Somewhere in your group right now, a BDC (Business Development Center) agent is pasting a customer’s email thread into ChatGPT to draft a faster reply. That thread might contain a name, a phone number, a trade payoff, maybe a line about the customer’s credit situation. Nobody told the agent this was wrong. Nobody told them it was fine, either. Nobody decided.
That’s the problem. Not the AI. The missing decision.
I’m a fractional CMO. I spent 3.5 years running marketing for La Mesa RV, an 11-store RV dealer group, where I moved $500K a month from unmeasurable broadcast into measurable digital. I’ve never worked at an auto dealer — but the mechanics here don’t depend on the inventory. A multi-store group, a CRM full of customer information, salespeople adopting tools faster than leadership can govern them: the architecture is the same whether the inventory has four wheels or a kitchen. What follows is a plain way to answer the question your team is already answering for themselves: can we put customer data into ChatGPT?
What’s already happening on my floor?
More than you think, and most of it is reasonable. Salespeople are drafting follow-up emails. BDC agents are rewriting the same appointment-confirmation message for the fortieth time. Someone in the internet department is punching up vehicle descriptions that the DMS (Dealer Management System) spit out as a comma-separated list of options.
None of that is the problem. The problem is what gets pasted alongside it. A follow-up email draft is harmless. A follow-up email draft that includes the customer’s full name, cell number, the vehicle they’re financing, and a note that “their co-signer fell through” is a different object entirely. The useful work and the risky data travel together in the same paste, and nobody on the floor has been given a line to draw between them.
This is where a decision has to be made and owned. If the group hasn’t decided what’s allowed, every individual employee is making the call alone, dozens of times a day, under time pressure, with no training. That’s not a people failure. It’s a policy gap.
What must never go into a personal AI account?
Here’s the short list, and I’d make it non-negotiable:
- Anything from a credit application
- Social Security numbers and driver’s license numbers
- Bank account or card details
- Deal jackets, in whole or in part
- Customer lists exported from the CRM or DMS
The legal frame — and have your counsel review this, because I’m a marketer, not a lawyer — is the FTC Safeguards Rule. The FTC’s position is that dealers who finance or facilitate financing for consumers are financial institutions under the Rule, and its June 2025 dealer FAQ covers what counts as customer information, how service providers must be handled, and the requirement for a designated qualified individual. An AI tool that receives customer information looks like a service provider under that framework — which means it would need to be vetted and contracted like one, not adopted by whoever signed up for a free account on their phone. (The same discipline applies to every other vendor touching the group’s data.)
That gives you a clean test, and it’s the one I’d put in the policy verbatim: customer information under the FTC Safeguards Rule goes only into a tool your group has vetted as a service provider. If the tool hasn’t been through that review, the customer data doesn’t go in. Full stop.
So what can my team put in?
Plenty. The goal of a dealership AI policy is not to kill the productivity gain. It’s to keep the gain and strip the exposure. Safe inputs include:
- Vehicle descriptions. Year, make, model, trim, options, mileage, price. It’s already on your website. Let the AI rewrite it forty times.
- Reply templates. “Write me five versions of a no-show follow-up for a service appointment.” No customer attached, no problem.
- Public inventory facts. Towing capacity, fuel economy, package contents — with the caveat that the team verifies the output against the actual spec sheet, because these tools state wrong facts confidently.
- A customer question with identifying details stripped. “A customer is asking whether they can trade in a leased vehicle before the lease ends — draft a plain-English explanation” is fine. The paste that includes the customer’s name, lease terms, and payoff is not.
The strip-the-details rule is the workhorse here. Most of what the floor wants from AI is phrasing, not facts about a specific customer. Teach the team to pose the situation generically and the risk mostly evaporates while the speed benefit stays.
Does it matter which tool and which plan we use?
It matters more than almost anything else in this decision, because the training defaults differ by plan — and the plan your salesperson signed up for on their personal phone is almost certainly the wrong one.
As of this writing (settings change; check the vendors’ current pages):
- ChatGPT Free, Plus, and Pro personal accounts share data for model training by default, with an opt-out in Data Controls (OpenAI help page). ChatGPT Business, Enterprise, and the API don’t train on business data by default (OpenAI enterprise privacy).
- Consumer Gemini chats can be reviewed to improve Google services unless Keep Activity is off (Google support). Gemini in Google Workspace is not used for training outside your domain without permission (Workspace privacy hub).
- Consumer Claude chats train only if the user opts in (Anthropic privacy). Claude for Work and the API are not used for training by default (Anthropic privacy).
The pattern across all three vendors: business plans default to not training on your data; personal plans either train by default or depend on a setting your employee has never opened. So the policy move is simple. Name one approved tool on a business plan, put the group on it, and keep personal accounts out of customer work entirely. Don’t write a policy that says “be careful on your personal ChatGPT.” Write one that says customer work happens on the group account, period. A rule that requires each employee to configure their own privacy settings correctly is a rule that will fail.
What does the actual policy look like?
One page. If it’s longer than a page, the floor won’t read it and the BDC won’t follow it. Five components:
- The approved tool and plan, named specifically. Not “AI tools” — the exact product, the exact tier, the group login structure.
- The never list. Credit applications, SSNs, driver’s license numbers, bank and card details, deal jackets, CRM and DMS exports. Verbatim, so there’s no interpretation.
- The strip-the-details rule. Pose customer situations generically. If a real name, number, or deal term is in the paste, it doesn’t go in.
- An owner. Under the Safeguards Rule framework you already need a qualified individual; this policy naturally belongs to them, or to the group’s marketing lead if they’re the one driving AI adoption. A policy without an owner is a suggestion.
- A quarterly usage check. Vendor settings change, employees drift, new tools appear. Someone looks at how AI is actually being used, every quarter, and updates the page.
That’s it. You could draft this in an afternoon, have counsel review it in a week, and roll it out at the next managers’ meeting. The cost of not doing it is that the decision keeps getting made ad hoc, by the newest hire, on the worst possible account.
Isn’t this an IT problem? Why is a marketing guy writing about it?
Because the same tools your BDC is using to answer leads are the tools that will write your group’s marketing — the inventory descriptions, the email campaigns, the service-drive promotions, the ad copy. The question of what goes into the AI and the question of what comes out of it are the same question, and marketing leadership has to own both ends.
The policy that keeps customer data out is also the policy that keeps the facts in: approved tools, verified inputs, someone accountable for what the machine is fed and what it produces. (The governance layer marketing AI needs and what to settle before an AI agent touches your marketing are the output side of this same policy.) A group that governs AI inputs well is a group that can trust AI outputs — and that trust is the precondition for actually using these tools to produce marketing at scale instead of dabbling with them.
This is where the money is left on the table. Groups that treat AI as an IT ticket get a blocked tool and a workforce using personal accounts in the parking lot. Groups that treat it as a leadership decision — tool, plan, policy, owner — get the speed without the exposure. At La Mesa RV, the lesson of moving $500K a month out of broadcast wasn’t about the channel. It was that measurement and governance are what turn a tool into an advantage. The same logic applies here.
Who owns this decision?
Someone at the group has to own this: pick the tool, write the page, run the quarterly check, and then put the same governed tools to work on the marketing itself. That’s exactly the kind of decision a fractional CMO at a dealer group exists to own — connecting the compliance question to the growth question, so the group gets the speed of AI without betting customer trust on an intern’s chat history. If nobody in your org chart owns that decision today, that’s the gap worth closing first.