Built a digital car salesperson that can actually negotiate a deal — looking for dealer feedback

[mikkotark]

I’ve been building a proof-of-concept for a different kind of dealership AI.


Instead of another chatbot that answers questions and then asks for a phone number, this one is designed to act as a digital salesperson.


The customer can negotiate the actual vehicle price online. The dealer defines the commercial boundaries, and the system can accept, counter, reject or escalate an offer within those rules.


The important part is that the AI does not control the pricing policy.


The AI handles the conversation, while deterministic application logic enforces things like minimum acceptable price, negotiation limits, required vehicle-condition acknowledgement and deal-state transitions.


The current demo includes:


• dealer-controlled negotiation boundaries
• ACCEPT / COUNTER / REJECT / ESCALATE decisions
• limited negotiation rounds
• required condition-report acknowledgement before progressing
• persistent deal state and audit trail
• defined handoff points for dealership systems and human review


Financing, payments, DMS/CRM and similar systems are integrations around the salesperson rather than decisions made by the AI itself.


I built this initially around European automotive retail, and I’m now interested in how US dealers see the idea.


The question I’d really like dealer feedback on is:


Would you allow software to negotiate and agree on an actual vehicle selling price with a customer, as long as it could never operate outside rules and limits set by the dealership?


If not, what would be the deal-breaker: pricing control, compliance, desking workflow, trade-ins, management approval, customer behavior, or something else?


Demo: https://kopilotti.online


I’m not trying to sell anything here. I’m mainly interested in whether the underlying model makes sense from the dealer side, and what I’m missing.

 

[carguyfiles]

I'd let software move inside a floor a manager set. My deal-breaker is further upstream than any of the ones on your list: the number the floor is set from.

I build software that posts dealer inventory to Facebook Marketplace and keeps price and sold status in step with the dealer's inventory feed, so the input side of your model is what I look at all day. In our data 22.6% of listings change price after they go live. A boundary set on Monday can be wrong by Thursday, and a bot will defend that line with total confidence. Sold units are the same story. A car delivers Saturday, the listing stays up, and on Sunday your digital salesperson is negotiating a car that isn't on the lot.

What I'd want to see in the demo before anything else:

Floors and asks read live from the DMS at the start of every negotiation, never from a cached copy.
A hard stop the moment the unit's status changes mid-conversation.
No counter ever above the advertised price. The FTC's pricing FAQs from this month say the advertised price should be one any consumer can actually pay, and a bot countering above it is exactly the transcript I wouldn't want with my store's name on it.

Trade-ins, which you already listed, are the other gap. In a lot of US used deals the money is made on the trade and in the finance office, so a bot that only moves the selling price is negotiating the one number the customer already researched.

Against myself: our software stays out of the conversation entirely, and the dealer's own people answer every message, so I'm judging this from the input end and I've never run a negotiator. Putting the AI on the words and deterministic logic on the price is the right split, and the audit trail is the part I'd copy first.

Full disclosure, I build software for dealers, so weigh accordingly.

Why we keep our tool out of the buyer conversation: Why AutoLander Doesn’t Answer Your Buyers | AutoLander
How the price and sold-status sync works: Facebook Marketplace Inventory Sync for Car Dealers

 

[mikkotark]

There’s also work around the negotiation engine that is relevant to your point. The dealer administration tooling includes import validation, a review step, explicit approval for floor changes, and an audit trail. That addresses control over the inputs, although it does not establish real-time DMS synchronization.

Trade-ins are also beyond the idea stage: we have a separate implementation tested with a test valuation provider. I’d distinguish that clearly from a connected, production-ready valuation service.

We’ve also built an experimental analytics layer to compare dealership observations with external market information, while keeping prior system recommendations separate from observed evidence. That isn’t connected to a live market feed or authorized to change Sales prices.

So there is more around the selling-price engine, but your core challenge still stands: we need to demonstrate that the connected system reacts correctly when its source data changes.