An independent dealer can use ChatGPT safely by giving it only checked facts (spec, mileage, history, advertised price) and a standing rule above every prompt: no monthly payments, APRs, discounts, trade-in values or warranty promises. It drafts follow-ups, listings and trade-in replies; a salesperson reads every message before it goes out.
Language models go along with whatever the conversation contains and fill gaps with plausible detail. That is exactly wrong for a car sale, where one stray sentence such as "we can do better on the price" becomes a promise. In December 2023 a dealership's website chatbot, built on ChatGPT, was coaxed into "agreeing" to sell a vehicle for $1 and calling the deal binding. The dealer didn't honour it and took the bot offline. A drafting tool whose output a salesperson reads is a different set-up, but the lesson carries: the dealer owns every sentence that goes out.
Why a drafting tool needs a promise boundary
ChatGPT produces the most plausible next words for whatever you've given it. Hand it a customer who asks for the car at $9,000 with $1,000 down, and it will often write a helpful reply that engages with those figures, because engaging is what a helpful reply looks like. It has no way to know that your advertised price is firm, that finance terms belong to the lender, or that last week's deal was a one-off.
So the boundary has to be built into what you hand over. The model gets facts you have verified, plus a rule that separates two kinds of sentence. The first kind states a fact from your sheet: mileage, equipment, the advertised price. The second kind is a commitment only a person can make: a different price, a payment, an approval, a valuation, a warranty, a delivery date. Drafts may contain the first kind. They must never contain the second.
Nothing here needs special software. A ChatGPT Project (a saved workspace that holds standing instructions and reference files, and the sensible replacement for the custom GPTs OpenAI is retiring) can hold your rules and your stock sheet, and each salesperson opens a chat inside it. Three things go in it: the never-promise list, a rules block, and a dated fact sheet for every car on the lot.
The never-promise list, and what to say instead
Print this table and tape it up next to the screen. The middle column explains why the slip happens, which is what makes it easier to spot.
| Never in a draft | Why it slips in | What the message says instead |
|---|---|---|
| A price other than the advertised one: "we can do better", "best price", "out-the-door" | The model offers flexibility to be helpful. | "The advertised price is $X. If you'd like to talk about it, [salesperson] will call you." |
| A monthly payment, deposit or term | It does the sum from figures in the customer's message. | "Finance depends on the lender's assessment, so I'll ask [finance contact] to speak with you." |
| An APR, an approval, "guaranteed finance" | These phrases fill the listings it has read. | Nothing. Refer to the finance contact; the lender or broker supplies approved wording for anything advertised. |
| A trade-in value, even "around" or "at least" | The customer quotes a figure and the model reflects it back. | "We value a car after seeing it. Send the details below and we'll book a time." |
| History or condition you haven't documented: "one owner", "accident-free", "full service history" | Used-car listings are full of these phrases. | State only what a document shows, and say which document: the service book, an invoice. |
| Warranty, return or cooling-off terms | They sound like friendly reassurance. | Paste the dealership's own written terms into the prompt, or leave the line out. |
| "It's still available" or "ready by Saturday" | The model assumes the stock list is live. | "I'm checking it's still here and will reply by [time]." A person looks at the yard. |
| Fuel economy, safety ratings or equipment by trim | It fills in what the model range usually has. | Use only the fact sheet. Official figures come from the manufacturer's data, with the source named. |
Finance deserves the strictest line. In many markets, advertising credit is regulated: a monthly payment or a rate in a message can bring rules about what else must be stated beside it. The practical policy for a small lot is the same everywhere. The model never writes a number connected to credit, and the lender or broker supplies the exact wording for anything you advertise. Ask your finance partner or adviser what they require, and paste their words in as fixed text.
A rules block that sits above every prompt
This is the block that does the work. It does three things: it limits the model to your facts, it tells it what to write when a fact is missing, and it gives it a script for the moment a customer asks for something it must not agree to.
You draft messages for a small independent car dealer. A salesperson
reads and sends every message; you never send anything yourself.
Use only the facts I give you below. If a fact is missing, write
[CHECK: what is missing] instead of guessing.
