If you ask a frontier model to write a follow-up text to a car buyer who has gone quiet, you’ll get something polite and fluent, close to what a decent salesperson would send. What you won’t get is any sense of whether that text will bring the buyer back, because the model has never sent one and waited to find out.
What the internet left out
Language models learned from what people wrote down, and that covers a remarkable amount: product pages, reviews, sales books, and years of forum threads about haggling over a trade-in. It’s more than enough to describe a car, explain how a lease works, or draft a polite follow-up.
What it leaves out is almost everything that actually decides a sale. Nobody publishes their sales conversations, and the few that end up online never include what the buyer did afterward. A model could read every dealership website in the country and still have no idea why one shopper booked a test drive while the next one stopped replying. Without the ending, all that text can teach is style, and style was never the hard part.
Most of selling isn’t in the words
If you watch a good salesperson for a week, you’ll notice how little of what they do comes down to phrasing. They can tell that "what’s my trade worth?" sent at 10pm means the buyer is close, and that "just looking" on a Saturday morning usually means the opposite. They know when price belongs in the first message and when bringing it up will end the thread, when a call will land better than a text, and when it’s time to stop following up altogether.
Those are judgments about timing and about people, and in a transcript they show up only as a choice someone made, with no record of whether it was the right one. People pick them up by trying things, watching what happens, and adjusting, and we think a model has to learn them the same way.
Where the missing data lives
The data that teaches this only exists in production, in conversations whose ending is known.
RevDesk agents reach out to leads for dealerships and service businesses every day, usually within seconds of a web form, and then keep the conversation going across texts, iMessage, WhatsApp, email, and calls until the buyer books, replies, or asks them to stop. When something comes of a conversation, whether that’s an appointment set, an appointment kept, or a unit sold, the outcome is recorded against it, along with whether the agent got there on its own or with a salesperson’s help.
That record is exactly what the internet never had. It also belongs to the businesses we work for, which is why the way we use it matters as much as the data itself.
Two models, one promise
We’re building two things that work together.
The first is a single global model that acts as the judgment layer, deciding when to reach out, what to say, and when to stop. It’s built on foundation models, licensed and public material, and conversations from businesses that choose to contribute them, which they can only do once an admin turns contribution on.
The second is an agent for each business, which learns that business inside its own workspace from its knowledge, its playbooks, and how its own conversations ended. It learns by remembering rather than training, and it only ever works for the business that runs it.
Put simply, your data and conversations never train anyone else’s models, and they train ours only if you choose to help improve them.
The trap in rewarding sales
Anything rewarded for sales will eventually find shortcuts. Pressure works once, and so does a deadline that isn’t real, but each of them spends the trust that brings a customer back for service, for their next car, and one day for their kid’s first car.
Some of the guardrails don’t depend on the model at all. Consent, quiet hours, and attempt limits are checked by rules before a message is ever written, so there’s no way for the model to talk its way past them. When someone replies STOP by text or iMessage, the calls, texts, and iMessages to that person stop, and every decision an agent makes is written down with its reason, so a manager can always see why it reached out.
The harder part is the score itself. We want a model that wins the second sale as well as the first, which means a sale that costs the customer’s trust can’t count as a win. How we measure that is something we’ll keep writing about here.
The bet
Language models learned to write by reading what people had written. We think the next skill they pick up will come the way people learn it, by trying something, seeing what happened, and doing a little better the next time. Selling is where that loop is easiest to see, because the ending gets written down: the test drive either happened or it didn’t.
Our agents hold these conversations every day for real businesses, and every ending is recorded. The work in front of us now is turning that record into a model that has actually learned to sell.
