predicts each decision before the expert makes it. It asks one question, at a natural pause, only when the expert surprises it. That captures the rules that never make it into a ticket or an SOP, and each rule is proven on sealed cases the expert never showed.
Your most experienced operator knows that equipment over a certain amount is capex, which vendor double-bills, and when to hold an invoice. None of it is in the SOP. When she leaves, it leaves with her. When you deploy an AI agent, it never knew it.
Process documents describe the steps, not the judgment. The exceptions live in people.
Experts can't list what they do without a case in front of them, and nobody has the hours to sit through it.
Both fail the same way: they follow the written process and miss the rule that was only ever said out loud.
The expert just works. only speaks up when its own prediction was wrong, and a rule is only trusted once later cases agree with it.
Before every save it writes down the fields and action it expects. When the guess is right, it stays silent. A wrong guess is the only thing it pays attention to.
Not a form or an interview. When her hands stop, it asks why. Her answer becomes a rule with a threshold, in her own words.
“Equipment over €5,000 net is always capex.”
demo expert · asked oncePredictions are frozen and hashed first, then humans label fresh cases around the boundary. A rule that fails gets re-fitted, not defended.
predictions committed before any label
A new hire about to book what the expert would hold is blocked, with the expert's sentence as the reason. The same Work Map exports to agents via API or MCP.
“No asset number, no capex booking.”
expert's words · Work Map rule“Equipment over €3,600 net is always capex.”
demo expert · AP · taught once
illustration from the built-in accounts-payable demo workflow
Our first audience: controllers and shared-services leads in accounts payable, facing expert turnover and pressure to deploy AI agents. Agent builders are second.
The expert teaches by working, with no documentation project. A new hire ramps against her real rules instead of a stale SOP.
Agents earn permission rule by rule, because every rule carries its evidence and its test result.
Read the Work Map through the API or MCP, rules and the expert's words together, instead of re-eliciting them for each client.
A pilot runs on a single workflow with one expert. It ends with a pilot report, so the decision to go further rests on measured results from your own cases.
How often its prediction matched her decision, overall and first half against second half, so you can see it learning.
Accuracy on fresh cases whose predictions were frozen and hashed before any human labelled them.
Which rules were learned, confirmed, approved or retired, plus how new hires fare on their first try.
Early access, plainly stated. Read the privacy policy and terms for the details.
try the demo now, or book a pilot and we'll run it on one of your workflows