Ask an AI assistant to “write an email to finance asking for the budget” and you may get three polite paragraphs with a blank where the deadline should be. Ask for “two sentences to Priya in finance: I need the Q3 travel budget by Friday noon for Monday’s board pack” and you have given it what a draft needs; one careful read may be all it takes before you send.

That difference is most of the skill. Using AI for writing emails works when you bring the substance and the assistant brings the wording. You decide who the reader is, what you need from them, by when, and which facts matter; the assistant turns your rough notes into a short draft with the request first; you check every fact, promise and tone before anything is sent. In five steps:

1. Pick the routine emails you send every week
2. Write the facts and the ask in rough notes first
3. Tell the assistant who will read it and what they should do
4. Swap names for roles and leave confidential details out
5. Read the draft against your notes, then as the recipient

How much time AI for writing emails saves

In a randomized trialrandomized controlled trial: A study that assigns participants to the treatment or the comparison group at random, so the groups start out alike and a difference in what happens next can be put down to the treatment. Random assignment makes the two groups comparable; it does not make the people in the trial representative of anyone else.Full entry in the glossary that Microsoft ran at large firms in 2023 and 2024, office workers who used an AI assistant built into their email spent less time on email each week, and the workers offered it still replied to as many threads.¹ The study, forthcoming in the American Economic Review: Insights, measured time, not the quality of what was sent.

The study

Moderate evidence

Did an AI email assistant save time at 66 large firms?

Microsoft recruited large firms that had bought early licenses for its Microsoft 365 Copilot assistant, and each firm gave access to a randomly chosen group of the workers it had enrolled, while their colleagues kept their usual tools. In the second half of the six-month trial, workers who used the assistant spent about 2 fewer hours a week on email than comparable colleagues. Across everyone given access, whether they used it or not, workers replied to as many email threads as their colleagues, and no more slowly.¹

Over a five-day week, the users’ saving is under half an hour a day: modest, but it recurs. Three of the four authors work for the company that makes the tool, the firms that signed up may adopt new technology more readily than most, and because the assistant could also summarize threads, some of the saving may have come from reading less rather than writing faster.¹

The authors read the result as the same amount of email work done in less time.¹ In an ordinary week, that means small savings on each message you write again and again: confirming a meeting, chasing a document, reporting progress. The UK government’s AI Playbook, written for public servants in 2025, likewise lists first drafts of routine email responses among the jobs AI can take on.² What this means for you: try the assistant first on the three or four emails you write most often, such as a weekly status note or a meeting request, not on the hardest message in your drafts folder.

Further reading

• 

  Writing for Busy Readers: Communicate More Effectively in the Real World

  by Todd Rogers and Jessica Lasky-Fink

  Rogers and Lasky-Fink's book on the short-request experiment cited here, with rules for writing messages busy readers act on.

As an Amazon Associate WiserHours earns from qualifying purchases.

Why the facts have to come from you

An AI assistant can only work with what you give it, so the facts in an email have to come from you. The UK government’s AI Playbook notes that AI systems lack contextual awareness and generate text by returning the most likely output, which can sound plausible and still be wrong.²

In an email, the most likely output is a generic one. The assistant does not know that the budget figure feeds Monday’s board pack, so a thin prompt comes back fluent but vague, with a placeholder or, worse, an invented date or promise, a failure covered in why AI assistants make confident errors. The Priya prompt at the top of this page works because it hands over the name, the ask, the deadline and the reason, and leaves nothing to invent.

Length is your call too. Harvard researchers Todd Rogers and Jessica Lasky-Fink, authors of Writing for Busy Readers, describe emailing thousands of US school board members a request to fill in a short survey: a version less than half as long drew nearly twice as many responses as the wordier one. They suggest some readers judged the effort by the length and skipped the long email, while others stopped reading before they reached the request at the end.³

That is one experiment with strangers, and it appears in the authors’ own magazine article and book, not in a peer-reviewed paper we could find. The lesson still carries over to a colleague’s inbox: tell the assistant how long the email may be and where the request goes.

1. Who will read it, and what they already know
2. The one thing you need from them
3. By when, and why that date matters
4. The facts only you know: figures, names of documents, what was agreed
5. What comes back: the request in the first lines, background last, and short enough to read at a glance

Prompts to copy for requests, updates and follow-ups

Each of the three prompts below asks for the request or the news first, a length limit and a subject line, and forbids facts you did not supply. Replace the words in square brackets with your own notes, and paste only what the email needs.

A prompt for a request

A request works when the reader sees within seconds what you need, by when, and why it matters to them.

Filled in for the finance email, the first line becomes “Write a short email to Priya, finance lead. I need the Q3 travel budget by Friday at noon because it goes into Monday’s board pack.”

