How to Use AI to Write Work Emails

Most advice about AI and email assumes you start with a blank page and ask for a draft. The largest study of how people actually use these tools says that is not what happens.

In How People Use ChatGPT, an NBER working paper analysing roughly 1.5 million messages from 130,000 users between May 2024 and July 2025, writing was the single most common work activity — around 40% of work-related messages. But the useful finding is buried one level down: two-thirds of those writing messages asked the model to edit, critique, translate or otherwise modify text that already existed. For the average professional, the tool is not a ghostwriter. It is an editor.

That reframes the whole task. If two-thirds of the value is in revision, then the highest-leverage thing you can do is stop asking for drafts and start handing over yours.

Give it your bad draft instead of your good brief

Compare two ways of asking for the same email.

The blank-page version: Write an email to a client explaining that the project will be two weeks late.

The revision version: Here's my draft. Tighten it, keep my voice, and don't add any commitments I haven't made: 'Hi Sam — bad news on timing, the integration took longer than we thought and we're looking at two weeks past the date we agreed. Still confident on scope. Can we talk Thursday?'

The second produces better output for a reason that has nothing to do with prompt engineering tricks. Your rough draft already contains the things the model cannot know: that Sam is the kind of client you can be blunt with, that scope is the reassurance that matters here, that Thursday is when you're free. The blank-page version forces the model to invent all of that, and inventing it is exactly where AI email goes wrong — it produces something fluent, generic, and subtly not yours.

Rough notes work as well as a draft. Three fragments and a tone instruction beat a paragraph of careful briefing.

Replace professional with dials

Make it professional is the least useful instruction in common use, because professional means something different in a law firm and a games studio. It reliably produces the same beige register regardless of context.

Specify the dimensions separately instead:

• Formality — first names or titles, contractions or none
• Warmth — how much acknowledgement before the substance
• Directness — ask in the first line, or build to it
• Length — a hard ceiling, in sentences, not concise

Formal in structure, warm in tone, very direct about the ask, four sentences maximum is four constraints the model can actually satisfy. Professional is zero.

The length constraint deserves special attention. Models pad. Ask for concise and you get five paragraphs of well-organised padding; ask for four sentences maximum and you get four sentences. Numbers work where adjectives don't.

The guardrail that matters most

There is one failure mode that turns an AI-drafted email from unhelpful into damaging: inventing commitments you have not made.

Ask for an apology email about an outage and a model will happily produce we've implemented additional monitoring to ensure this doesn't happen again — a claim about your infrastructure that may be false, in writing, to a customer. Ask it to decline a request and it will offer let's revisit this next quarter when you have no intention of revisiting it. Ask for a status update and it will fill an unknown date with a plausible one.

The instruction that prevents this is short, and it belongs in every email prompt where the stakes are real:

Do not invent dates, remedies, commitments, or facts I haven't given you. If something load-bearing is missing, mark it [NEED: ...] and leave it for me.

The bracket convention matters more than it looks. Without it, a model resolves missing information by guessing, because producing complete-looking text is what it was trained to do. With it, gaps arrive visible and unmissable in the draft rather than invisible in a sent email.

This is the same discipline behind every prompt in our business communication pack — every one accepts either rough notes or a draft to revise, takes explicit tone dials rather than vague register instructions, and refuses to promise anything you have not confirmed.

Where AI genuinely outperforms you

Three cases where the model is not just faster but better:

The email you're too annoyed to write. Not because it writes it better, but because it writes it without the edge you'd leave in. Paste your furious draft, ask for the same content with the temperature removed and the substance kept, then check that the substance actually survived.

The message that needs three versions. Declining an invitation warmly, neutrally, and firmly are three different emails, and picking between drafts is much easier than getting one right in the abstract. Ask for all three.

The long thread nobody read. Summarising fourteen messages into what was decided, what's outstanding and who owes what is genuinely tedious and genuinely mechanical. This is the strongest email use case there is, and it goes underused because it isn't what AI writing sounds like.

Where it doesn't

Email carrying real relational weight — a hard performance conversation, a resignation, a genuine apology for something you got wrong — is not a drafting problem. The words are hard because the situation is hard, and smoothing the language usually makes it worse rather than better. A model can pressure-test your reasoning before you write. It should not write it.

The same is true of anything a lawyer would want to see first. Fluent text reads as considered text, and that is precisely the risk when the content is a commitment.

The short version

1. Hand over a draft or rough notes; don't start from a blank page — that's two-thirds of the observed value
2. Use discrete dials — formality, warmth, directness, a sentence count — instead of professional
3. Ban invented commitments explicitly, and require [NEED: ...] for gaps
4. Ask for variants when tone is the hard part
5. Use it hardest on thread summarisation, which is the least glamorous and most reliable win

The pattern underneath all five: the model supplies fluency, and you supply the facts, the relationship and the commitments. Every failure mode above is a case of letting it supply one of yours.