Autonomy & Approvals

Choosing a level

Three questions decide how high an action class should go: how bad is a wrong action, can it be undone, and how often does it happen.

Autonomy ladder

Autonomy is set per action class (for example "Assign housekeeping task" or "Create work order"), never as one global switch. Every class starts at L0.

Why a ladder and not a switch? Trust in people grows step by step, and so should trust in a system. Each level lets you see how the class behaves before the next level lets it do more, and a bad week costs you one step, not everything.

Moving up and down

• Promotion is always a decision by an administrator. The screen shows the acceptance rate, error trend and number of cases behind the request.
• Demotion happens by itself. If acceptance or accuracy drops below the threshold, the class falls one level and the administrator is notified.
• Kill switch. One control returns a module, or the whole tenant, to L0 immediately.

Limits

Limits are conditions an action must satisfy at every level above Suggest. They are checked every time, even for actions that have run a thousand times before.

Shadow mode before promotion

Before you promote a class, run it in shadow mode. Enterprise Mind makes its suggestion silently while people carry on as usual, then compares its suggestion with what people actually did.

• Nothing is shown to staff and nothing runs.
• The Autonomy screen shows how often the silent suggestion matched the real decision.
• Cases where they differ are listed, so you can see who was right.

Approval queue

Suggestions that need a decision wait in the approval queue. Each card is complete on its own: you should not have to open another screen to decide.

What a good approval queue looks like

• Short: cards that need no decision are never shown.
• Self-contained: you can decide without opening another screen.
• Owned: every card has a named approver role, and each role has a deputy.
• Honest: a low-confidence card says so and escalates instead of pretending.

• Approve, edit or reject. Editing and rejecting both ask for an optional reason that goes into the ledger.
• Escalation. Low confidence moves a card to the approver role you configured. Escalated cards are marked.
• Safe to retry. Approving the same card twice runs the action only once.
• Unknown is not failed. If an action gets no answer from an outside system, the card shows "unknown" and asks for a check instead of reporting success or failure.

Escalation

A suggestion is escalated when it needs more authority than the first approver has, or when the engine is not sure enough.

Action catalog

Enterprise Mind can only run actions that an administrator listed in the catalog of the module. Anything not on the list cannot run, whatever the level.

Related pages: Memory · Governance, Privacy & Voice

Worked example: promoting a class

The hotel operations manager wants "Assign housekeeping task" to stop waiting for clicks.

1. The class has been at Suggest for four weeks. The Autonomy screen shows 92 percent acceptance across 310 cases.
2. The manager runs a week in shadow mode for the Approve level. The silent suggestion matches what staff did in 94 percent of cases.
3. The class moves to Approve. Approvers now accept with one click.
4. Three weeks later acceptance is still above 90 percent, so the manager allows Limited auto for tasks under $50 that can be reassigned.
5. In week two of auto-run, acceptance dips to 68 percent after a staffing change. The class falls to Approve by itself and the manager is notified.

Nothing dramatic happened. The ladder did what it is for: autonomy followed the evidence up, and down.

Reviewing what ran

Every action that ran on its own is listed on the Autonomy screen with what it did, which limits it passed, and a button to undo it where undoing is possible. Review it weekly while a class is new to auto-run.