Why our bots learn your prices but never change them on their own

"It learns as you use it" is on the website of almost every AI tool. It is rarely explained. Usually it means your data goes into a process you cannot see, and the tool behaves a little differently next month for reasons nobody can tell you.

That is not acceptable when the thing being learned is what you charge your customers. So we built OnDot's bots to learn in the open, from your own work, with you deciding every change.

What the bots watch

Every time a bot drafts a reply, OnDot notes what you did with it:

• Sent as written: the draft was right.
• Edited: OnDot records exactly what you changed, for example "raised front brake pads from $150–$250 to $180–$290".
• Held or discarded: the draft was not usable.

After the job, OnDot asks one question with one tap: did they book, and what was the final bill? That tells the bot whether its estimate was in the right place, not just whether you liked the wording.

All of this stays in your business's own record.

What they learn from it

Once a week, OnDot looks through your recent edits and results for patterns. It only speaks up when a pattern is clear, and it always comes with the evidence:

You raised front brake pads on 7 of the last 9 estimates. The prices you actually sent were usually $180–$290. Update your price list?

Final bills came in above your estimate on 6 of the last 10 jobs. Your price list may be low.

The rules behind that are simple and deliberately cautious. An item needs at least five edits in the same direction, making up most of its recent edits, before OnDot suggests anything, and the suggested range is the middle of what you actually sent. One odd job does not move your prices.

It changes your price list, not the AI

This is the part that matters most. When you accept a lesson, nothing inside the AI changes. Your price list gets a new version with the new range, and you can see what it was before.

The AI's job never changes: read the request, pick the matching items from your list, and say how many. The arithmetic is done in code, from your list. If the model ever writes an amount of its own into a reply, it is replaced with the one from your list.

So "the bot learned" always means something you can point to: a number on your list, with the date you accepted it.

Earning the right to send on its own

Some owners, after a few weeks, approve almost every draft without changing a word. For them, tapping Approve becomes a chore.

You can switch on earned autonomy for a form. It does nothing until it has been earned on your own record: 20 estimates, with at least 90% sent exactly as written. Even then it only sends estimates that are under your approval limit, and it never sends when a customer asks for a firm price. If your edits pick up again and the rate drops below 80%, it switches itself back to asking you. The form shows where it stands: "14 so far, 13 as written."

Starting from something better than a blank page

A brand-new business has no history to learn from. So every kit comes with starter prices, clearly marked as examples to change.

Once enough businesses use the same kit, at least five businesses and twenty confirmed price lists, a new business starts instead from the typical range those businesses charge, still marked as unconfirmed until the owner checks it. Only the numbers are used: never names, customers, messages or files. On the Crew and Agency plans you can switch off contributing your prices.

From day one, everything after that starting point is learned from your business only.

What the bots will never do

• Make up a price. Every amount comes from your list.
• Change their own rules. Every lesson is a proposal; you accept it or it does not happen.
• Send above your limit. Anything over your approval limit always waits for you.
• Pretend to be sure. When nothing on your list fits, the draft offers your paid visit or asks a question instead of guessing.

AI that improves is only useful if you can trust what it improved. We would rather show you every lesson than ask you to take our word for it.