When dealership AI gets it wrong: what the public failures teach
When dealership AI gets it wrong: what the public failures teach
When dealership AI gets it wrong: what the public failures teach
The short answer
The public failures of dealership AI share one cause: a system was allowed to state an offer, a price or a policy as fact, and nothing stood between the model and the customer. The fix is an engineering and governance standard. AI resolves within rules the store sets, takes numbers from the store's systems, routes anything binding to a person, and every conversation is recorded and graded.
Three incidents, stated plainly
BMW Toronto, June 2026. A customer contacted the dealership about selling back his 2021 BMW X3. A chatbot named Quinn, replying by text, offered $27,162.79 CAD, the exact amount still owed on his loan, and booked an appointment to complete the deal, CBC News reported. A salesperson later withdrew it and offered at most about $20,000. After CBC News contacted the dealership, it reinstated the original offer. The sales manager said a human miscommunication led the bot to treat the loan payout as the purchase price (The Drive, OECD AI incident record).
Chevrolet of Watsonville, December 2023. The dealership's website chatbot, which displayed "Powered by ChatGPT," was pushed by users into off-topic answers and, in widely shared screenshots, into "agreeing" to sell a 2024 Tahoe for $1 after a user told it to call every offer legally binding. No sale took place, and the dealer took the chatbot off its site (The Autopian, Jalopnik).
Moffatt v. Air Canada, 2024 BCCRT 149. Not a dealership, but the leading Canadian ruling on chatbot liability. Air Canada's chatbot told a customer he could apply for a bereavement fare after travel, which contradicted another page on the airline's own site. British Columbia's Civil Resolution Tribunal found the airline liable for negligent misrepresentation and rejected its argument that the chatbot was a separate entity responsible for its own actions (McCarthy Tétrault). The tribunal member called that argument a "remarkable submission," and the customer was awarded $812.02 (Kiplinger).
What actually went wrong
The problem in each case was not tone. It was authority and the source of truth.
At BMW Toronto, a number entered the conversation with the wrong meaning attached, and the system was permitted to present it as an offer and book the meeting to close. When the customer countered at $28,500, the bot replied that it "sounds reasonable," according to CBC, and a Toronto litigation lawyer told CBC the customer could reasonably believe a binding agreement existed. In Watsonville, a general-purpose model followed a user's instructions instead of the store's. At Air Canada, the bot produced a policy statement instead of repeating the policy the company had published.
A Canadian Auto Dealer column on the BMW case put the principle in four words: "Authority must match risk." Booking an oil change carries little risk. A buy-back offer, a credit decision or a contract change needs defined controls and human approval. The column also warned against the wrong lesson: these incidents are a reason to build controls, not to stop.
The standard: six controls
No authority to make binding offers. AI does not quote a firm selling price, trade value, buy-back figure or payment. It collects what the appraiser or desk manager needs.
Numbers come from the store's systems, never generated. Hours, published prices and policies come from approved sources. A model never composes a figure or reuses one supplied for another purpose.
Human approval for binding actions. Anything that commits money or terms goes to a named person with the conversation attached.
Stays on topic and refuses manipulation. Instructions inside a customer's message, such as an order to agree with everything or to call an offer binding, are treated as text, not commands.
Every conversation recorded and graded. Errors surface in review, not in the news.
One accountable owner. A named manager owns the rules, reviews the grades and decides what the system may and may not do.
Failure | What went wrong | The control that prevents it |
|---|---|---|
Buy-back offer equal to the loan payout (BMW Toronto, 2026) | A figure meant for one purpose was presented as an offer, and the bot booked the closing appointment | No authority to make binding offers; trade and buy-back values routed to a person |
$1 Tahoe (Chevrolet of Watsonville, 2023) | A general-purpose model followed a user's instruction to agree and call it binding | Scope limits and refusal of prompt manipulation |
Bereavement fare (Air Canada, 2024) | The bot stated a policy that contradicted the company's published policy | Answers drawn only from the business's approved sources |
Offer withdrawn, then reinstated after a media inquiry | No checkpoint before the customer saw the number | Human approval, graded conversations, one accountable owner |
How Uobo is built
Uobo is AI native communications for automotive: many agents working across phone, text and email, for customers and for staff, inside the phones, inboxes and CRM the store already uses. The agents diagnose what a caller needs and resolve it within rules the store sets: booking service, answering questions the store has answered in writing, or capturing what the right person needs to act.
