Practical guidance
AI playbook
A starting point for deciding what to measure, what to change, and where help could be useful.
1. Start with the work
List repeatable tasks, how often they happen, who handles them, and what they cost. Select a workflow with a clear owner and a measurable outcome before choosing a model.
2. Assist before automating
Begin with drafts and human review. Record the corrections and failure cases. Expand automation only when the evidence supports it and the consequences are manageable.
3. Evaluate before expanding
Use representative examples and define what an acceptable result means for this workflow. Test exceptions as well as ordinary cases, and rerun checks after changes.
4. Include the full cost
Estimate model usage, hosting, integrations, maintenance, and human review. Compare the result with usable time saved or measurable incremental value.
5. Design the fallback
Validate inputs and outputs, restrict tool access, and make escalation explicit. A person should be able to identify and recover from a failed run.
6. Give it an owner
Document how the system runs, how to monitor it, and who can change or pause it. Review quality, usage, cost, escalations, and one business outcome regularly.