How do you rewrite a work email so it doesn't sound passive aggressive?

“Per my last email.” “Just a friendly reminder.” “Thanks in advance.” If you have a draft open with one of those in it and a feeling it reads sharper than you want, the fix is usually simple. Keep everything the email needs to get done, and delete the sentence whose only job is to show you’re annoyed. In practice that means saying why you’re writing, stating what happened, asking for one thing by a set time, cutting the jab, and closing without assuming they’ll say yes.

I co-founded Subtext, which reads a draft for exactly this before you send it, so I’ll be clear about where a tool helps and where it can’t. A passive aggressive draft is usually missing its request. The rewrite that works tends to be more direct than the original. Softening it into something vague is an easy way to get the fix wrong.

This is about your own draft. If you’re trying to work out whether someone else’s email was passive aggressive, the research on that, including full stops and emoji, is in Is that email passive aggressive, or just short? And if your email is short and firm, it may already be fine.

What makes your own email read as passive aggressive?

A sentence reads as passive aggressive when it carries irritation you could deny, usually by pointing back at something the reader failed to do. “Per my last email” is the standard example. It adds nothing to the task, since they already have the earlier email. What it adds is the suggestion that they didn’t read it.

That deniability sits close to how organisational researchers define incivility. Andersson and Pearson described it as low-intensity behaviour “with ambiguous intent to harm the target”¹. The ambiguity is what the reader takes away, and it’s the part you can’t see in your own writing.

Kruger and colleagues showed why². In the largest of their five experiments, 154 pairs of University of Illinois students tried to get sarcasm, seriousness, anger or sadness across to each other. Across email, voice and face to face, senders expected to succeed 88.8 per cent of the time and managed 70.4 per cent. The gap was widest over email, where readers got 62.8 per cent right. Being friends made no difference. You hear your intended tone while you type and can’t switch it off. These were students and constructed one-line statements, so I’d lean on the mechanism and hold the exact numbers loosely.

The phrase rankings you see quoted come from vendor polls. The best known is a Preply survey of 1,006 US adults in June 2023, run by a language-tutoring company, in which “per my last email” came top and 83 per cent said they’d received a passive aggressive email from a colleague³. That tells you which phrases people recognise as loaded. It says nothing about what the senders meant.

Brevity on its own is not the problem. In a study of real messages between people who knew each other, Pollmann and Roos found readers got the intended tone of work and study emails about right. In their companion study of text messages, length made no difference, and for the emails they had too few complete cases to test it⁴. Misreads concentrate somewhere else. Sillars and Zorn had 276 adults in New Zealand and Australia supply work emails they had found upsetting, and the recipients rated those emails more negatively than outside observers did, especially where communication was already poor and the recipient was junior⁵.

So a short email with a firm deadline is probably fine. The sentence worth worrying about is the one that keeps score. It is worth checking for. In a lab study, 84 US undergraduates doing maths problems for a fictional supervisor reported worse mood and got fewer answers right when that supervisor’s emails were uncivil⁶. It’s one small student sample, so it shows the effect can happen, and says little about how big it is at work.

These are the cues I’d look for in your own draft:

• A pointer back at their failure: “per my last email”, “as previously stated”, “as I already mentioned”, “not sure if you saw”
• A kind word doing an unkind job: “just a friendly reminder”, “just a gentle reminder”, “with all due respect”
• Thanks for something they haven’t agreed to, when it’s there as pressure
• A hedge that isn’t one: “correct me if I’m wrong, but”, “I think you’ll find”
• A comparison: “some of us”, “as everyone else seems to know”
• A lecture about the future: “going forward, please make sure”
• A manager copied in on a first, small issue
• Sarcasm: “sure, we’ll try”, “great, thanks for that”
• Frustration with no request attached

The five-step rewrite: goal, fact, ask, jab, close

Run the draft through five checks. Each one is a yes or no question, so you can do it in under a minute.

The step people skip is the ask. Most of the phrases in that list are standing in for a request that never got written. “Just a friendly reminder that this is overdue” hints. “Could you send it by 2pm?” asks.

Keep the firmness. The deadline, the number and the consequence all stay. Rewriting “The copy was due yesterday and I don’t have it” as “The date seems to have slipped” removes the blame and also removes who is responsible, which makes the next conversation harder. Across three studies of MBA students and their colleagues, Ames and Flynn found an inverted U: people rated too low or too high in assertiveness were both seen as weaker leaders. In their third study, too little assertiveness hurt results and too much hurt relationships⁷. The middle they describe is roughly what this framework aims for: a firm ask with the sting taken out.

