How your team learns
Your team teaches Business-Gym by doing the work, not by writing rules. Every time someone approves, edits or dismisses a proposal, that decision is kept as a labelled example, and the next proposal of the same kind moves closer to it.
Two things bound that learning. Money, legal and external send always wait for a human — permanently, and that part is live today. The ladder described below is how the learning is designed to work: today the system proposes and you execute.
Do you have to write rules?
Section titled “Do you have to write rules?”No — that is the design. There is nothing to configure up front, because every operator action is already a labelled example.
Each example is filed against three things together: the intent (what the item is asking for), the sender, and the action taken. Confidence is tracked for that combination, not for the system as a whole. Handling one supplier’s invoices well says nothing about how another sender’s items are treated.
How is confidence meant to grow?
Section titled “How is confidence meant to grow?”Up four rungs, one at a time, for each intent × sender × action:
- SHOW — read-only. You see what it would have proposed; nothing is offered.
- SUGGEST — one click to accept.
- PRE-STAGE — prepared and waiting on one approval.
- AUTO-EXECUTE — runs without you, for that one combination only.
A combination only climbs after enough consistent examples. It never skips a rung, and it never carries its standing across to a different sender or a different action.
What decides whether something may proceed?
Section titled “What decides whether something may proceed?”A confidence × risk gate. Risk is checked first: if the action moves money, carries legal weight, or sends outside your company, it goes to a human and confidence is never consulted. Only reversible internal steps are even eligible for the ladder.
The confidence × risk gate: risk is checked before confidence, so money, legal and external send never enter the auto region — only reversible internal steps can climb, and a bad outcome sends them back down.
What never graduates, however confident it gets?
Section titled “What never graduates, however confident it gets?”Three classes never graduate, no matter how many clean repetitions the system has seen behind them:
- Money — anything that spends, pays or commits funds.
- Legal — anything with legal weight.
- External send — anything that leaves your company.
This is not a default you can raise later. Confidence has no path to these; they sit outside the auto region by construction, and that floor is in force today.
Can something lose its autonomy again?
Section titled “Can something lose its autonomy again?”Yes — graduation is designed to be reversible. One bad outcome demotes that combination back down the ladder, and it has to earn its way up again.
The demotion is as narrowly scoped as the promotion was: it affects that intent, that sender and that action, and nothing else.
What works today, and what does not?
Section titled “What works today, and what does not?”Today the system proposes; it does not auto-execute. Autonomy is scoped per mailbox, and the general ladder above is design rather than live behaviour.
Editable per-tenant playbooks, and a playbooks library and editor to manage them, are also designed but not built. There is no screen for them today.
Next: read the inbound → approve → execute loop this learning sits inside.