Design an AI Workflow That Sticks · Lesson 6 of 7

Lesson 6 — Testing with real users

30 minutes. Real-world testing before rollout — without putting customers at risk.

---

Before you let the workflow touch real customer-facing work, run it in shadow for one week:

• The workflow runs every time the trigger fires.
• AI proposes the output.
• A human reads the output.
• The human does *not* send the AI version. They write their own reply / take their own action.
• The two are compared at end-of-week.

You learn three things from a shadow week:

• How often does AI get it right? If AI's output would have been usable as-is in 7 of 10 cases, you're ready to go live with human approval. If 2 of 10, your prompt needs more work.
• What does AI miss? The cases where you wrote something AI didn't propose are gaps in the prompt's context or instructions.
• What does AI add that you don't? Sometimes the AI's draft is *better* than what you'd write. Note those — they make your final prompt stronger.

When you go from shadow to live, pick one person to use it for two weeks before rolling to the team. Not the most enthusiastic person — the most skeptical.

If the skeptic uses it consistently for two weeks, you have a workflow. If they bail in three days, listen to why.

→ Lesson 7 — Rolling it out without scaring anyone

Previous: Lesson 5 — Approval loops and audit trails · Next: Lesson 7 — Rolling it out without scaring anyone