How Much AI Help Is Too Much in a Message From a Manager?

You are a manager, you have a congratulations note to write, and the AI draft is sitting right there. It is polished, warm, and done in four seconds. Your own version would take twenty minutes and probably be worse. So why does sending the draft feel slightly off?

Research on exactly this question suggests the feeling is tracking something real. The issue is not whether you used AI. It is how much of the message your team believes came from you.

What The Research Actually Found

A study by Peter Cardon (University of Southern California) and Anthony Coman (University of Florida), published in the International Journal of Business Communication, showed professionals manager emails with the AI-written or AI-edited portions highlighted, at low, medium, and high levels of AI assistance. The University of Florida’s summary describes the headline result: AI-assisted messages were generally rated as clear, efficient, and professional, but trust in the sender dropped as AI involvement grew.

Press coverage, including HR Dive’s write-up, reports that more than 8 in 10 respondents saw a supervisor’s lightly AI-assisted message as sincere, while only roughly 40% to 52% said the same about heavily AI-assisted ones. Treat those figures as approximate: they come from secondary reports of a sample of about 1,100 US professionals, and I have not seen the paper’s full tables. Check the journal article before quoting a number.

The pattern is the useful part. Fixing grammar and polish carried little penalty. Letting the tool write most of the message did, and the penalty was strongest for messages meant to carry care: praise, congratulations, empathy, motivation, personal feedback.

The Line Is Between Editing And Authoring

Most of us treat “used AI” as one category. The people reading your message do not. They sort it into two:

• You wrote it and a tool cleaned it up. Nobody minds. This is spellcheck with better taste.
• A tool wrote it and you approved it. For routine information, still fine. For anything personal, the reader feels the absence of effort.

The second case is the trust problem, and it is a perception of effort, not a detection of software. A generic “Congratulations on the milestone, your hard work is truly inspiring” reads as machine-made whether or not a machine made it.

Before:

Great job on the launch everyone. Your dedication and hard work have been truly inspiring and showcase the best of our team’s collaborative spirit.

After:

The launch went out on time, and Priya, the rollback plan you wrote in week two is the reason the Tuesday bug cost us ten minutes instead of a day. Thank you. Everyone else, you also earned the quiet weekend.

The second version is better because only the writer could have written it. No amount of polish adds that; it has to start with you.

Sort Your Messages By What They Carry

A quick way to decide how much help to take:

• Information messages (schedule changes, process updates, meeting logistics): let the tool draft, then check the facts. This is where AI saves the most time and costs the least trust.
• Judgment messages (decisions, priorities, trade-offs): use the tool to tighten, but make the decision and the reasoning yourself.
• Relationship messages (thanks, praise, difficult feedback, condolences): write the substance yourself. Use the tool only to fix mistakes and trim.

If you want the longer argument for keeping some messages fully human, when not to use AI for work messages covers privacy and accountability as well as tone.

Give The Tool A Narrow Job

The practical move is to ask for editing, not authoring. Write the messy version first, with the specific detail in it, and have the tool fix the surface.

A Wrivio Context for relationship messages could say:

Edit this message for clarity and grammar only. Keep my wording, my sentence order, and every name, date, figure, and specific detail exactly as written. Do not add praise, adjectives, or sentiment that is not already in the draft. Do not make it more formal than it is.

Press Ctrl+Shift+Space, paste your draft, and compare the result with what you wrote. If the output contains a sentence you would never say out loud, delete it. For a fuller routine, see how to review AI rewritten text.

Your own voice matters more than the polish, which is why making AI emails sound like you starts from your examples rather than adjectives.

Disclosure Is A Separate Question

The research measures what readers assume when they are told, or shown, how much AI was involved. It does not settle whether you must tell anyone. That depends on your employer’s policy and, in some jurisdictions, the law, so read whether you should disclose you used AI to write an email and check your own workplace rules.

What it does suggest is a simple internal test: if you would be uncomfortable with your team seeing exactly how the message was made, you have probably let the tool author it.

Common Questions

Does using AI to write emails hurt a manager’s credibility?

Light use appears to cost little. In the Cardon and Coman research, lightly assisted manager messages were mostly judged sincere, while heavy AI authorship lowered trust in the sender, especially for messages about praise, empathy, or feedback.

Which manager messages should I write myself?

Anything whose value is the evidence that you paid attention: thanks, congratulations, personal feedback, and sensitive news. Use AI for logistics and routine updates, and only for light editing on the personal ones.

Is it fine to use AI for grammar and polish only?

Yes, according to the same research, and it removes errors that can themselves hurt credibility. Ask the tool to keep your wording and add nothing.

Do these findings apply to employees writing to peers?

The study focused on manager messages, so do not assume the exact numbers carry over. The underlying logic, that readers discount effort they cannot see, likely applies, but that is an inference rather than a finding.

Download Wrivio for Windows to tighten your own drafts in a single keystroke without handing over the parts that have to sound like you.