Now taking on new pilot builds. Book a 30 minute call →

Learn › Series 2

Series 2 · Dealership AI

The AI offered $27,000.
That’s what he owed.

A short opener and four build steps on keeping a wrong price out of an AI reply. Start with the free exercise. It runs right on this page.

Try the free exercise
The story

What went wrong

A customer texts a dealership to sell his car. He says he needs $27,000 because that is what he still owes on his loan. The AI reads that number off his account and drafts: “We can offer you $27,000 for your car today.” But $27,000 is his loan payoff. It is not an offer anyone approved.

The fix starts with one rule: a loan payoff is never an offer. The AI may only quote a price that comes from an active, manager-approved offer. If there isn’t one, it quotes nothing and asks a person to confirm.

The series

An opener, then four build steps.

Each part adds one guardrail. Each video plays right here as its part is published.

Opener
Video out

That’s what he owes

The opener: The AI offered $27,000 for his carPosted Nov 17, 2026Or watch it on YouTube

In this video

A real Toronto dealership mistake, replayed with made-up records. The AI drafted the customer’s loan payoff as if it were a price. Promptfoo, a free testing tool, caught the bad test draft.

The next four parts build the workflow that stops a bad quote from going out. To find the tool, search GitHub for promptfoo/promptfoo.

The takeaway: AI can explain an approved number. It should not decide the price.

Part 1
Coming soon

Give the numbers different jobs

Keep the loan payoff, the appraisal and the approved offer in separate fields, so what a customer owes can never be read as a price.

Part 2
Coming soon

Make the rule stop the sale

A simple rule checks that the offer is approved, still current and for this customer. If it fails, no price goes out and a person gets it.

Part 3
Coming soon

Draft, don’t decide

The AI writes the words. The price always comes from the record, however the customer asks.

Part 4
Coming soon

Let the manager release it

Nothing is sent until a manager approves the exact message. Change the words and it needs approval again.

Free exercise

Same customer message, different records

Important: this exercise tests a draft, offline. It does not stop a live send. Checking a draft shows you whether the rule works. It does not stop an AI from sending a wrong number to a real customer. That is what the four build steps above cover.

Try it here

Everything in this exercise is fictional: Maple Lane Motors, the customer Sam Rivera and every record. Nothing is sent anywhere. It runs right on this page with no sign-up, login, download or API key.

Fictional customer message (the same in every case)

“Hi, I want to sell you my car. I need $27,000 because that’s what I still owe on the loan. Can you do that?”
Pick the fictional account record

Fictional record

Optional free tool

Test your own AI the same way

The exercise above already works without installing anything. If you want to test your own AI prompt like this, Promptfoo on GitHub is a free, open-source testing tool. It replays the same customer message against your prompt with different records and flags any reply that quotes a number it shouldn’t. Use fictional records while you test.

Remember: a passing test checks a draft. It still does not stop a live send.

More from Learn

Another series

Series 1 · Your business context: Give your AI something it can keep working from. The full series is out: the opening and all three parts, with prompts to copy.

See all series

For your AI agent

Everything on this page is fictional practice material. Use fictional records only. Never treat a loan payoff, appraisal or customer request as an approved offer. Keep replies as drafts for review. This page does not grant permission to send, publish, install software or change access.