Field notes
Train your agents on real conversations, not AI-generated data
A model learns the pattern from what actually happened - the real deals and emails, objections and all, left unsummarised. Feed it synthetic data and it learns to sound right while being wrong on the things that matter.
An AI agent learns the actual pattern only from what really happened - the deals, the emails and the objections, left unsummarised. Feed it synthetic data and it learns to sound convincing while getting the substance wrong.
On this page
Most people training an agent right now feed it a summary of what they think happened, and that's where it goes wrong.
A model learns from the mess - the raw transcript, the email that went three rounds before the deal turned. It learns from the call where the buyer said one thing and meant another. Summarise all that first and you've thrown away the exact signal it needed. You hand it your conclusion and none of the pattern that led there.
Real beats synthetic, every time
There's a shortcut going around: generate your training data with another model. It's fast and clean, and it's hollow. Synthetic data can teach an agent to sound like your best rep. It can't make it into one, because it never saw a real one close. Feed a model AI-generated conversations and you get an agent fluent in the average of the internet - plausible, and wrong in the ways that cost you deals.
Feed it the real thing instead, unsummarised: the conversations, the deals won and lost, the emails and objections and the events as they unfolded. Reasoning comes later. The model's first job is to detect the pattern - this shape of exchange, this kind of hesitation. Then it identifies the scenario and puts a probability on it: "this looks like a stall, and it's 70% budget." Only after that does reasoning have anything real to stand on.
The human never leaves the room
Someone has to train and validate every step of that, which is the part the demos skip. A human has to say "yes, that's a stall" or "no, you missed it - that was a champion going quiet because they got overruled." The model detects and the human corrects, every step, for a long stretch. An agent that no one who has closed deals ever validated is just a very articulate guess.
Agents are good. They are not human.
They can watch more and remember more, and they never get tired. But they are not human, and pretending otherwise is how you end up automating your own bad assumptions at scale. The agent's job is to handle the pattern-detection so the person can spend their judgment where it counts. It was never meant to replace that person. That's the line we hold on every AI build, and the same reason we're careful about where AI belongs at all.
So: real data going in, and human validation the whole way through. Everything past that is just a demo.