A prompt for a status update

An update is read for its bottom line: is the work on track, and does the reader need to do anything? The prompt asks for that answer first and background last, the order US plain-language guidance recommends.⁴

A prompt for a follow-up when nobody has replied

Silence can simply mean a crowded inbox. In a 2015 study of more than 2 million Yahoo Mail users who wrote to each other regularly, most replies that came at all arrived within a day, and on days when people received more email, they replied to a smaller share of it.⁵ The data come from one email service more than a decade ago and count only replies that arrived, so they hint, rather than prove, that a late answer is unlikely to turn up unprompted. A short nudge after a couple of working days, easy to answer from a phone, is our rule of thumb, not a tested one.

Five steps from rough notes to a sent email

Five habits turn rough notes into an email worth sending: pick routine messages, write the substance yourself, brief the assistant, strip out anything confidential and read the draft as its recipient would.

1. Pick the routine emails you send every week

List the emails you write every week: meeting requests, reminders, status notes, answers to the same few questions. A checked draft pays off most on these, and a slightly plain tone does them little harm.

2. Write the facts and the ask in rough notes first

Before you open the assistant, jot down who the email is for, the one thing you need, the deadline and the facts it rests on. Fragments are fine; the assistant was not in last week’s meeting. Keep to one ask per email: a second request tacked on at the end risks the fate of the survey link at the bottom of the wordy email, which, the researchers suggest, some readers never reached.³

3. Tell the assistant who will read it and what they should do

Brief the assistant as you would a new colleague: the reader’s role, what they already know, what you want them to do, and the tone. US federal plain-language guidance starts from the same questions about the audience and advises stating the purpose and bottom line first, with background toward the end.⁴ Give a length limit it can follow, such as one short paragraph or five bullet points.

4. Swap names for roles and leave confidential details out

A forwarded thread carries names, email addresses, phone numbers in signatures and often details about clients or colleagues. In 2024 the Dutch data protection authority said it had received a number of data breach notifications caused by staff entering personal data of patients or customers into AI chatbots. Most chatbot companies store what is entered, it notes, and when staff do this on their own initiative, against their employer’s rules, it counts as a data breach.⁶

That is the Dutch regulator’s reading of EU data protection law; wherever you work, your employer’s rules apply. Use the assistant your employer has approved, replace names with roles (“the client”, “Supplier A”), paste only the part of the thread the reply needs, and leave out health, pay, performance and financial details.

5. Read the draft against your notes, then as the recipient

Check names, dates, figures, any attachment the draft mentions, and any commitment you never made, such as a delivery date. The UK AI Playbook tells public bodies that a person must review an AI tool’s irreversible actions, such as sending an email, and the rule suits any outbox.² Then read it once as the recipient, because tone is hard to judge from your own side of the screen, a problem explained in how email senders misjudge their tone.

WiserHours

Before an AI-drafted email goes out

• It is a routine email I send often
• The facts, the ask and the deadline came from my own notes
• The assistant knew who the reader is and what they should do
• Names and confidential details were removed before pasting
• Every name, date, figure and promise checked against my notes
• Read once as the recipient, for tone

Find this checklist online:
https://wiserhours.com/ai-productivity/ai-for-writing-emails/#checklist-ai-for-writing-emails-steps

Readers judge the sender, not only the email

Readers can find an AI-assisted message professional and still think less of the person who sent it. In a 2025 survey experiment with 1,100 working professionals, Peter Cardon and Anthony Coman found that when a manager’s message congratulating a team had medium to heavy AI help, readers began to question the manager’s sincerity, caring, confidence and ability.⁷

We read only that study’s abstract, which does not say how readers learned how much AI had been used. The authors describe such a message as a routine communication task, so the risk is not limited to sensitive messages.⁷ Disclosure does not remove the cost: across 13 experiments published in 2025, Oliver Schilke and Martin Reimann found that people who disclosed using AI were trusted less, a drop they trace to a sense that relying on AI was less legitimate, though being exposed by someone else cost more trust.⁸

Nor does hiding it. In a 2023 Scientific Reports experiment, online workers who suspected a chat partner of using AI-suggested replies rated that partner less cooperative, although their guesses tracked actual use only weakly and actual use made partners seem more cooperative, not less.⁹

Picture a team lead who asks an assistant to “congratulate the team on the launch” and gets a warm note that could go to any team. Two sentences naming what each person did carry what a thank-you exists to show: that you noticed. So match the help to the message.

Where your employer or profession has a rule on disclosing AI use, follow it. Refusals, apologies and bad news need more care than a template, and are planned as a separate guide in the AI for personal productivity category.

How much weight each finding can bear

The firmest evidence here is one large randomized trial of time saved, run by the tool’s maker. The writing advice rests on one experiment and government guidance, and the findings on readers’ reactions come from online and survey experiments, two of them read only in abstract.

The bottom line

Hand an AI assistant the emails you send every week, and keep the substance yours: the facts, the ask, the deadline and the final read. Ask for short drafts with the request first, take names and confidential details out before you paste, and write anything personal yourself.