Then read it aloud in the tone you are afraid it has, as if you were annoyed. In the fourth Kruger experiment, 54 students emailed statements meant to be sarcastic or serious. Those who first read them aloud in the opposite tone lost their overconfidence completely, while those who read them as intended stayed overconfident². It was a sarcasm task with students, but it’s the closest thing to a tested editing trick here.

That gap between what you meant and what the words carry is the part Subtext checks. It marks the exact words that could land as passive aggressive, sarcastic or loaded, explains what each one is likely to signal, and names the feeling underneath, before it offers any rewrite. On the draft “Per my last email, the file is attached.” Subtext tagged the opener Passive-aggressive and named the feeling as Frustration. Its note said the phrase is widely read as a polite way of saying “I already told you this”, which can sound irritated. The plainest of its three versions read “The file is attached to this email.” That is the app’s real output, unedited.

Five work emails, rewritten

Chasing an overdue reply

Before Hi Dan, just following up again. Per my last two emails, I need the updated figures to finish the forecast. Would appreciate a response this time. Thanks in advance.

After Hi Dan, I need the updated Q3 figures to finish today’s forecast. Could you send them by 2pm? If that won’t work, let me know by 11 and I’ll plan around a rough estimate. Thanks, Maya

The score-keeping goes: “again”, “last two emails”, “this time”. The urgency stays, with a reason, a time and a way out. “Thanks in advance” goes too, though I wouldn’t ban it. In Preply’s sign-off survey of 1,005 US adults, fielded in December 2022, it was the fourth most used sign-off and the least preferred, picked by 5 per cent⁸. In Boomerang’s analysis of more than 350,000 mailing-list threads, emails ending with it got the highest reply rate of the common closings, 65.7 per cent against a baseline of 47.5⁹. Both are vendor data, and replying to an email is a different thing from liking its sign-off. Next to a polite request it’s fine. Next to “this time” it reads as a demand.

A missed deadline

Before Hi Alex, as previously stated, the copy was due yesterday. I’m not sure why we keep having this problem, but going forward I need deadlines to actually mean something.

After Hi Alex, the copy was due yesterday and I don’t have the final version yet. Could you send it by 11am today so design can start this afternoon? If something’s blocking it, tell me this morning and we’ll decide whether to move the hand-off. Thanks

The missed deadline stays in, as a fact with a date. “We keep having this problem” judges a pattern without closing today’s gap, and “going forward” is a lecture. The consequence, design starting this afternoon, carries the urgency on its own. If there is a pattern, that’s a conversation for a call.

Pushing back on scope

Before Happy to add all of this if everyone’s fine with the launch slipping, since there obviously isn’t time to do everything.

After Hi Ana, we can add the three new screens, but they won’t fit before the 30 March launch. Two options: launch on the 30th with the agreed scope and add the screens in April, or add them now and move launch to 20 April. Which would you prefer?

The objection was sound. “Obviously” and the sarcastic “happy to” were the problem. The rewrite turns a complaint into a decision with two real options and asks the person who owns it to choose.

Correcting a colleague’s mistake

Before Hi Jo, just FYI, the revenue figure is wrong again. As I mentioned yesterday, it should be £1.8m, not £2.1m. I’ve copied Rob so everyone’s on the same page this time.

After Hi Jo, slide 14 shows £2.1m. The approved revenue figure is £1.8m, and the source is in the attached finance sheet. Could you update it before the 3pm client review? I’ve copied Rob because he signs off the final deck. Thanks

The correction is as precise as before. “Wrong again”, “as I mentioned” and “this time” are gone. Rob stays on copy only because he has a part in the decision, and the email says so. If he didn’t, I’d take him off.

Declining a request

Before Hi Priya, I’d love to help, but given this landed on me two days before the deadline, I’m not sure what I can realistically do. As you know, I already have plenty on my plate.

After Hi Priya, I can’t take this on for Thursday alongside the launch work I’ve committed to. I could look at it next Monday, or Sam may have room before then. Let me know which works.

The no is clearer here than in the original, and the complaint about late notice is gone. It gives one reason and two alternatives, then stops. When to say no at work in the first place is its own question, covered in How do you say no to extra work?

Phrases worth swapping

Treat these as options. Sometimes the original is the right call: “As discussed on Monday, the approved budget is £8,000” can be plain record-keeping in a thread where people need the number.

Does an AI rewrite cost you anything with the reader?

Sometimes, if they notice. On routine work email the studies so far find little cost from light AI help. The penalty shows up when a message that’s meant to carry warmth or feeling is mostly AI-written and the reader knows it.

