Trust ladders for business agents

Companies don't need "an AI agent" — they need specific tasks done with a level of autonomy they're comfortable with. I prototyped a control plane around that idea; these are the concepts that held up.

The ladder

1. Observe — the agent watches and summarises; it changes nothing.
2. Suggest — it proposes actions in a queue.
3. Draft — it prepares the action (an email, a record) for a human to send.
4. Act with approval — it executes after a one-click approval.
5. Act and report — it executes and reports; humans audit after the fact.

Each task climbs independently, based on measured outcomes — not on how impressive the demo was.

Policy booklets

Every agent gets a plain-language booklet: allowed and prohibited actions, tone, data sources, escalation triggers, an autonomy level and a cost ceiling — versioned like code, readable by a manager.

Cost routing

Most requests don't need the biggest model. Routing rules send simple classification to small models and reserve large ones for drafting and reasoning, with spend limits per agent.

Open questions

• How do you demote an agent gracefully after an incident?
• What is the right unit of audit — a prompt, a decision or an outcome?