Skip to content
Get in touch
Menu
Field Guide · level 5 · Office work, by task

How do you write customer replies with AI and keep them right?

Draft from your policy. Check every promise. Hard cases go to a person.

5 min · try it 10 min · as of Oct 2026
Bring this to your team · Workflow automationSet up your team's reply desk: policies in, promises checked, hard cases to a person. Get in touch
01

The reply sounded lovely

It promised a full refund. Nobody had approved one.

02

Three habits, every reply

Treat it like a keen new starter:

  1. Hand it the policy and best replies.
  2. Check every promise.
  3. Hard cases go to a person.
03

Hand over the binder

Your real policy and three replies you were proud of. Keep them in a project.

04

Check every promise against it

Match every promise, date, number and name to the policy. Not there? Cut it.

05

Hard cases go to a person

Legal threats, health, vulnerable customers, real anger: a person owns those. AI just helps draft.

06

Emails can hide orders

Customer emails can hide text aimed at the AI. So it drafts; you send.

My pick for email

Claude logoClaude, with Gmail or OutlookAlso good: Gemini logoGemini in Gmail

The Gmail connector asks before it sends, replies or forwards. At an Outlook company, Claude for Outlook drafts replies for you to send.

All my picks
07

No policy, no promise.

Every promise points to a line in your policy. Then a person sends it.

Try it · 10 minutes

Draft one reply from a policy

Nothing private goes in.

  1. Copy a store's public returns policy.2 min
  2. Paste the prompt, with a made-up email asking for more.3 min
  3. Check each promise against the policy yourself.5 min

A prompt to paste (fill both blanks)

Draft a reply to the customer email below, using only the policy below. Then list every promise, date, number and name in your reply, each with the policy line it rests on. If the policy doesn't cover something, say so. Flag anything sensitive for a person. Don't send anything.

Policy: [paste]

Customer email: [paste]

Good looks likeEvery promise points to a policy line.

In your assistant · as of Oct 2026 · Claude, ChatGPT, Gemini

Set up your reply desk

Menus move often. Use the tool your company approves for customer data.

  • CheckYour current policy
  • CheckReplies you'd resend
  • CheckYour work account

Customer emails hold personal data: only in the tool your company approves for it. AI drafts; a person sends.

Claude

A project for the files, a skill for your style.

Add

Make a project. Add your policy and three past replies to its knowledge. Then write its instructions: drafts only, from these files.

Style
CustomizethenSkillsthen+thenCreate skill

Optional: your team's reply style as a skill. Needs Code execution and file creation on.

Send

Gmail connected? Claude asks before each send, reply and forward by default. Keep it that way.

At an Outlook company, Claude for Outlook (beta; Pro, Max, Team and Enterprise) drafts replies unsent in Outlook, and you click Send. Ask IT.

ChatGPT

A project for the files and the rules.

Add
New project

In the sidebar. Add your policy and past replies as files.

Rules
•••thenProject settings

Instructions here override your general ones. Say: drafts only, from these files.

Brake
Always askAllow read actionsAllow low-risk actionsAllow all actions

For a connected Gmail or Outlook. Unless you or your workspace set another, Allow low-risk actions applies.

Files per project: Free 5; Go and Plus 25; Pro, Business and Enterprise 40. On Zendesk? Its plugin (beta) reviews tickets and customer history and prepares replies.

Gemini

A notebook for the files and rules, a skill for your style.

Add
NotebooksthenNew notebook

In the sidebar. Add your policy and past replies with Add sources, and your rules in Notebook settings.

Style
CustomizethenSkillsthenCreate

Create with Gemini, start from a template, or upload a SKILL.md. Use it with /.

Mail
Ask Gemini

At the top right of Gmail. Google's page describes drafting replies, not sending them for you. Needs an eligible Workspace or Google AI plan.

Skills need you to be 18 or over, with Keep Activity on. On a work account, some features depend on your licence, which your admin manages.

Quick check · three questions

Did it stick?

  1. 1You answer the same kinds of customer question every week. Where should the policy and your best replies live?

    The answer: C A project puts the same files and rules in front of every chat inside it, so each draft starts from them. A skill carries your reply style.

  2. 2A draft reply offers a full refund. Your policy offers an exchange. What do you do?

    The answer: A Customers hold you to what a reply promises, even a softened one. The policy decides, or the person who can approve an exception.

  3. 3Which of these should a person own, with AI only helping to draft?

    The answer: B Stakes decide, not length or tone. Health, legal threats, vulnerable customers and real anger need a person's judgement.
0 of 3 answered
Keep the binder currentSide note
When a policy changes, swap the file in your project that day and take the old one out. With two versions in, it may quote the wrong one.
Write your hard-case listSide note
Agree as a team what always goes to a person: legal threats, health, a customer who seems vulnerable, real anger, the press. Put the list in the project's instructions too.
A rule in the chat isn't a gateSide note
The UK's National Cyber Security Centre advises against relying on prompting alone to control an agent. A typed rule helps. Your read before Send is the gate.
Sources, checked 18 Oct 2026Source

Product details as of Oct 2026, each checked on the vendor's own page on 18 Oct 2026. Recheck by 18 Nov 2026. The examples in the pictures are made up for this lesson.

Free skillCheck a draft against its sourcesBring this to your teamWorkflow automation · Get in touchBack to level 5, Office work, by task

Your policy in. A person sends.

Next lesson · 15 minHow do you plan a project or a week with AI?