Most dealership websites have a chat box now, and more of them are run by AI. The question an owner should ask is not whether the AI can answer a shopper. It is what happens when the shopper asks about money. Price, trade-in value, credit and financing are the four questions a chatbot is worst placed to answer. This article explains why, using one well-reported incident, and then sets out the rule Staffhand applies to every dealership build.
What happened in December 2023
In December 2023, Business Insider reported on a Chevrolet dealership in California whose website chat window was described by a user as “powered by ChatGPT.” A software engineer shopping on the site asked the bot to write a Python script, and it did. He posted screenshots, and other users then tried to get the bot to act against the dealership’s interests. According to the report, one of them got the bot to agree to sell a car for $1. The article notes that this was not legally binding. The software vendor’s chief executive told Business Insider the company shut down the bot for that dealer’s site; he said the logs showed many attempts to misuse it, that it had resisted and had not disclosed confidential dealership data, and that most real shoppers ask routine questions. Source: the Business Insider report as archived in the AI Incident Database, incidentdatabase.ai/reports/3536.
Two points stand out. First, the $1 agreement was not enforceable. Second, it did not matter. The screenshots travelled, the dealership’s name travelled with them, and the vendor pulled the bot. The damage was to trust, and trust is what a dealership chat window is supposed to build.
Why money questions are different
A chatbot can answer “are you open on Sunday” with little risk. Money questions are different for four reasons.
The answer depends on facts the bot does not have. A real price depends on the specific vehicle, current incentives, the trade, the buyer’s credit and what the manager will approve today.
The answer can be read as an offer. Even when it is not legally binding, a shopper who was told a number by the dealership’s own website will expect that number when they arrive.
Credit and financing are regulated conversations. The wording and disclosures matter, and the person asking deserves a person answering.
People will test the bot. Some visitors are shoppers. Some are there to see what they can make it say.
None of this means AI has no place on a dealership site. It means the AI needs a hard boundary around money.
Staffhand’s rule
Every Staffhand build runs under the same control rules, and one of them exists for this problem: price, trade value, credit and financing questions always go to a person. They are never answered by the AI. Not with a range, not with a “starting at.”
This applies to every channel the build touches: web chat, text, email and the follow-up messages that desks draft after a lead comes in. It applies to all 22 desks in a dealership build. If the question is about money, the desk’s job is to recognize it and hand it off.
How routing works in the approval queue
A Staffhand build has one approval queue. Desks draft; a person approves, changes or holds. Money questions follow a stricter path through it.
- The desk recognizes the question. Best price, trade value, approval odds, monthly payment.
- The desk does not answer it. It can acknowledge the message and say that someone at the dealership will respond. That acknowledgement is itself customer-facing, so it goes through the approval queue like everything else.
- The question is routed to a person with context attached. The lead’s history, the vehicle of interest and the conversation so far, so the person does not start cold.
- The person answers. A sales manager, finance manager or owner replies in their own words.
- The action is logged. The live action log records what the desk did and when, so the owner can see every hand-off.
The same hand-off shows up in the morning brief under what needs a decision. People at Staffhand review what the desks do as part of every build; nothing runs unattended. For the full set of controls, see How it works.
What the AI can still do
Removing money questions does not leave the AI idle. In a dealership build the desks draft lead replies, text back missed calls, book test drives, send confirmations and status updates, follow up with quiet leads, request and respond to reviews, keep the website and local listings current, draft inventory listings and price-drop alerts for a person to approve, and work repeat customers and referrals as a distinct lead source. All in approval mode, all in the dealership’s own accounts. The desks are listed at AI for car dealerships.
FAQ
Can the AI at least give a price range from the website listing?
No. Anything that reads as a price, trade value, credit decision or financing term goes to a person. A listed price is the website’s job, not the chat desk’s.
What does the shopper see while they wait for a person?
An acknowledgement that someone from the dealership will respond, drafted by the desk and approved by a person before it is sent.
What if a shopper tries to trick the AI into agreeing to something?
Nothing customer-facing is sent without approval, so there is no unapproved reply for a prank to produce. Attempts show up in the action log.
Does the rule slow down the response?
It adds a person to the money questions only. Routine replies, bookings and confirmations still move through the approval queue, and the morning brief flags anything waiting on a decision.
How your current chat, web forms and lead handling would hold up to a shopper asking about price is part of the lead handling score in a Staffhand digital audit. Start with a digital audit.