The workplace study most people cite is Cardon and Coman’s, which had 1,100 working professionals rate messages written with different levels of AI help¹⁰. According to the University of Florida’s summary, only 40 to 52 per cent of employees saw supervisors as sincere when they used high levels of AI, against 83 per cent for low-assistance messages. Ratings of professionalism fell too, from 95 per cent to between 69 and 73¹¹. The study used a congratulatory message, and participants were told how much AI had been used, which real inboxes don’t do.

Sahebi, Formosa and Bankins tested human, AI-assisted and fully AI-written text with 547 participants¹². Fully AI-written messages were trusted least. In the work email condition, though, AI-assisted and human-written emails were rated about the same for trustworthiness. Purcell and colleagues ran incentivised trust games with 1,637 US participants and found AI-assisted writers wrote faster and earned about as much trust, even when the AI use was disclosed¹³. One of their two studies did show a small trust edge for writing alone, and the assisted messages scored lower on a text measure of authenticity.

The clearest line on warmth comes from Khadpe and colleagues¹⁴. In the second of their two studies, with 280 of their 399 US participants, an AI label made a thank-you or an apology read as noticeably less warm, while a brag or a message blaming someone read about the same either way. Each of these is a single paper that nobody has replicated yet, and most used strangers or hypothetical messages.

Read together, AI that helps you spot and remove a jab in a status chase looks low risk. AI that writes your thank-you or your apology is where readers mark you down. More on both in Do AI message rewriters actually work? and Does AI make your texts sound like a robot?

My advice is to use the tool to find the problem and suggest options, then keep your own words. Subtext shows how much of your draft each version changed, so you can see when a rewrite has drifted, and Manual edit lets you fix the one word that’s off and leave the rest as you wrote it.

Which AI tools can rewrite a work email?

Most of them will rewrite it. Fewer tell you what was wrong first, and for passive aggression that’s the useful part, because you want to know which sentence was the jab. Grammarly Pro is $12 a month billed yearly, or $30 billed monthly¹⁵. Checked 12 and 13 October 2026:

I co-founded Subtext, so weigh that row accordingly. Longer comparisons are in Every phone rewrites your messages now and Most tone checkers never check your tone.

If you use a general chatbot, ask it to point before it rewrites. Something like:

Here’s a work email I’m about to send. List any phrase that could read as passive aggressive and say why in one line each. Then give me two rewrites that keep every fact, date and request, add no apologies or warmth I didn’t write, and stay about the same length.

The last two instructions matter. In the Purcell study, AI-assisted messages came out warmer in tone and longer than unaided ones¹³. That suits a trust game. In a firm work email it’s how a clear request turns vague.

What should you check before pasting a work email into an AI tool?

Check whether your employer allows it and whether the email holds other people’s information. Then check whether the tool trains on what you paste, which depends on your account as much as the product.

• Conversations on personal ChatGPT accounts (Free, Plus and Pro) may be used for training unless you turn off “Improve the model for everyone”. Business, Enterprise and Edu accounts aren’t, by default. On a personal account, a thumbs up or down can send that whole conversation for training even after you have opted out. On Business, Enterprise and Edu it doesn’t²³.
• Grammarly’s training setting is on by default for individual accounts and off by default for enterprise and education accounts²⁴.
• Google says Gemini in Gmail doesn’t use your content to train its generative AI models, for Workspace and Google AI plan users²⁵. Microsoft says prompts and responses in its work Copilot aren’t used to train foundation models²⁶.
• Subtext doesn’t train on what you send, and the AI providers that process your text aren’t allowed to either. Your text goes to Anthropic for processing in the United States. Conversations delete five days after you last use them, or 90 if you pin one. Its terms license it for personal use and add: “If you use it for work, you confirm you are allowed to do so”²⁷.

People skip these checks often. In a TELUS Digital survey of 1,000 US adults who use AI assistants at companies with 5,000 or more employees, 57 per cent admitted entering sensitive information into public AI assistants, and 29 per cent said their company had a policy prohibiting it²⁸. A PagerDuty survey of 1,250 office workers in the US, UK, Australia and Japan found 43 per cent had shared emails or other correspondence with public AI tools²⁹. Both are vendor surveys built on self-reports.

If your company has not approved the tool, keep the email out of it. On an approved tool, still strip names, client details and figures the rewrite does not need. The full policy comparison is in Which AI writing assistants train on your messages?

What can’t a rewrite fix?

A rewrite changes the words. It can’t see the history between you, predict how this particular person will take it, or solve the workload or performance problem the email is about.

History and status change how the same words land. In the Sillars and Zorn study, the gap between recipient and observer was widest where communication was already poor and the recipient was junior⁵. A tool reading your draft sees none of that unless you tell it.

