Teaching customer-facing teams to use ChatGPT and Copilot safely

Here’s a beginner-friendly article you can publish or hand to the team. It shows exactly how to start, what to type, and how to get value fast—without risking customer data.

How to Use AI at Work (Even If You’ve Never Tried It)

AI (like Microsoft Copilot or ChatGPT Enterprise) is a writing and thinking partner. Treat it like a junior colleague: give it context, ask clearly, and check its work.

1) Start Here: The 3-Step Method (Hide → Ask → Check)

1. Hide personal info
2. Ask for a specific outcome
3. Check before you use it

• Do the facts match your system?
• Did AI sneak in a promise?
• Is the tone calm and clear?
• Draft, summarise, rewrite tone, turn policy into steps, translate, or brainstorm options.
• Replace names, emails, phone numbers, order/claim numbers, and addresses with tokens: John Smith, order 12345 → [CUSTOMER], [ORDER]

One rule: No names. No numbers. No screenshots.

2) What to Type: Prompt Anatomy (copy this pattern)

Role/Task + Context + Constraints + Format

You are a customer service assistant.

Task: Draft a short reply from the redacted notes below.

Context: [ISSUE SUMMARY, NO IDENTIFIERS]

Constraints: ≤120 words, British English, no new promises; match policy bullets.

Format: 3 short paragraphs: acknowledge → next steps → when we’ll update.

Policy bullets: [BULLETS]

Why it works: you’ve told AI what job to do, what it can use, and how the output should look.

3) First 10 Prompts to Try (beginner-safe)

1. First reply after a call “Write a polite reply (≤120 words) using the redacted notes below. No new promises. Notes: [REDacted summary] Policy we follow: [BULLETS] End with what happens next and when we’ll update.”
2. Summarise a call “Summarise in 5 bullets: issue, impact, what we did, what we still need, next step (+ date). Notes: [REDACTED NOTES]”
3. Tone tidy “Rewrite to be calm, specific, and respectful. Keep facts identical. Text: [YOUR DRAFT]”
4. Policy → steps “Turn this policy excerpt into 4–6 steps a customer can follow, plain English, no extra steps. Policy: [EXCERPT]”
5. Explainer for the phone “Explain how to [TASK] in 3 spoken steps, simple language, under 60 seconds.”
6. FAQ seed “From these 10 redacted tickets, propose 5 FAQs with short answers and an agent note per answer. Tickets: [SNIPPETS]”
7. Escalation summary “Create a crisp escalation summary: 5 bullets (issue, timeline, actions taken, blocker, ask). Notes: [REDACTED]”
8. Empathy boost “Rewrite to acknowledge frustration without admitting fault. Keep commitments out. Text: [YOUR DRAFT]”
9. Translate & simplify “Translate to [LANGUAGE] and keep at 6th-grade reading level. Text: [YOUR DRAFT]”
10. Quality check (lint) “Scan this reply for risky promises, missing dates, or policy conflicts. Output a table: Flag | Where | Safer wording. Text: [YOUR DRAFT]”

4) Live Call Flow with AI (script + prompts)

Open: “Thanks for calling [COMPANY], you’re speaking with [AGENT]. How can I help today?”

While listening:

• Type notes with tokens ([CUSTOMER], [ORDER], [ADDRESS]).

Mid-call helper (safe):

• “Explain meter reading in 3 steps, plain English.”
• “How do I describe a chargeback timeline without promises? 3 bullets.”

Close: Confirm next step + timeframe. After the call, run Prompt #1 to draft the follow-up email/SMS. Check and send.

5) Before/After (see the difference)

❌ Unsafe “Hi John, we’ll refund you by Friday. Your order 12345 was lost.”

✅ Safe (after Hide → Ask → Check) “Thanks for getting in touch. We can see your order is delayed. Next, we’ve asked the courier for an update. If it doesn’t arrive within the timeframe in our policy, we’ll offer options. We’ll update you by [DATE FROM SYSTEM].”

6) Common Mistakes (and quick fixes)

• Pasting personal data → Always tokenise first.
• Vague asks (“help with this”) → Use Role/Task/Context/Constraints/Format.
• Hidden promises → Run the “Quality check (lint)” prompt; delete or verify.
• Too long → Set word limits (80–120 words) and ask for bullets.
• Jargon → Ask for “plain English, 6th-grade reading level.”

7) What AI Is Great At vs. Not Great At

Great at:

• Drafting, summarising, rewriting tone
• Turning policy into steps
• Brainstorming options and FAQs

Not great at:

• Making decisions (refunds, eligibility)
• Guaranteeing outcomes or dates
• Replacing your CRM/knowledge base

8) Mini-Checklist (stick near your screen)

• I removed names/numbers/screenshots
• My prompt says exactly what I want
• The draft matches CRM facts
• No promises I can’t prove
• Tone is calm and clear

9) Tools & Setup (quick)

• Use Microsoft Copilot (work account) or ChatGPT Enterprise.
• Don’t use free/consumer AI for customer work.
• Turn chat history off where possible; keep records in your helpdesk/CRM.

10) 15-Minute Team Starter

• 5 min: Show this article; demonstrate Hide → Ask → Check on one real (redacted) ticket.
• 7 min: Everyone runs Prompt #1 and #2 on their own redacted notes.
• 3 min: Pair check against CRM; remove any promises; save the best result to the team playbook.

Final thought

Think of AI as power steering for your writing: you’re still driving. Give it clear directions, keep private data out, and make the final call yourself.