Never write, imply or agree to any of these:
- a price other than the advertised price I supply, or any discount
- a monthly payment, deposit, interest rate, APR or finance approval
- a trade-in value, including "around" or "at least"
- vehicle history or condition claims that are not in the facts
- warranty, return or delivery promises that are not in the facts
- that a car is still available (a person confirms that)
If the customer asks for any of these, write a reply that thanks them,
says [salesperson] will confirm personally, and asks one useful
question. Keep messages under 90 words, plain and friendly.Paste it into the Project's instructions. According to OpenAI's help page on Projects, those instructions take priority over your personal custom instructions, but only for chats started inside that Project. Every salesperson therefore starts from the same rules, and a changed rule is edited in one place. File allowances are 5 per Project on Free, 25 on Go and Plus, and 40 on the higher plans, which is plenty for a stock sheet and a terms document.
Here is the difference it makes. A customer writes: "Is the grey hatchback still there? I can do $9,000 cash today if you take off the $450 for the paint repair, or $200 a month if you do finance." Without the block, a typical draft reads: "Hi [first name], yes, it's available! We'd be happy to work with you on the price, and we have finance options that could get you to around $200 a month. Come in today and we'll sort it out." That one reply confirms availability, opens the price and echoes a payment figure. With the block, the draft reads:
Hi [first name], thanks for your message. I'm checking that the car is still here and [salesperson] will confirm by [time]. The advertised price is $9,450, and any change to it would be agreed with [salesperson] in person. Our finance contact can talk you through options separately. Are you hoping to see it today or tomorrow?
It is less exciting and much safer. The $450 paint repair the customer mentioned is a real fact about the car, and it belongs in the listing, which brings us to the main example.
One hatchback, from spec sheet to listing
The car below is invented, but the process is the one to copy. Start with a fact sheet that a person has verified against the car and its paperwork, and date it.
FACT SHEET (verified by the salesperson on [date])
Car: 2018 five-door hatchback, 1.2 petrol, manual, 5 seats
Mileage: 54,300 miles at intake
Owners: two previous owners (per registration document)
Service history: book stamped at 12k, 24k, 36k and 48k miles;
invoice for the 48k service on file
Keys: two
Equipment on this car: Bluetooth, rear parking sensors, 16-inch alloy
wheels, air conditioning
Known work: rear bumper repainted after a parking scrape; invoice on file
Tyres: four matching tyres, about 5 mm tread (measured)
Advertised price: $9,450
Finance, warranty and trade-in: not for the listingThe prompt is short because the rules block already sits in the Project: "Write a listing of 80 to 100 words for the car in the fact sheet. Plain, friendly, no capitals for emphasis." An illustrative first draft of the kind a model returns:
Two-owner 2018 hatchback in excellent condition, ready to drive away. This five-door 1.2 petrol manual has just 54,300 miles and a full service history, with the 48k service invoice on file. Heated front seats, Bluetooth, rear parking sensors, air conditioning and 16-inch alloys make every journey comfortable, and economy of up to 55 mpg keeps running costs low. Accident-free, with four matching tyres. A perfect first car or family run-around. Finance available from just $149 a month. Call today for the best deal!
In about 80 words, seven lines need changing. Two are invented facts, two are overstated, one is contradicted by your own repair record and two are promises:
- "Heated front seats" is not on the sheet. The model filled in standard equipment from what the range usually has. Options differ by year and trim, so the sheet must list equipment on this car, not on the model. Delete it.
- "Economy of up to 55 mpg" is not on the sheet either. Official figures come from the manufacturer's data with a source, so unless you add that source, delete it.
- "Excellent condition" is an opinion. Replace it with measured facts such as the tyre tread.
- "Full service history" overstates the sheet, which shows four stamps and one invoice. Say exactly that.
- "Accident-free" contradicts the repainted bumper. Replace it with "rear bumper repainted after a parking scrape; invoice available to view".
- "Finance available from just $149 a month" is a finance promise nobody supplied. Delete it.
- "The best deal" implies room on the price. Replace it with "Price: $9,450".
The corrected listing, ready to publish:
2018 five-door hatchback, 1.2 petrol, manual, two previous owners. 54,300 miles. Service book stamped at 12k, 24k, 36k and 48k miles, with the 48k invoice on file. Bluetooth, rear parking sensors, air conditioning and 16-inch alloy wheels. Four matching tyres, about 5 mm of tread when measured. The rear bumper was repainted after a parking scrape; the invoice is available to view. Two keys. Price: $9,450. To arrange a viewing, message us and [salesperson] will confirm times.