Culture is often summarised as low-context versus high-context, with whole countries sorted into each. It’s a useful reminder that norms differ, but the evidence for the country labels is thin. Cardon reviewed 224 articles in business and technical communication and found the contexting model had little empirical support³⁰, and Kittler, Rygl and Mackinnon found only seven of 26 studies had tested the country classifications empirically³¹. In a 2026 preprint, Sakai and colleagues found that when 175 Japanese and American employees wrote work emails with AI drafts, their emails shifted towards the draft’s style, far more when the draft clashed with their own norms³². It’s a single study and hasn’t been peer reviewed, but it suggests a rewrite can quietly change a register you chose on purpose. If English isn’t your first language, Do my English texts sound polite? covers what you can and can’t check.

Sometimes the right move is a call. If the email is the third round of a disagreement, or the relationship is already strained, voice carries tone that text drops, which is what Kruger’s voice conditions showed². Even a good tone check is only a second reading. Subtext’s safe-to-send score is its own estimate from the words in front of it, and a warning means look again at the words it marked²².

If you’d like that second reader before you send, Subtext reads your draft or a screenshot of the thread, marks the words that could land as passive aggressive, and gives you up to three versions that keep the formality you wrote in and any greeting or sign-off. You pick one and send it yourself.

Try Subtext in your browserDownload on the App StoreGet it on Google PlayGet the app

---

Think I’ve read a study wrong, or know research on this I’ve missed? Tell me on LinkedIn.

Samet Durgun is the co-founder of Subtext, an app that catches the emotional tone of your messages and rewrites them in your own voice. He’s based in Berlin.

Sources

Numbered in the order they appear above. Product features, prices and policies checked 12 and 13 October 2026. Where a figure comes from a summary rather than the paper itself, the entry says so.

