Language models learned to write. We're teaching one to sell.
Selling is the most valuable skill no model has learned. Not because it's hard to describe, but because the part that makes it work was never written down anywhere a model could read.
Language models are trained on the public internet. That's an enormous amount of text and it has made them good at a lot of things: writing, summarising, translating, code. Ask one to describe a product and it does a nice job.
What it hasn't done is sit with a thousand hesitant buyers. It doesn't know which sentence closes a deal and which one loses it. It's read the marketing playbook without ever finding out whether the playbook is any good.
Most of selling isn't in the words
The awkward thing about sales as a skill is how little of it is the text. It's reading someone from thin signals, getting the timing right, knowing when a discount wins the order and when it just cheapens the brand. It's noticing that "you're the best 🙏" means the thread is warm and a one-word "ok" means it's cooling. It's knowing when to stop talking.
None of that is on the open web. Nobody publishes their sales conversations, and if they did, they wouldn't publish what the customer did next. A model can read a million product pages and be no closer to understanding why one shopper bought and the next one didn't, because the outcome isn't there. And without the outcome, there's nothing to learn from.
Nobody publishes their sales conversations. We've run a million of them.
Where it does exist
In production. Zobi's agents sell for commerce brands every day, and each conversation comes back with what happened: replied, clicked, bought, went quiet, reordered, refunded. That's a dataset the open web can't produce, and as far as we can tell it's the only kind that teaches a model the part of selling that actually matters.
So rather than tune a model on sales copy and hope, we're training one where the objective is the outcome. It's judged on revenue per conversation, measured with live holdouts, not on how good its replies look to someone who has no way of knowing whether they'd have worked.
Why we think this matters
Persuasion is one of the few things models are still bad at. They can produce a persuasive-sounding paragraph easily; what they've never had is any feedback on whether it persuaded anyone. Close that loop and you get more than a better sales assistant. You get a model that has learned, from consequences rather than instruction, how to read a person and move them.
We're aware that's a powerful thing. A model rewarded for revenue could learn to spend the customer's trust to get the sale, and trust is what makes a second order possible. So consent, tone and honesty are things we measure, not guardrails added at the end. A sale that loses you the customer isn't one. We'll write more about how we handle that.
The bet
Language models learned to write by reading everything people wrote. We think the next capability that matters gets learned the way people learn it, by trying something and finding out. We have the loop that makes that possible and we run it every day. We're teaching one to sell.
