Scenario pack: customer service
Use AI for first drafts, not final customer answers.
When a small team answers similar low-risk questions repeatedly, AI can help prepare a useful first draft. A person still checks facts, tone, customer context, and every exception before anything is sent.
Use this scenario when
- The team repeatedly answers common, low-risk questions about public information, approved policies, hours, or standard processes.
- A reviewer can compare each draft with an approved source before it reaches a customer.
- The first test can use approved, non-sensitive examples rather than a live customer record.
- The purpose is better first drafts and less repetition, not tracking response volume or replacing customer judgment.
Do not use this scenario yet when
- The reply involves an account, payment, refund, legal issue, health information, complaint, employment matter, or private customer details.
- No person can reliably review the output before it is sent.
- The team has no approved information source, concern route, or named owner.
- The change is being introduced to raise individual quotas or monitor employee activity.
Run the smallest useful test
1. Decide
Check the workflow
Start with the customer-reply example in the First AI Workflow Kit. Confirm the task, owner, review step, data boundary, and likely time saved.
Check workflow fit2. Review
Make the pre-send check explicit
Use a review card for facts, tone, data, escalation, and the circumstances where a draft must not be sent.
Build review card3. Hand off
Map the exceptions
Define when the routine draft stops, who takes over, and what must happen before the work returns to the normal path.
Map exceptions4. Launch
Tell the team what is changing
State the purpose, human control, capacity intent, team influence, and no-penalty concern route before live use begins.
Build launch note5. Learn
Compare real examples
Test five to ten representative examples. Record time after review and rework, quality, and the team signal before deciding what changes next.
Record evidenceWhat a good result looks like
A good result is not “more replies per person.” It is a faster, reviewed first draft for routine questions; fewer repeated explanations; clear escalation for exceptions; and visible capacity for complex conversations or customer recovery. Stop or narrow the test if correction takes longer than the old process, quality drops, private information enters an unapproved tool, or the team experiences pressure or hidden catch-up.
Next decision after the test
Use the work-quality review to decide whether the workflow improved work, not simply whether it produced text faster.