A safe rollout plan

Start small

The best way to adopt agents is one narrow job at a time, with a clear way to know it works.

A four-stage rollout

Stage 1: Pilot with reading only. Choose a job where the agent only reads and reports, such as a daily briefing or inbox categorisation. No write tools. Run it by hand for a week and compare its output with what a person would have produced.

Stage 2: Add drafts and approvals. Give it a tool that proposes an action, such as a reply, behind approvals. Track how often you approve unchanged, approve after edits elsewhere, or deny. Fix the instructions when denials cluster.

Stage 3: Schedule it. Add a schedule or webhook only after the by-hand runs look good. Set budgets and guardrails first. Watch Reports daily for the first week.

Stage 4: Review regularly. Each week check Reports, Recent failures and Usage. Each month re-read the instructions, the tools and the key's rules, and run the test cases again.

Measure what matters

Choose two or three numbers before you start:

• Quality: how often is the output accepted without change?
• Time saved: how long did the job take before?
• Cost: typical cost per run times runs per month.
• Safety: number of denied approvals and failures.

Reports gives you success rate, typical run time and cost, and Approvals counts how many requests were approved or denied.

Know when not to scale

If quality is poor, fix it before adding volume. If approvals are mostly denied, the agent is aimed at the wrong job. If nobody reads the results, retire it.

Try it yourself

Sketch a rollout for one job in your team: what stage 1 reads, what stage 2 proposes, the numbers you will track and the date of your first review.

Recap

• Roll out in stages: read only, drafts with approvals, scheduled, regularly reviewed.
• Decide the measures before you begin.
• Fix quality before adding volume, and retire agents nobody uses.