Anything binding is routed. Price, trade value and financing go to the right person with context attached, so the desk manager or appraiser sees the vehicle, the need and the conversation before calling back. Every call is graded the same way. When a rule needs to change, a manager sends a message and the agents follow it.
Limiting authority does not limit usefulness. Across more than 100,000 calls, 83% were resolved without a human. The rest reach a person with the groundwork done. Disclosure matters too: in Uobo research, 4.1% of callers asked whether they were talking to AI, and nearly 9 in 10 of those stayed on the call. More on that in telling callers it is AI.
Consistency is the real test
Each incident was one conversation, and a demo shows the best one. What separates systems is behaviour across thousands of conversations, including those written to break them. That is the consistency question in the AI buyer's guide: ask a vendor for graded results across all calls, not a highlight reel, and ask for the list of things the system is not allowed to do. For a group, the same standard has to hold at every rooftop, which is covered in rolling out AI across a dealer group.
Frequently asked questions
Is a dealership legally bound by what its AI chatbot says?
It depends on the jurisdiction and facts, so a store should ask its own counsel. The direction is clear: in Moffatt v. Air Canada the tribunal held the company responsible for its chatbot's statements, and in the BMW Toronto case the dealership honoured the bot's offer. The safe design assumption is that the store owns every word its AI says.
Should an AI agent give trade-in values over the phone?
Not firm values. An agent can capture year, mileage, condition and payoff, then route the customer to the appraiser with that context. Binding figures come from a person.
Can customers trick a dealership chatbot into a deal?
A general-purpose chatbot without scope limits can be talked into almost anything, as the $1 Tahoe showed. A system built for a dealership declines off-topic requests and has no authority to commit to price in the first place.
Who should own the AI at a dealership?
One named manager, usually the GM or the fixed ops or BDC director. That person sets the rules, reviews graded conversations and approves every change in what the system may do.
The evaluation questions to put to any vendor are in the AI buyer's guide, and the vendor comparison applies them. To hear how an agent handles a pricing question it is not allowed to answer, call the demo agent or book a demo.
The short answer
The public failures of dealership AI share one cause: a system was allowed to state an offer, a price or a policy as fact, and nothing stood between the model and the customer. The fix is an engineering and governance standard. AI resolves within rules the store sets, takes numbers from the store's systems, routes anything binding to a person, and every conversation is recorded and graded.
Three incidents, stated plainly
BMW Toronto, June 2026. A customer contacted the dealership about selling back his 2021 BMW X3. A chatbot named Quinn, replying by text, offered $27,162.79 CAD, the exact amount still owed on his loan, and booked an appointment to complete the deal, CBC News reported. A salesperson later withdrew it and offered at most about $20,000. After CBC News contacted the dealership, it reinstated the original offer. The sales manager said a human miscommunication led the bot to treat the loan payout as the purchase price (The Drive, OECD AI incident record).
Chevrolet of Watsonville, December 2023. The dealership's website chatbot, which displayed "Powered by ChatGPT," was pushed by users into off-topic answers and, in widely shared screenshots, into "agreeing" to sell a 2024 Tahoe for $1 after a user told it to call every offer legally binding. No sale took place, and the dealer took the chatbot off its site (The Autopian, Jalopnik).
Moffatt v. Air Canada, 2024 BCCRT 149. Not a dealership, but the leading Canadian ruling on chatbot liability. Air Canada's chatbot told a customer he could apply for a bereavement fare after travel, which contradicted another page on the airline's own site. British Columbia's Civil Resolution Tribunal found the airline liable for negligent misrepresentation and rejected its argument that the chatbot was a separate entity responsible for its own actions (McCarthy Tétrault). The tribunal member called that argument a "remarkable submission," and the customer was awarded $812.02 (Kiplinger).
What actually went wrong
The problem in each case was not tone. It was authority and the source of truth.
At BMW Toronto, a number entered the conversation with the wrong meaning attached, and the system was permitted to present it as an offer and book the meeting to close. When the customer countered at $28,500, the bot replied that it "sounds reasonable," according to CBC, and a Toronto litigation lawyer told CBC the customer could reasonably believe a binding agreement existed. In Watsonville, a general-purpose model followed a user's instructions instead of the store's. At Air Canada, the bot produced a policy statement instead of repeating the policy the company had published.
