The autonomy dial: set it per decision, not per system

Set per decision type, not per system

“Turn on the AI” is the framing that makes careful buyers nervous — one global setting, one leap of faith. ADIP doesn't have that setting. Autonomy is defined where the risk actually lives: at the level of the decision type. Re-dating a routine purchase order inside an agreed window is one policy. Reallocating supply across contract customers is another. Anything that touches a penalty clause can be a third.

Each decision type carries its own written answers to three questions — explicit, reviewable policy your team owns, not behavior buried somewhere in a model:

• What the agent may do without asking — the bounds: which fields, what magnitudes, which windows, how much spend.
• What it must attach when it acts — the evidence: the triggering signal, the alternatives priced, the constraints applied.
• What sends the decision to a person instead — the escalation triggers: thresholds crossed, patterns the policy doesn't recognize, consequences it doesn't name.

Routine runs; everything else escalates

Inside its bounds, a routine decision simply runs: the agent senses the change, prices the options, acts within the envelope, and logs the full trail — proposed, not smuggled, even when no approval gate was required. This is the work your planners burn mornings on today: the fortieth unremarkable re-date, not the one that matters.

Everything outside the bounds escalates by default — and escalation is designed behavior, not an error state. The decision arrives in front of a person with the choice set already priced and the reason for the escalation stated. The dial's resting position is ask.

Write authority is earned read-first

A new agent starts the way a new planner would: read-only. It watches the live decision flow, drafts recommendations, and is judged on them — while a human still makes every call. Only when its recommendations for one decision type have been watched and trusted does that decision type graduate to acting within bounds. Nothing about the graduation is automatic; your team promotes the agent, one decision type at a time.

And the dial turns both ways. Tighten bounds, demote a decision type back to read-only, widen authority where the record has earned it — each is an ordinary policy change, applied from the next decision it touches and logged like everything else. Autonomy is never something you granted once and hope about; it's something you are always currently choosing.

The questions to ask any agentic vendor

If you're evaluating agentic AI — anyone's — make this concrete. Where is autonomy set, and how finely? What happens to a decision the policy doesn't cover? How does an agent gain write access, and how fast can you take it back? Vague answers to those questions are where the cautionary tales come from.

ADIP's answers are structural: per decision type, escalate by default, read-first, revocable at once. Judgment stays human — the dial only decides which decisions are worth a person's morning.