Default to draft mode: the agent suggests actions, the user approves, edits, or rejects. This is not a limitation — it is the feature that gets enterprise deals signed. No VP approves a tool that sends emails autonomously on day one.

The autonomy progression: (1) all outputs require approval, (2) low-risk actions auto-execute (internal notes, data lookups), high-risk require approval (external emails, financial changes), (3) most actions auto-execute with exception review, (4) full autonomy with monitoring. Most customers spend 2-4 months at each stage.

Build approval UX that is fast, not bureaucratic. One-click approve, inline edit, and 'approve all similar' batch actions. If approving an agent draft takes longer than doing it manually, users stop using the agent.

Every agent output should include evidence: citations to source documents, links to API records, diffs showing what changed, and a step timeline showing what the agent did and in what order.

'Based on Refund Policy v2.3, Section 4.1, this customer is eligible for a full refund' beats 'This customer is eligible for a refund.' The first is verifiable in 10 seconds. The second requires blind trust.

Step timelines are especially powerful for complex tasks: '1. Retrieved customer account (0.3s) → 2. Searched 847 policy documents (1.2s) → 3. Found matching clause in Policy #442 (0.8s) → 4. Verified eligibility criteria met (0.5s) → 5. Drafted refund response (2.1s).' Users who see this process trust the output.

Users need three override options on every agent output: edit (change the draft before approving), reject (discard and try again or handle manually), and teach ('always do it this way') feedback that improves future outputs.

The 'teach' option is your most valuable product feedback loop. When a user edits an agent draft, capture the diff and use it to improve prompts, add examples, or adjust playbooks. 'User changed formal tone to casual tone for this customer segment' is a product improvement signal.

Track override rates by workflow and customer. High override rates (>30%) indicate the agent is not calibrated for that use case. Low override rates (<10%) with high approval rates mean you can offer more autonomy. These metrics guide both product development and the autonomy progression for each customer.