheet: say plainly what is a placeholder and what is a commitment.

3. Define who can bind the company

Make sure your standard terms, email footers, and website clearly indicate who has authority to commit the business — by role if not by name — and that all agreements must be signed or approved through specific channels. Naming the signing channel is not bureaucratic box-ticking; it is how you convert a vague "appearance of authority" into a documented limit the other side was on notice of.

4. Align vendor contracts with your risk tolerance

When you hire an AI or automation vendor, scrutinize their default settings and their contract terms before you deploy. Make sure the tools are configured to respect your approval process, and build in obligations that support your compliance and record-keeping. Do not assume you can simply pass a loss up to the vendor — most AI vendor agreements are drafted to limit the vendor's liability heavily, so the protection has to come from the front end.

5. Train your people (and document it)

Make sure your team understands that AI is a tool, not a decision-maker. Give simple examples of what the system may not do on its own — like sending final pricing, or agreeing to changes in a limitation-of-liability clause — without a lawyer or manager looking first. Documenting that training does double duty: it prevents mistakes, and it evidences a standard of care if a dispute ever arises.

My AI already sent a bad contract — what should I do now?

If it has already gone wrong, respond promptly and in a measured way: tell the other side in writing that the document was sent in error, clarify the terms you actually intended, preserve a clean record of the AI's output and your systems, and get tailored legal advice quickly. If you are reading this after something has gone wrong, you are not alone. Move calmly and deliberately.

Act quickly but calmly. As soon as you discover a problematic AI-generated contract or email, acknowledge the issue to the other party and clarify that it was sent in error. Avoid broad, emotional statements; focus on the specific document and what went wrong. Prompt correction also weakens any argument that the other side reasonably relied on the mistake.

Clarify what you did intend. Offer a corrected draft, or a clear description of the terms you actually meant to propose, and give the other party a reasonable opportunity to consider them before anyone takes further steps.

Preserve a clean record. Keep copies of all AI outputs, internal communications, and system logs related to the incident. If the matter escalates, those records will be important in showing what happened and how you responded. This is the same evidence discipline that decides real breach of contract fights.

Get tailored legal advice. Whether the document is ultimately binding depends on the facts — timing, reliance, prior dealings, and the terms themselves. A short, focused consultation can often transform a potential dispute into a manageable conversation. If the other side has already escalated, my guide on how to respond to a demand letter in Ontario is a useful next read.

Is the AI-sent contract actually enforceable? What a court weighs

Enforceability is not automatic either way — a court weighs whether the other party reasonably relied on the document, whether there were red flags that should have prompted questions, and whether you corrected the error quickly. The more your system looks like an ordinary channel for doing business, the harder it is to distance yourself from what it produces. But that cuts both ways, and several fact-specific questions decide the outcome.

Did the other party act reasonably in relying on the document? Reliance has to be reasonable. A sophisticated counterparty who pounced on an obviously mistaken offer is in a weaker position than one who sensibly took a normal-looking deal at face value.

Were there red flags? Odd terms, an unusual format, pricing wildly out of step with your prior dealings, or an explicit "draft" label are all things that should have prompted questions. Where a reasonable person would have paused, a court is less likely to reward reliance.

Did you correct the error quickly, or sit on it? Prompt, clear correction helps you; silence and delay hurt you, because they let the other side's reliance harden. If a genuine dispute develops from here, the practical path forward looks a lot like any other contract fight — see how to enforce a contract in Ontario for the mechanics, and remember that a contract dispute lawyer can tell you early whether you are likely bound at all.

Treat AI as a powerful junior, not as a partner

The underlying theme is simple: AI can be an excellent junior assistant, but it should not quietly become your de facto negotiator or signing officer. If you treat it as a tool — one that drafts, summarizes, and reminds you, while a human makes the key commitments — you can capture most of the upside while keeping the risk where you can see and manage it. If instead you let it speak with your voice, using your channels, without clear limits, you may find yourself bound by "decisions" you never made.

