AI email writing can make a workplace message sound warmer or more polished without making the recipient more likely to open it or reply. That is the useful warning from a field experiment involving 121 employees at six companies. Researchers examined 16,880 emails during a three-week crossover study, comparing messages people wrote unaided with those rewritten by an AI tool in either a playful or professional style. Neither AI treatment produced a direct statistically significant improvement in opening, replying or response time.
It is a familiar disappointment for anyone who has spent ten minutes asking a chatbot to perfect a message that goes unanswered. A more elegant sentence does not necessarily make the subject urgent, relevant or worth the recipient’s time. The study does suggest that emotional tone is connected with engagement, but that is not the same as proving that switching on an AI rewrite button causes somebody to reply.
What the AI email writing experiment actually tested
The researchers used a randomised crossover design. That means participating employees experienced different writing conditions over the study rather than having one group permanently assigned to AI and another to ordinary writing. The conditions were unaided email composition, an AI rewrite intended to be playful and an AI rewrite intended to be professional. The approach helps compare behaviour within the same broad population while retaining features of real workplace communication.
The resulting messages were not all identical except for a decorative flourish. A playful rewrite could change the warmth of a request; a professional rewrite could make it cooler or more formal. The researchers found that the playful condition raised measured positivity and the professional condition lowered it. These were changes to how the emails were expressed, not evidence that the information being sent had become more relevant or accurate.
The headline outcome was more restrained. Neither AI-assisted condition showed a direct change in the likelihood that an email would be opened, that it would receive a reply, or how long the reply would take. This is a field experiment, which gives it more practical texture than an isolated questionnaire, but it remains one study of particular employees and organisations. It does not establish that every type of AI writing assistance is ineffective.
It is also important to distinguish the effect of the tool from the effect of the final wording. The authors report that more positive wording from the same sender was associated with higher engagement. Their analysis suggests a possible indirect route through tone. That finding is interesting, but the correlation between positivity and response behaviour should not be recast as a guaranteed recipe for generating replies.
A polished email is not necessarily a persuasive email
Consider an ordinary request for a missing invoice. One version might be immaculate, formal and full of reassuring phrases. Another may be short, clear and give the recipient the account name and the date needed. The second may be easier to act on even if the first sounds more professional. This is an illustration of the underlying problem, not a scenario the researchers individually tested.
The recipient has reasons of their own for responding. They may be away, waiting for an answer, unable to help or unconvinced that the requested action belongs to them. No amount of wording can remove all those obstacles. AI can refine a message but cannot automatically create a legitimate reason for the other person to care.
Other research points to limitations that go beyond grammar. A separate 2025 study of business refusal emails found AI-generated texts could be formulaic and less nuanced than human writing. That work assessed a particular kind of correspondence and did not replicate the workplace reply experiment. Together, the studies caution against confusing smooth prose with a successful human exchange. The field experiment itself was circulated as a preprint, and further independent testing would help establish how widely its results apply.
One practical danger is that people begin optimising an email for an imaginary score rather than for an actual colleague. The model may suggest adding pleasantries to a time-sensitive request, removing useful directness or introducing a familiarity that feels out of place. Those changes might be harmless in an internal note, but an artificial tone can be jarring when a person is chasing payment, discussing a mistake or dealing with a sensitive complaint.
When sounding warmer helps and when it gets in the way
A more positive tone has plausible advantages. An email that acknowledges someone else’s workload may be easier to receive than one that sounds impatient. A useful thank-you can communicate genuine appreciation. The experiment’s association between positive language and engagement fits those ordinary observations, but its findings do not justify inventing warmth or pretending that a bad situation is cheerful.
There is also a difference between a friendly email and a vague one. A request can be courteous while stating exactly what is needed, who should do it and what deadline applies. AI tools are often good at proposing a diplomatic formulation. The human sender still needs to check that the result preserves the practical ask, especially when several people have responsibility for different parts of a project.
LiveAIWire has examined how large organisations are introducing AI into everyday correspondence. When writing help becomes a routine workplace feature, the interesting question shifts from whether the software produces grammatical sentences to whether the communication actually works. A fluent output may save drafting time while leaving the underlying bottleneck untouched.
A separate concern is the recipient’s experience. If every work message becomes smooth, lengthy and similarly enthusiastic, the distinctive signals that used to carry useful meaning may fade. A concise human note that mentions a specific document can be more valuable than a generic paragraph of professionally worded reassurance. This is an editorial interpretation rather than a result measured in the study, but it is a sensible question for teams adopting automatic rewrites.
How to use a chatbot without losing your own voice
The strongest use of AI may be as an editor rather than a stand-in author. A sender can provide the intended outcome, essential facts and relationship context, then ask for a clearer version. Before sending, they should check whether the request is still specific, whether the tone feels appropriate and whether the message contains any claims or commitments that the machine has added without permission.
It is worth comparing a short original with the rewritten version rather than assuming the longer answer is superior. An effective edit should make the next action easier to understand. If a paragraph adds nothing the recipient needs, it can often go. If the message asks for a decision, place that decision where it can be seen rather than burying it under a grand introduction.
People who use workplace AI should also respect their organisation’s rules for confidential material. Feeding an internal dispute, customer record or unannounced contract into a tool may create a separate risk from the wording of the final email. LiveAIWire has discussed the expanding ambitions of assistants that work across email and calendars; the more access such systems receive, the more carefully their permissions and actions need to be understood.
For managers, the temptation is to measure success by the volume of messages drafted or the time saved. Those measures do not answer whether colleagues understand requests, take appropriate action or feel overwhelmed by automated correspondence. Response quality and avoidable follow-up may offer a more useful picture of whether the communication process is getting better.
The workplace burden can even shift rather than disappear. If AI enables people to send more emails, their colleagues may have more reading and replying to do. LiveAIWire has reported on the tension between faster AI-assisted work and growing responsibilities. A faster composition tool may improve an individual task without reducing the number of tasks in everybody’s inbox.
There is no universal ideal amount of AI rewriting. Someone writing in a second language may gain confidence from grammar corrections, whereas an experienced colleague may only need help shortening an awkward sentence. The experiment does not establish that these groups benefit equally, nor does it explain every possible workplace relationship. That is why a sender should judge an edit against its purpose, rather than adopting a default promise to make every message warmer, more cheerful or more corporate.
The sensible conclusion is narrower than either enthusiasm or rejection. AI can be genuinely useful for tightening a clumsy message, finding a respectful tone and removing errors. But the observed experiment did not find that the simple act of rewriting emails produced a direct engagement benefit. The questions that matter most remain human ones: is the message necessary, does it say what the recipient needs, and is there a clear reason to respond?
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.