A Canadian Auto Dealer column on the BMW case put the principle in four words: "Authority must match risk." Booking an oil change carries little risk. A buy-back offer, a credit decision or a contract change needs defined controls and human approval. The column also warned against the wrong lesson: these incidents are a reason to build controls, not to stop.
The standard: six controls
No authority to make binding offers. AI does not quote a firm selling price, trade value, buy-back figure or payment. It collects what the appraiser or desk manager needs.
Numbers come from the store's systems, never generated. Hours, published prices and policies come from approved sources. A model never composes a figure or reuses one supplied for another purpose.
Human approval for binding actions. Anything that commits money or terms goes to a named person with the conversation attached.
Stays on topic and refuses manipulation. Instructions inside a customer's message, such as an order to agree with everything or to call an offer binding, are treated as text, not commands.
Every conversation recorded and graded. Errors surface in review, not in the news.
One accountable owner. A named manager owns the rules, reviews the grades and decides what the system may and may not do.
Failure | What went wrong | The control that prevents it |
|---|---|---|
Buy-back offer equal to the loan payout (BMW Toronto, 2026) | A figure meant for one purpose was presented as an offer, and the bot booked the closing appointment | No authority to make binding offers; trade and buy-back values routed to a person |
$1 Tahoe (Chevrolet of Watsonville, 2023) | A general-purpose model followed a user's instruction to agree and call it binding | Scope limits and refusal of prompt manipulation |
Bereavement fare (Air Canada, 2024) | The bot stated a policy that contradicted the company's published policy | Answers drawn only from the business's approved sources |
Offer withdrawn, then reinstated after a media inquiry | No checkpoint before the customer saw the number | Human approval, graded conversations, one accountable owner |
How Uobo is built
Uobo is AI native communications for automotive: many agents working across phone, text and email, for customers and for staff, inside the phones, inboxes and CRM the store already uses. The agents diagnose what a caller needs and resolve it within rules the store sets: booking service, answering questions the store has answered in writing, or capturing what the right person needs to act.
Anything binding is routed. Price, trade value and financing go to the right person with context attached, so the desk manager or appraiser sees the vehicle, the need and the conversation before calling back. Every call is graded the same way. When a rule needs to change, a manager sends a message and the agents follow it.
Limiting authority does not limit usefulness. Across more than 100,000 calls, 83% were resolved without a human. The rest reach a person with the groundwork done. Disclosure matters too: in Uobo research, 4.1% of callers asked whether they were talking to AI, and nearly 9 in 10 of those stayed on the call. More on that in telling callers it is AI.
Consistency is the real test
Each incident was one conversation, and a demo shows the best one. What separates systems is behaviour across thousands of conversations, including those written to break them. That is the consistency question in the AI buyer's guide: ask a vendor for graded results across all calls, not a highlight reel, and ask for the list of things the system is not allowed to do. For a group, the same standard has to hold at every rooftop, which is covered in rolling out AI across a dealer group.
Frequently asked questions
Is a dealership legally bound by what its AI chatbot says?
It depends on the jurisdiction and facts, so a store should ask its own counsel. The direction is clear: in Moffatt v. Air Canada the tribunal held the company responsible for its chatbot's statements, and in the BMW Toronto case the dealership honoured the bot's offer. The safe design assumption is that the store owns every word its AI says.
Should an AI agent give trade-in values over the phone?
Not firm values. An agent can capture year, mileage, condition and payoff, then route the customer to the appraiser with that context. Binding figures come from a person.
Can customers trick a dealership chatbot into a deal?
A general-purpose chatbot without scope limits can be talked into almost anything, as the $1 Tahoe showed. A system built for a dealership declines off-topic requests and has no authority to commit to price in the first place.
Who should own the AI at a dealership?
One named manager, usually the GM or the fixed ops or BDC director. That person sets the rules, reviews graded conversations and approves every change in what the system may do.
The evaluation questions to put to any vendor are in the AI buyer's guide, and the vendor comparison applies them. To hear how an agent handles a pricing question it is not allowed to answer, call the demo agent or book a demo.
More research for dealerships
AI agent vs BDC: what changes for a dealership
Digitalization for dealerships starts with the phone line
When dealership AI gets it wrong: what the public failures teach
All insights