Illustrative example, not a client engagement

Example playbook: customer email

This shows the shape of a playbook from a Human–AI Workflow Audit. The team, figures and rules below are invented to illustrate the format.

1. Snapshot

Workstream: incoming customer email for a small support team. AI already classifies messages and drafts replies from the knowledge base; people review and send. The aim of the playbook is to make clear who leads each step, when a person steps in, and who is accountable for the reply.

2. Roles

3. Review and override rules

• No AI draft is sent without a person approving it.
• Anything tagged refund, legal or complaint always gets a person’s review.
• A person may override any AI classification or draft; the override is logged with a short reason.
• When the AI is unsure or the case is unusual, it summarises the history only and hands over.

4. Accountability

The person who approves a reply owns it. The team lead owns escalated cases and the weekly review of overrides. The AI is never the accountable party.

5. Baseline metrics illustrative values

• Override rate (AI drafts changed substantially)e.g. 30% (placeholder)
• Rework (replies reopened within 7 days)e.g. 8% (placeholder)
• Median time to first replye.g. 6 hours (placeholder)
• Escalations to team lead per weeke.g. 12 (placeholder)

Placeholder values to show the format. A real playbook uses figures measured in your own workstream.

6. Hotspots

• Refund requests that the AI classifies as general questions.
• Drafts that quote outdated knowledge-base articles.
• Long threads where earlier promises to the customer are missed.

7. 30-day rollout

1. Week 1: agree the roles table and review rules with the team; start logging overrides.
2. Week 2: apply the refund, legal and complaint tags and the hand-over rule.
3. Week 3: first weekly override review led by the team lead; fix the top hotspot.
4. Week 4: re-measure the baseline metrics and adjust the rules.