Inbox and correspondence: triage, replies and chasing with AI

Email is where an office day leaks away. Sixty unread messages, three threads that need a decision from someone else, a supplier who has gone quiet and one reply you have been putting off because it is awkward. AI helps with most of this. It sorts, summarises and drafts quickly. It does not know your relationships, your organisation's history or what your manager would actually agree to, so the judgement stays with you.

This guide follows Maya Roberts, Office Manager at Fernway Group (our fictional practice company), through a Monday morning inbox. Fernway sells and services office equipment and supplies to businesses across the South and Midlands.

First decide where the email lives

The tool you use changes what is possible and what is safe.

An assistant built into your mail. Copilot in Outlook and Gemini in Gmail can read the thread you have open, and sometimes search your mailbox, without you copying anything. It usually needs a paid plan, and for work mail, a work account your organisation has licensed. When you are signed in with a work account, your organisation's permissions and data terms apply. If you don't see the option, the most likely reason is the licence, not a fault. The detail lives in Copilot in Outlook and Teams and Gemini in Gmail and Docs.

A general chat assistant. ChatGPT, Claude, Gemini or Copilot Chat in a browser only see what you paste. That is fine for your own drafting and for material with nothing sensitive in it. It is the wrong place for a customer's complaint with their account details, or an HR thread, unless your organisation has approved that tool for that kind of information.

A connector. Some assistants can connect to a mailbox. Connecting a work mailbox to any AI tool is a decision for IT and your data protection lead. Ask before you switch anything on.

Your organisation's policy and approved tools still take priority over anything in this guide.

Worked example: Maya's Monday inbox

Maya doesn't paste 58 whole emails anywhere. She copies the sender, the subject and the first line of each unread message into a list. That gives the tool enough to sort, and keeps most of the detail out. Where a first line contains something sensitive (a name in a sickness absence email, a customer's account problem), she replaces it with a short label such as "HR matter, see inbox".

I'm the office manager for a UK company that sells and services office equipment. My managing director is in meetings all morning and I'm covering anything urgent for her.

Below is a list of unread emails: sender, subject and first line. Sort them into four groups: needs a reply from me today; I can reply with a short standard answer; for information only; can wait until later this week. Put anything that might need the managing director in a separate group at the top.

For each email give one short line saying why it's in that group. If you can't tell from the subject and first line, put it in a group called "open to check" rather than guessing.

[paste list]

The reply comes back in the groups Maya asked for. Two items sit under "might need the MD": a customer asking to move a contract meeting, and a note from the landlord's agent about the lease. Eleven are for information only. Four land in "open to check" because the first line said nothing useful.

This is where Maya's own knowledge matters. The tool put a message from Ravi Menon in IT under "can wait". Maya knows Ravi is chasing the laptop imaging for three new starters who begin next week, so she moves it up. The list is a starting point. She spends two minutes adjusting it, then works from it.

She follows up in the same chat:

From the same list, which emails look like someone is waiting on me for something I promised? Quote the words that suggest it. Don't include anything where it's unclear.

Drafting replies that sound like you

A generic AI reply is easy to spot: too long, too formal, full of phrases you would never use. You fix this by giving the tool something to copy.

Keep two or three of your own sent emails that you think read well, with names and details removed. Paste them in as a style sample when you draft. Tell the tool what you want to say in your own rough words, and let it tidy rather than invent.

Here are some emails I've written before, so you can match my style: short, friendly, no jargon, no exclamation marks.

[paste two or three of your own emails, with names and details removed]

Now draft a reply to the email below. What I want to say: [your points in rough words]. Keep it shorter than the original. If you need a fact I haven't given you, leave a gap in square brackets instead of making one up.

[paste the email you are replying to, with anything sensitive removed]

If you draft on behalf of someone else, as Maya sometimes does for Erin, work from a few of their sent emails instead. Anything that commits them to a date, a cost, a position or an apology goes to them for approval before it is sent, however polished the draft looks.

Chasing without nagging

Chasing is repetitive and slightly uncomfortable, which makes it a good fit for a draft. The tool needs to know what was asked, when, what happens if it doesn't arrive, and how firm to be.

Draft a short chaser email. Context: on [date] I asked [role, for example "our cleaning contractor"] for [what you asked for]. I haven't had a reply. I need it by [date] because [reason]. This is my second chase.

Give me two versions: one friendly and light, one firmer that still sounds polite. Each under 70 words. Make the ask and the deadline impossible to miss. No guilt-tripping.

Keep a note of what you chased and when. The tool won't remember the history for you unless you are using a saved project or assistant, and even then your own tracker or the sent folder is the record.

Tone for difficult messages

Complaints, refusals, pushback from a senior colleague, a message that needs to say no: these are where people most want help, and where AI needs the closest watching. Its default is either stiff corporate language or over-apology. Neither helps.

Two ways to use it well:

Check the tone of your own draft. Write what you actually mean first. Then ask the tool how it might land with the reader, and what to change. You keep the content and your judgement; it points out the sentence that reads as sharp.

Draft from clear facts. If you want a first draft, give it the facts, what you can and can't offer, and the reader's situation. Tell it what to avoid: "no apology for things we didn't do", "don't promise a refund".

I need to send the message below to [who they are and the relationship, for example "a long-standing customer who is annoyed about a late delivery"]. I've written what I mean.

Read it as that person would. Tell me which sentences could sound defensive, cold or unclear, and suggest a replacement for each. Don't change the facts or what I'm offering. Don't make it longer.

[paste your draft]

For a full worked version with a real complaint, try Reply to a difficult complaint.

What not to paste

An email often carries more than it seems. Before anything goes into a tool that isn't approved for it, look for:

• People's personal details: home addresses, phone numbers, dates of birth, anything about health, family or absence.
• HR matters of any kind: performance, grievances, disciplinaries, pay.
• Customer account details, contract terms and pricing that isn't public.
• Anything marked confidential, and anything a sender would not expect to leave your organisation.
• Passwords, security codes and bank details. These never go into any AI tool.

Health and some other kinds of information are what UK data protection law calls special category data, which needs extra protection. Removing a name doesn't always make something anonymous: "the only person in accounts who was off for three weeks in March" identifies someone in a small office. If you are unsure, the Can I put this in ChatGPT? tool walks you through the decision. Your organisation's policy and approved tools still take priority, and your data protection lead has the final say.

What to check before you send

For triage lists, spot-check a few of the "can wait" items. Those are the ones where a wrong call costs you.

Common mistakes

• Trusting the triage. The tool judges urgency from words, not relationships. It doesn't know the quiet email from your biggest customer matters more than the loud one from a newsletter. Glance over every group before you act on it.
• Letting drafts grow. Ask for "a reply" and you often get three paragraphs for a one-line answer, with the key fact buried. Set a word limit in the prompt. The email that got longer shows how easily it happens.
• Accepting invented detail. If you didn't give it a date, a price or a name, it may supply a plausible one. Tell it to leave gaps, then check.
• Pasting whole threads by habit. You usually need the last message and a line of context, not the full chain with everyone's signatures and phone numbers.
• Sending on autopilot on someone else's behalf. A draft that reads well can still commit your principal to something they didn't agree. Approval comes from them.
• Using the wrong account. A personal ChatGPT or Gemini account on a work laptop is an easy way for work data to leave the organisation. Use what your organisation has approved.