Cut Stock Phrases from an AI Email Without Losing the Message
Here is an email Gemini wrote when asked to tell colleagues that the shared drive was available again. The crossed-out words are my suggested cut; everything else is the model’s wording, including the blanks for the sender.
Subject: RESOLVED: Shared Drive is back up and running
Hi team,
The issue with the shared drive has been resolved, and it is now fully accessible again.
Thank you for your patience
while we worked to get this fixed. Please let IT know if you experience any further issues.Best regards,
[Your Name]
[Your Title]
The cut removes a second reference to fixing the problem. It leaves the thanks, the useful instruction about further trouble and the ordinary sign-off. You could make this email shorter still, but there is no need to delete every familiar phrase to make it worth sending.
There is another edit a phrase list would miss: “we worked” puts the sender on the team that fixed the drive. Our prompt never said who the sender was. If you are passing on IT’s news, do not adopt that claim just because the sentence reads smoothly.
Which familiar phrases actually appeared
Before generating any emails, we saved ten phrases we expected to see, in an expected order. “I hope this email finds you well” came first in that prediction; “Best regards” came last. The results put the sign-off far ahead. The famous opening appeared in only six of the 100 emails.
The prompts covered ten work situations, ten requests each: asking for something, following up, apologising, refusing, thanking, arranging, updating, giving instructions, reporting a problem and making an announcement. Each request produced one response from Gemini 3.5 Flash. We counted the first email in each response, once per phrase per email.
For perspective, we applied the same searches to 79,569 distinct sent emails extracted from the Enron email collection released by Carnegie Mellon. These are mostly from 2000 and 2001, at one company. They show an older workplace’s usage, not what people write now.
| Phrase | Generated out of 100 |
Historical out of 79,569 |
|---|---|---|
| I hope this email finds you well | 6 | 0 |
| Please let me know if you have any questions | 6 | 205 |
| Thank you for your understanding | 7 | 0 |
| I wanted to reach out | 0 | 0 |
| Please don't hesitate to contact / reach out | 1 | 4 |
| I look forward to hearing from you | 6 | 83 |
| I am writing to | 19 | 17 |
| at your earliest convenience | 2 | 35 |
| Thank you for your time | 3 | 9 |
| Best regards | 79 | 138 |
The denominators matter: 79 generated emails means 79%; 138 historical emails means about 0.17%. Neither figure makes “Best regards” bad English. The sets differ in date, purpose and selection, so the gap cannot be attributed to AI alone.
Nor is this a list of the ten most frequent phrases in AI writing. It is the result for ten phrases we chose in advance. The exact wording matters: zero for “I wanted to reach out” does not mean no email used “reach out”. A different tense, a contraction or a shorter version is a different match.
What was counted, and what was left out
The generation run took place on 12 October 2026 through Google’s API, with no system instruction, default temperature and minimal thinking. There was one response per prompt, not repeated trials. Where a response offered alternatives, only the first complete email was counted. Six responses stopped in later alternatives or notes; their first emails were complete and retained.
Searches ignored letter case and normalised whitespace. The “don’t hesitate” row accepted either “contact” or “reach out” after it and either straight or curly apostrophes. Other rows used the displayed wording. Real messages were deduplicated and cut at recognised quoted-reply or forwarding markers before this comparison; sign-offs were retained. No real email was sent to the model. The results are not an authorship test: a person can write every phrase in this table.
Make a decision about the sentence, not the phrase
Cut repetition. In the shared-drive email, the repair has already been announced. The words after “patience” add little. Likewise, “I am writing to ask whether you can send the file” can become “Can you send the file?” The revised question still needs a filename and, if there is one, a deadline. These are editing examples, not further model results.
Replace a vague request with the actual one. “At your earliest convenience” leaves the timing to the reader. If you need a reply by Thursday afternoon, say that. If you have no deadline, do not invent one merely to make the email sound direct. “Please let me know if you have any questions” can stay when questions really are welcome, but it cannot do the job of “Please confirm the delivery address.”
Keep courtesy that you mean. I would keep “Thank you for your patience” in a message to people whose work had been interrupted. I would be more reluctant to keep “Thank you for your understanding” after a disputed decision: the reader may not understand or agree. The difference is what I am thanking them for, not how often a model writes the line.
“Best regards” needs no unusual replacement. A sign-off is allowed to be conventional. The same restraint applies to deleting “just”: decide what it does in that particular sentence.
For the shared-drive message, a finished edit could read as follows. This is our wording, assuming the sender is reporting the repair rather than claiming to have made it:
Hi team,
The shared drive is available again. Please tell IT if you still have trouble accessing it.
Thank you for your patience.
Best regards,
[Your name]
The useful change is that the status and the next action are easy to find. Changing the sign-off to prove that a human wrote it would add no useful information.
