Software simulation · Hand-coded HTML
AI Copilot for Customer Support
A support-console simulation where agents handle six customer chats with an AI copilot, deciding when to send, edit or escalate its suggestions.
The challenge
A contact centre rolled out an AI reply assistant. Agents either sent every suggestion unread, or ignored the tool entirely. Neither is safe, and a policy document can't teach the judgement in between.
My approach
- A realistic console: chat queue, transcript, customer record and a copilot panel, so practice looks like the real job.
- Six chats where the copilot is right, cold, wrong, unsafe or overconfident, so learners can't win by always accepting or always rejecting.
- Every decision requires checking the customer record or policy on screen, which is the habit the course exists to build.
- Immediate customer reactions and coaching, a shift report with the best handling for each chat, and a printable tone and escalation guide.
Objectives and how each is assessed
The simulation is the assessment: each chat is worth 2 points for the best handling, or 1 for safe but slower handling.
| Learning objective | Assessment |
|---|---|
| Use AI-suggested replies while keeping the brand tone. | 2 chats: send a good suggestion as is, and rewrite a cold one |
| Spot an incorrect or risky AI suggestion before sending it. | 2 chats: a wrong returns policy and an unsafe product-safety answer |
| Escalate cases the AI should not handle. | 2 chats: account takeover and a customer in financial difficulty, routed to the right team; pass mark 9 of 12 |
Results and reflection
Add pilot results, reviewer quotes or what you would do differently next time.
Design artefacts
- Support tone and escalation guide (printable)
- Storyboard (available on request)