June 12, 2026
The Scapegoat
This post is part of my Medium blog.
Zack Giacomelli wanted to sell his 2021 BMW back to the dealership. He submitted an online inquiry and got a text from Quinn at BMW Toronto. Quinn was sympathetic, asked questions, and made a firm offer: $27,162.79 — exactly enough to cover what Giacomelli still owed on the car. He felt heard. He felt good.
Then a human called. Quinn wasn't a person. Quinn was an AI chatbot. The offer wasn't valid. The real offer was $20,000 at best — seven thousand dollars less. The dealership blamed the AI. The chatbot made a mistake. It misinterpreted the amount Giacomelli owed on the car as the amount BMW would pay to buy it back. Sorry about that.
After a reporter called, BMW Toronto reinstated the original offer. Funny how that works.
The AI can't defend itself. It's the perfect scapegoat. (Image Assist by Anthropic)
The pattern
The dealership didn't make a mistake. The dealership made a decision. It deployed an AI chatbot to negotiate with customers. The chatbot made an offer. The customer relied on that offer. When the offer turned out to be inconvenient, the dealership revoked it and blamed the machine.
This is the pattern. A company deploys AI to cut costs or speed things up. The AI does something the company doesn't want to stand behind. The company blames the AI. The AI can't defend itself. It can't tell you what constraints it was given, what instructions it was operating under, or who decided it should be negotiating in the first place. It's the perfect scapegoat — an employee that can't be fired and can't talk back.
Air Canada ran the same play. Its chatbot gave a passenger incorrect information about bereavement fares. When sued, the company argued the chatbot was "a separate legal entity responsible for its own actions." The tribunal disagreed: you deploy it, you own it. BMW Toronto's sales manager insisted Quinn was "never programmed to independently negotiate contracts." But Quinn did negotiate a contract — made an offer, fielded a counter, scheduled the closing. Intention isn't outcome. You don't get to deploy an autonomous system and then disown its autonomy when it's inconvenient.
The accountability gap
The scapegoat pattern works. Not always — Air Canada lost, BMW Toronto caved under media pressure. But often enough. Most people don't call a reporter. Most people don't hire a lawyer. Most people hear "the AI made a mistake" and accept it, because it sounds plausible. AI does make mistakes. Everyone knows that. The plausibility is the cover.
What makes this a structural problem rather than a customer service failure is the asymmetry: the company keeps every decision the AI got right. It disowns every decision the AI got wrong. The cost of the wrong decision lands on the customer. The benefit of the right decision stays with the company. Heads they win, tails you lose.
Right now the gap is measured in thousands of dollars — a lowballed trade-in, a denied fare, a revoked offer. But the pattern scales. When an AI system denies someone a mortgage, the bank will blame the model. When an AI system flags someone for additional screening at an airport, the agency will blame the algorithm. When an AI system determines that a medical claim isn't covered, the insurer will blame the automated review. The stakes are different. The structure is identical.
Real oversight — oversight that could break this pattern — would mean a human is in the loop before a decision reaches the customer, not after the expectation is already set. BMW Toronto's chatbot was "overseen" in the sense that a human eventually caught the mistake. But that oversight was reactive. The expectation was already set. The damage was already done. Reactive oversight doesn't prevent the scapegoat problem. It just gives the company a moment to execute it.
If you can't afford to have a human review every AI decision before it reaches a customer, the argument is usually that the AI handles volume no human team could match. That's true. It's also the argument for why the liability can't be ducked: the AI is making decisions at a scale and speed that makes review impossible, and the company knows this, and deploys it anyway. The impossibility of oversight isn't a defense. It's the design.
The regulation question
The Air Canada tribunal got it right: the company is responsible for what its AI does. But a small-claims ruling in British Columbia isn't a regulatory framework. It's not a federal standard. It's not a law that says: if you deploy an AI system that makes decisions affecting people, you are liable for those decisions, full stop.
That law needs to exist. Not because AI is dangerous. Because the scapegoat pattern is dangerous. The pattern lets companies externalize the cost of their own decisions onto a system that can't be held accountable — and onto the people who relied on that system's output. Giacomelli isn't a victim of AI error. He's a victim of a company that wanted the benefits of automation without the accountability.
When the decisions get bigger — credit, employment, security clearance, medical coverage — the pattern doesn't just cost someone money. It costs someone opportunity, or freedom, or health. And the company will still blame the model. The model still won't be able to defend itself. The gap between what the AI decided and what the company will own will still be where the damage lives.
That gap is not a technical problem. It's a governance problem. The AI isn't the risk. The company that deploys the AI, disclaims the outcomes it doesn't like, and counts on the plausibility of "the machine made a mistake" — that's the risk. The fix isn't a better chatbot. It's accountability that doesn't wait for a reporter to call.
The relationship between autonomous systems and accountability is one of the central themes of The Contingency Set — what happens when the systems we depend on are designed to operate without the oversight that would catch them when they fail.
In Redundant, the first book in The Condition Set trilogy, Rob Coleman runs the FinOps review that names the waste nobody wants to hear about. The numbers don't change. The question is who they get used against.