For many owner-managers, this is the moment to pause, audit how AI is being used in the business, and put a simple governance framework in place before something goes wrong.

Key takeaways

• An AI can bind your business. Ontario law judges contracts by outward, objective behaviour — not by whether a human wrote the words — so an AI that sends what looks like a real offer from your business can create a binding deal.
• "Blame the machine" is weak. Courts treat an AI as one of your tools; you chose and configured it, so responsibility tends to land on the business, not the software (the principle behind Moffatt v. Air Canada).
• Authority is the fault line. Actual authority is what you gave the system; apparent authority is what your conduct led the other side to reasonably believe — and you can create it without meaning to.
• Design beats disclaimers. Require human approval before anything reaches a counterparty, label AI drafts clearly, name who can sign, align vendor contracts, and train your team.
• If it goes wrong, move fast. Correct the error in writing, clarify what you intended, preserve the records, and get advice early — prompt correction undercuts the other side's reliance.

Frequently asked questions

Can an AI assistant legally bind my business to a contract in Ontario?

Potentially, yes. Ontario contract law asks whether there was an offer, an acceptance on those terms, consideration, and an intention to create legal relations — and it judges these by outward, objective behaviour, not by what you privately intended. If your AI sends something that a reasonable person in the other side's position would read as a genuine offer from your business — on your domain, with your signature block — and they accept and rely on it, a court could treat it as your contract. The fact that "a machine wrote it" does not, on its own, undo the deal. Whether you are actually bound depends on the specific facts: how the document looked, whether the other side reasonably relied on it, and whether you gave the AI apparent authority to act for you.

Can I avoid liability by saying "the AI made a mistake, not me"?

Usually not. From a court's perspective the AI is one of your business tools — like your email server or your template library. You chose it, configured it, and decided how much freedom to give it, so the law tends to put responsibility on the business, not on the software. In Moffatt v. Air Canada (2024), the airline argued its chatbot was a "separate legal entity" responsible for its own statements; a British Columbia tribunal flatly rejected that, holding the company responsible for information its own system produced. That decision is not binding in Ontario, but the underlying principle — you own your tools' output — is exactly how Ontario courts approach agency and responsibility. "Blame the machine" is a weak position.

What is "apparent authority" and how does it apply to AI?

Apparent (or "ostensible") authority is when your conduct leads a third party to reasonably believe someone — or something — is authorized to act for you, even if you never formally granted that power. The classic example is a salesperson who looks authorized to close deals. With AI, the same logic applies: if you let an AI send proposals, respond to counterparties, and finalize terms on your channels, and you tell or allow others to treat what it sends as your business speaking, you may have created apparent authority. A counterparty who reasonably relies on that appearance can hold you to the deal, even if internally you never meant to let the AI commit you.

Does the other side have to act reasonably to hold me to an AI-sent contract?

Yes — reasonableness runs both ways, and it is often where these disputes turn. A court looks at whether a reasonable person in the counterparty's shoes would genuinely have believed this was a real, authorized deal, and whether they actually relied on it. If the AI's output had obvious red flags — bizarre pricing, an unusual format, terms wildly out of step with your prior dealings, or an explicit "draft only" label — the other side may not be able to claim reasonable reliance. The messier your systems make it to tell a draft from a done deal, the more sympathetic a court will be to the counterparty. Clear labels and consistent process protect you precisely because they defeat "reasonable" reliance on a mistake.

How do I stop my AI tools from binding my business by accident?

Design and policy, not fine print. The most effective safeguard is to require human approval before anything an AI drafts goes to a counterparty — treat the AI as a junior drafter, not an autonomous negotiator. Automatically label AI-generated drafts with prominent language like "Draft — subject to final review and formal execution by [Company]." Make clear, in your standard terms and email footers, who actually has authority to bind the business and through what channel (for example, a signed agreement). Configure your tools so they cannot send final pricing or agree to changes in key clauses without a human. And train your team, in writing, on what the AI may and may not do on its own.