1. Andersson, L. M., and Pearson, C. M. (1999). Tit for Tat? The Spiraling Effect of Incivility in the Workplace. Academy of Management Review, 24(3), 452 to 471. Definition quoted from p. 457, as later papers quote it; I could not open the original page.
2. Kruger, J., Epley, N., Parker, J., and Ng, Z.-W. (2005). Egocentrism Over E-Mail: Can We Communicate as Well as We Think? Journal of Personality and Social Psychology, 89(6), 925 to 936. The 88.8 and 70.4 figures are Study 3, with 154 pairs of University of Illinois students, pooled over email, voice-only and face-to-face conditions; email accuracy alone was 62.8 per cent. The read-aloud result is Study 4, with 54 students. Supported by a University of Illinois Board of Trustees grant and National Science Foundation grant SES-0241544.
3. Zajechowski, M., for Preply (2023, updated 3 July 2024). Survey: 83% of Americans hit by passive-aggressive emails from coworkers. Online survey of 1,006 US adults aged 18 to 76, fielded 22 June 2023. A vendor survey by a language-tutoring company; sampling and weighting are not published.
4. Pollmann, M. M. H., and Roos, C. A. (2025). “I get u”. People correctly interpret the tone of text messages and emails. Computers in Human Behavior Reports, 18, article 100689. The email study, in a work or school context, had 361 receivers and 61 matched senders. Recent and not yet replicated.
5. Sillars, A., and Zorn, T. E. (2021). Hypernegative Interpretation of Negatively Perceived Email at Work. Management Communication Quarterly, 35(2), 171 to 200. The sample of 276 adults comes from Zorn’s own summary for The Conversation. Participants chose emails they had already found upsetting.
6. Giumetti, G. W., Hatfield, A. L., Scisco, J. L., Schroeder, A. N., Muth, E. R., and Kowalski, R. M. (2013). “What a rude e-mail!” Examining the differential effects of incivility versus support on mood, energy, engagement, and performance in an online context. Journal of Occupational Health Psychology, 18(3), 297 to 309. 84 undergraduates at a US university in a within-subjects lab experiment. Supported in part by National Science Foundation grant SMA 1004413.
7. Ames, D. R., and Flynn, F. J. (2007). What breaks a leader: The curvilinear relation between assertiveness and leadership. Journal of Personality and Social Psychology, 92(2), 307 to 324. Three studies with MBA students rated by colleagues, or rating their managers.
8. Zajechowski, M., for Preply (2023, updated 24 March 2025). Study shows the most savage ways to begin or end an email. Online survey of 1,005 US adults aged 18 to 76, fielded 9 December 2022. Vendor survey.
9. Greenley, B., for Boomerang (2017). Forget “Best” or “Sincerely,” This Email Closing Gets the Most Replies. Boomerang blog, 31 January 2017. More than 350,000 threads from the public mailing-list archives of 20+ online communities, not workplace email. Vendor analysis; observational.
10. Cardon, P. W., and Coman, A. W. (2026, first published online 16 July 2025). Professionalism and Trustworthiness in AI-Assisted Workplace Writing: The Benefits and Drawbacks of Writing With AI. International Journal of Business Communication, 63(4), 1825 to 1857. Survey experiment with 1,100 working professionals.
11. Alsup, A., University of Florida News (6 August 2025). Is writing with AI at work undermining your credibility? The sincerity and professionalism percentages come from this release; the paper itself is paywalled and was not checked.
12. Sahebi, S., Formosa, P., and Bankins, S. (2026). The AI penalty and disclosure paradox: Trust, authenticity and knowledge uptake in AI-mediated communication. Computers in Human Behavior: Artificial Humans, 8, article 100304. Preregistered, 547 participants recruited on Prolific. Funded in part by Facebook through the Ethics in AI Research Initiative for the Asia Pacific, and by the Australian Research Council.
13. Purcell, Z. A., Jakesch, M., Dong, M., Nussberger, A.-M., and Köbis, N. (2025). Writing with AI boosts trust-building efficiency. iScience, 28(12), article 114092. Two preregistered experiments with 1,637 US participants. The paper states Study 2’s sample as 971, but its two groups add up to 941, which matches the overall total.
14. Khadpe, P., Wenzel, K., Loewenstein, G., and Kaufman, G. (2025). Explaining the Reputational Risks of AI-Mediated Communication: Messages labeled as AI-assisted are viewed as less diagnostic of the sender’s moral character. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 8(2), 1413 to 1423. Two vignette studies with 399 US participants.
15. Grammarly. How much does Grammarly Pro cost? Grammarly Support. $144 a year, an average of $12 a month, or $30 billed monthly.
16. Google. Draft emails with Gemini in Gmail. Gmail Help.
17. Microsoft. Get email coaching with Copilot in Outlook. Microsoft Support.
18. Microsoft. Where can I get Microsoft Copilot? Microsoft Support.
19. Microsoft. Frequently asked questions about Copilot in Microsoft 365 subscriptions. Microsoft Support, updated 9 September 2026.
20. Grammarly. Grammarly Prices and Plans.
21. Apple. Use Writing Tools to write, improve, and summarize content on Mac. Apple Support. The iPhone guide lists the same tones.
22. Subtext. Subtext app: AI writing assistant for texts and replies. Company page; no independent evaluation of Subtext’s accuracy exists.
23. OpenAI. How your data is used to improve model performance. OpenAI Help Center.
24. Grammarly. Product Improvement and Training Control. Grammarly Support.
25. Google. Learn how Gemini in Gmail, Calendar, Chat, Docs, Drive, Sheets, Slides, Meet, Vids & Pics protects your data. Gmail Help.
26. Microsoft. Data, Privacy, and Security for Microsoft Copilot. Microsoft Learn, last updated 18 August 2026.
27. Subtext. Subtext | Terms, Privacy and your account. Terms last updated 26 July 2026; privacy policy last updated 10 October 2026.
28. TELUS Digital (26 February 2025). TELUS Digital Survey Reveals Enterprise Employees Are Entering Sensitive Data Into AI Assistants More Than You Think. Pollfish survey, January 2025, of 1,000 US adults at companies with 5,000 or more employees who use an AI assistant at work. Vendor survey.
29. PagerDuty (11 June 2026). PagerDuty Report Finds Two-Thirds (66%) of Office Professionals Have Used Unauthorized AI Tools at Work. Wakefield Research survey of 1,250 office professionals outside IT roles, at organisations with $500 million or more in revenue: 500 in the US and 250 each in the UK, Australia and Japan. Vendor survey.
30. Cardon, P. W. (2008). A Critique of Hall’s Contexting Model: A Meta-Analysis of Literature on Intercultural Business and Technical Communication. Journal of Business and Technical Communication, 22(4), 399 to 428. The abstract calls it a meta-analysis of 224 articles; I read only the abstract.
31. Kittler, M. G., Rygl, D., and Mackinnon, A. (2011). Beyond culture or beyond control? Reviewing the use of Hall’s high-/low-context concept. International Journal of Cross Cultural Management, 11(1), 63 to 82.
32. Sakai, S., Gao, A., Shoda, Y., and Reinecke, K. (2026). AI-Generated Email Drafts Shift Culturally Distinctive Communication Styles in Professional Email. arXiv preprint, posted 22 September 2026. Not peer reviewed; 175 full-time employees in the analysis.