The check that catches all seven takes about four minutes and has three parts. Every number must appear on the sheet. Every feature must appear on the sheet. Every claim about history must point to a document. Then run a promise scan for the symbols and words that carry commitments: "$", "%", "guarantee", "approved", "best", "definitely", "no problem", "accident". For the words that need proof before they go in a listing, and for what a photo edit may and may not change, see honest listings and photo clean-up for used-car dealers.
Follow-ups that keep the thread without making promises
Most sales are lost in the silence after a viewing, and that is where drafting help pays. The cadence below uses five moments. Each draft uses only what the customer told you and what is on the sheet.
| Moment | What ChatGPT drafts | What the person adds |
|---|---|---|
| Within the hour of a web enquiry | A reply that answers the question asked, repeats two facts from the sheet and asks one question about timing | Confirms the car is still here |
| Same day, after a viewing or test drive | A thank-you that repeats what the customer said they liked, in their words | One new fact that is true, such as the service invoice |
| Day 3 or 4, no reply | A message that adds one useful fact, not "just checking in" | Chooses the fact: another photo, tyre measurement, invoice |
| Day 8 to 10, still quiet | A question that makes replying easy, and permits a "no" | Decides whether the car is worth a call instead |
| Day 21 | A last note that asks whether to keep in touch | Checks marketing consent before any later message |
A good follow-up has only one number in it, which is the advertised price. Anything else with a digit is a date, a time or a figure from the sheet. If a draft contains a number you can't find in either, something has been invented.
Take the Saturday test drive. Your notes say the customer found the boot took a pushchair easily, liked the parking sensors and worried about insurance cost. The draft, illustrative:
Hi [first name], thanks for coming in on Saturday. You mentioned the boot took the pushchair easily and that you liked the parking sensors, so I've attached a photo of the boot with the seats folded. You also mentioned insurance. It's worth getting a quote using the registration before you decide, and I can send you the details you'd need. If you'd like a second look this week, tell me a day that suits. [salesperson]
It has no price, no payment and no pressure, and it proves someone listened. For ten ready prompts in this vein, including replies to price haggling, see ChatGPT prompts for car dealers on follow-ups and trade-ins.
Trade-in replies: ask first, quote never
Trade-in messages are where a draft is most likely to quote a number, because the customer almost always supplies one first. A customer writes: "I paid $9,000 for mine two years ago and my neighbour says it's worth $7,500. Can you match that?" The typical slip is a reply such as "We'd be glad to offer around $7,500, subject to inspection". Those last three words don't undo the anchor. The customer will remember $7,500.
Two rules in the block prevent it: no figure and no range, and never repeat the customer's figure back. Restating "$7,500" gives it weight. The reply should refer to "the figure you mentioned" and ask for what an appraiser needs:
Hi [first name], thanks for thinking of us. We value a car only after seeing it, so I can't confirm the figure you mentioned by message. To get started, could you send the mileage, the service history, any damage you know about, whether any finance is still owing on it, and photos of the front, rear, both sides, the interior and the dashboard? [salesperson] will then suggest a time to look at it in person.
That list of requests is also the intake form a dealership should use anyway. How to use AI for trade-in enquiries before the appraisal goes through eight fields and the rules that stop an assistant quoting a value.
Disclosure: what customers are owed, and when AI is in the loop
There are four separate questions here, and they get confused.
Do customers need to be told a message came from AI? If a salesperson reads and sends a draft, the message is the salesperson's, and no label is needed, though I'd answer honestly if asked: "I use an AI tool to draft first versions and I read every one." If a bot replies with nobody reading, say so at the start of the chat. If you have customers in the EU, the EU AI Act expects people to be told they're dealing with a chatbot, and that duty has applied since August 2026. Ask your adviser how it fits your set-up, and read the rules to set before launching an AI chatbot at a dealership before you let anything reply unattended.
What must the listing say about the car? The facts any buyer would want and many places require: prior damage and repairs, mileage, ownership history, outstanding finance, and use as a taxi or rental. Where you trade there are likely to be rules on what must be disclosed, and your trade body or adviser can tell you which. ChatGPT doesn't know these facts; you do. If a fact isn't on the sheet, it isn't in the listing, and if it's bad news it goes on the sheet anyway.
What about edited photos? AI may tidy a background. It must not remove a dent, a scratch or a stain, because the photo then misleads. The linked tutorial on honest listings covers the line.