My AI already sent a bad contract and the other side accepted. What do I do?

Act quickly but calmly. As soon as you spot the problem, tell the other side, in writing, that the document was sent in error and does not reflect terms your business authorized — focus narrowly on the specific document rather than making sweeping statements. Then clarify what you did intend, ideally with a corrected draft, and give them a fair opportunity to consider it. Preserve everything: the AI's output, internal messages, and system logs, since timing and reliance will matter if it escalates. Whether the deal is ultimately binding depends on the facts, so get tailored legal advice early — a short, focused review often turns a potential dispute into a manageable conversation. Prompt correction also weakens the other side's reliance argument.

Is a "Draft — not binding" label enough to protect me?

It helps a lot, but it is not a magic shield. A clear, prominent label — "Draft, subject to final review and formal execution by [Company]" — is strong evidence that a reasonable person should not have treated the document as a done deal, and it directly undercuts any claim of reasonable reliance. But a label buried in tiny print, or contradicted by how you actually behave (for example, if you routinely let "drafts" turn into real deals without anyone signing), carries less weight. Labels work best as part of a consistent system: AI drafts are labelled, real commitments go out only through a defined signing channel, and your conduct matches what your labels say.

Who is liable if my AI vendor's tool caused the problem — me or the vendor?

As between you and the counterparty who relied on the document, the business that sent it is generally the one on the hook — the third party dealt with you, not your software vendor. Whether you can then recover from the vendor is a separate question governed by your contract with them: their limitation-of-liability clause, warranties, indemnities, and configuration obligations. Most AI and automation vendor agreements are drafted to limit the vendor's exposure heavily, so do not assume you can simply pass the loss up the chain. The practical fix is front-end: scrutinize the vendor's default settings and contract terms before you deploy, and build in obligations that respect your approval process.

Does it matter that an AI, and not a human employee, sent the contract?

Legally, less than people expect. Ontario contract law focuses on offer, acceptance, and objective intention — it does not require that a human typed the words. And agency principles were built to handle exactly this: a business acting through an agent. Courts have historically extended those principles to new tools and technologies rather than letting businesses escape responsibility by pointing at the tool. So the "it was AI, not a person" argument tends to change the framing more than the outcome. What matters is how the document appeared, whether the other side reasonably relied, and whether your conduct gave the system authority to speak for you.

Should my business have an AI use policy before something goes wrong?

Yes — for most owner-managers this is the moment to put a simple governance framework in place before there is a problem. A short internal policy that says what AI tools may draft, what always needs human sign-off, who can bind the company, and how AI drafts must be labelled does two things: it prevents the accidental-contract scenario in the first place, and it gives you a documented standard of care if a dispute ever arises. You do not need a fifty-page policy. You need clear rules, matching configuration in your tools, and a record that your team was trained on them. A business lawyer can help you build one that fits how you actually operate.

Final thoughts

"My AI assistant signed a bad contract — who's liable?" is a genuinely new question with a surprisingly old answer. Ontario did not need a bespoke AI law to handle it, because the doctrines that decide the case — objective contract formation, actual and apparent authority, and the principle that a business owns the output of its tools — were built for exactly this kind of problem. The technology is new; the framework is not.

The parts most likely to hurt you are also the parts you control: whether your systems let AI content reach counterparties without a human checkpoint, whether anyone outside your business can tell a draft from a deal, and whether your conduct quietly hands the machine authority to speak for you. Fix those, and the scary Friday-night scenario mostly disappears.

If you want to audit how AI is being used in your business, tighten up your standard terms, or you are staring at a deal an AI sent and are not sure it is one, a contract lawyer in Toronto can pressure-test your setup before it costs you — and a business lawyer more generally can help you put a sensible governance framework in place. If a counterparty is already pushing to hold you to the deal, a commercial litigation lawyer can tell you where you stand. Call 416-554-1639 or book a free consultation, and we can map it out in a short conversation.