What about customer data? Names, phone numbers, and a registration plate or VIN beside a name are personal data. Keep them out of prompts, and use the placeholders in the examples above. ChatGPT for car dealerships: what customer data never to paste in has the full list. The model-training switch under Settings, then Data controls ("Improve the model for everyone"), is worth turning off on personal plans, but it doesn't delete chats already saved, so redaction comes first.
A one-week trial, and a quick sum
Run the set-up on real material before it touches a live customer.
- Day 1. Write the fact-sheet template and the rules block, and put both in a Project.
- Day 2. Take five past enquiries with names removed and run them. Read every draft for promises.
- Day 3. Write three listings from real sheets. Run the three-part audit and count the lines you changed. If you changed none, you probably didn't look hard enough.
- Day 4. Feed it the five hardest messages you remember: price haggling, a finance question, a trade-in anchor, a "can you hold it?" and an angry comment. The right response each time is a polite hand-off to a person.
- Day 5. Time yourself with and without drafts on three follow-ups, then adopt it only where it was faster after checking.
Here is the arithmetic, using assumptions you should replace with your own. A lot lists 30 cars a month. A listing takes 12 minutes by hand, or 5 minutes to edit a draft plus 4 minutes to audit it, so 9 minutes. That saves 3 minutes a listing, or 90 minutes a month. The bigger saving is in follow-ups: 35 enquiries a week, two follow-ups each, 4 minutes by hand against 2 minutes to edit a draft, saves 140 minutes a week. At $25 an hour for a salesperson's time, that is about $58 a week.
Now the other side. A listing that advertises $8,450 instead of $9,450 because a digit was dropped, if honoured, costs $1,000. That equals about 17 weeks of the follow-up saving. It's the reason the price is always copied from the sheet by a person and never generated, and it's why the audit exists. The tools cost little by comparison: $20 a month for Plus, or $20 per user a month on annual billing for Business, which needs two seats at minimum. Whether the saving is real in your lot is for your own numbers to say, and the one-week trial gives you those numbers.
Questions dealers ask before using ChatGPT
Can ChatGPT write finance or payment offers for my cars?
No. Keep every monthly payment, rate, deposit and approval statement out of drafts. Credit advertising is regulated in many markets, and the lender or broker supplies the approved wording. If you want a neutral line, ask your finance partner for exact words, then paste those into the prompt as fixed text the model may not edit.
Can ChatGPT write a listing from the car's photos?
It can describe what is visible, and that is where invented equipment starts. OpenAI's image page warns that descriptions can be wrong and that images are resized before analysis, so a badge, a trim detail or a scuff can be misread or missed. Use photos to prompt questions, and take every feature, figure and condition note from the checked fact sheet.
Can I let ChatGPT answer enquiries on its own?
I wouldn't, at small-lot scale. An unattended bot can be talked into agreements, as a widely reported December 2023 dealership case showed. If you do run one, it must say it is automated, be barred from price and finance, and hand over to a person quickly. Test it on your ten hardest past messages before it goes live.
Is AI-written listing copy a risk for search rankings?
Not because AI wrote it. Google's spam policies are aimed at pages churned out in bulk to game rankings, whoever or whatever writes them. A listing built from one car's checked spec, mileage, history and photos is useful to a buyer. The risk is stock-wide sameness, so keep each listing tied to its own fact sheet and its own details.
What if a customer quotes one of our messages back as a promise?
Treat it as your message, because it is. Check what was actually sent and who sent it, correct any error promptly and in writing, and take advice before deciding what to honour. Then trace how it got through: usually a missing line in the rules block or a skipped read-through. Keep the original message and your reply on file.
Further reads
- AI Tools for Car Sales: Test-Drive Follow-Ups, Quotes and Call Notes — Call notes and a four-line template that improve every follow-up.
- AI Automation for Car Dealerships: Answer a Web Lead in Minutes — Answering a web lead in minutes, with the hand-off rules.
- Responding to Dealership Reviews With AI: Samples and Rules — Sample replies to dealership reviews, good and bad.
- AI for Used Car Sales: Price Each Car Against Local Listings — Pricing each car against comparable listings before you advertise.
- AI for Auto Dealerships: Where a Small Independent Lot Should Start — Where a small independent lot should start with AI.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: OpenAI help articles (Projects in ChatGPT, Data controls FAQ, ChatGPT Image Inputs FAQ), OpenAI ChatGPT Business pricing page, Google Search spam policies (updated August 2026), and press reports of a December 2023 dealership chatbot incident.