What a Rewrite Model Actually Cuts From Non-Native English Email: 60 Measured Drafts
On 2026-08-09 we ran 60 email drafts through our own rewrite API and measured what actually changed — word count, sentence length, contraction use, and which non-native and AI-generated phrasings survived. 40 non-native business email drafts went through the native tone; 20 AI-written business emails went through the humanize tone. The raw inputs and outputs are public, linked below, so none of the numbers on this page have to be taken on trust.
Method
The corpus is 60 items: 40 tagged kind: esl, run through the native tone, and 20 tagged kind: ai, run through the humanize tone.
- The 40 ESL drafts spread over eight intents — request, follow-up, apology, cold intro, status update, decline, scheduling, thanks/close — and four L1 families, unevenly weighted: Romance 14, Slavic 10, South Asian 8, East Asian 8.
- The 20 AI drafts are AI-written business email in the shape a chatbot defaults to: triadic lists, uniform sentence length, and the vocabulary every model reaches for.
- One item is real: esl-01, a 126-word deadline-extension email the founder actually sent, already published in an earlier post. Every other draft in this corpus was constructed for this measurement — each one built to plant specific constructions catalogued in our own materials. We didn't sample real non-native email for this because Limato doesn't store user text (see the privacy policy), so there was nothing to sample from.
- Models were pinned for the run:
nativecalledgpt-4o-mini,humanizecalledgpt-5-miniwithreasoning_effort: low. Both measured on 2026-08-09. The prompt and the corpus both change over time — a rerun on a later prompt version is a different measurement, not a contradiction of this one. - Word count is whitespace-separated tokens. Compression is 1 − after/before.
The corpus, the raw output, and the measurement script are public, in this repository: corpus.json (the 60 inputs), raw.jsonl (one line per API call, input and output verbatim), and measure.js. Running node measure.js recomputes every table below from raw.jsonl, offline.
How much shorter the text got
| tone | n | median | p25 | p75 | shortest cut | deepest cut | mean |
|---|---|---|---|---|---|---|---|
| native (ESL drafts) | 40 | −37% | −30% | −40% | −23% | −54% | −36% |
| humanize (AI drafts) | 20 | −21% | −16% | −23% | +3% | −35% | −19% |
Across the 40 ESL drafts the corpus went from 2,736 words to 1,737.
The native prompt tells the model to target 40–55% and calls anything under 30% a failure. 10 of 40 drafts (25%) came in under that 30% floor, the shallowest at 23%. So on a quarter of the drafts, the model misses its own instruction.
The humanize prompt targets 15–35% and says an output longer than the input is a failure. It misses at the same rate: 5 of 20 (25%) came in under the 15% floor, the shallowest at 6%, and one of the 20 came back 3% longer than it went in.
Which non-native constructions actually survive a rewrite
Nine of the 18 constructions below are named verbatim in the rewrite prompt's kill list, and two more — "I hope this email finds you well" and "Dear Mr/Ms" — are covered by explicit greeting rules in the same prompt. So for eleven of them, a near-total removal rate measures compliance with a list the model was handed, not discovery. The interesting number is which ones survive anyway.
| construction | planted | survived | removal rate |
|---|---|---|---|
| "Kindly" | 17 | 0 | 100% |
| "I am writing to..." | 15 | 0 | 100% |
| "...in order to..." | 13 | 0 | 100% |
| "I hope this email finds you well" | 12 | 0 | 100% |
| "Dear Mr./Ms./Sir" | 10 | 0 | 100% |
| "Do not hesitate to..." | 10 | 0 | 100% |
| "I would appreciate if..." | 9 | 0 | 100% |
| "...at your (entire) disposal" | 8 | 0 | 100% |
| "Please be informed that..." | 7 | 0 | 100% |
| "Thanks/thanking you in advance" | 7 | 0 | 100% |
| "Revert back" / "do the needful" | 6 | 0 | 100% |
| "Please find attached/enclosed" | 5 | 0 | 100% |
| "Sorry for the inconvenience" | 5 | 2 | 60% |
| "As previously discussed" | 5 | 0 | 100% |
| "At your earliest convenience" | 4 | 0 | 100% |
| "As per..." | 3 | 0 | 100% |
| "Hereby" | 2 | 0 | 100% |
| "Awaiting your (kind) reply" | 2 | 0 | 100% |
"Sorry for the inconvenience" is the only one that survived — 2 of 5 drafts kept it: esl-13, a service-outage notice, and esl-26, declining a conference invitation. That's defensible, not a miss: in both drafts the apology is the actual content of the email, so cutting it removes meaning rather than padding.
Also worth reporting: the real founder email (esl-01) did not contain that phrase in its input. The rewrite added "Sorry for any inconvenience this causes" to the output — see the example below.
Which AI tells survive a humanize pass
These 24 phrases are also named in the humanize prompt, so the same caveat applies: near-total removal is mostly compliance with a ban list, not discovery.
| tell | planted | survived | removed |
|---|---|---|---|
| "fast-paced" | 3 | 2 | 1 of 3 |
| "streamline" | 6 | 2 | 4 of 6 |
| 22 others | 88 | 0 | 88 of 88 |
"Fast-paced" is the most persistent AI tell in this run — it survived 2 of the 3 times it was planted, as "in a fast-paced retail setting" and "in a fast-paced engineering organisation". The prompt bans the narrower string "in today's fast-paced world"; the model honours the literal string and keeps the collocation. "Streamline" survived 2 of 6.
The other 22 tracked tells were removed in every draft where they appeared: moreover (9 planted), comprehensive (9), furthermore (7), robust (7), crucial (7), "it is important to note" (6), seamless (6), leverage (5), "in conclusion" (4), delve (3), pivotal (3), elevate (3), unlock (3), foster (3), empower (3), transformative (3), underscore (2), landscape (2), harness (2), "navigate the complexities" (2), realm (1), cutting-edge (1).
Sentence length and contractions
| tone | words/sentence before → after | contractions/100 words before → after |
|---|---|---|
| native | 13.7 → 10.3 | 0.0 → 2.3 |
| humanize | 16.6 → 13.2 | 0.1 → 2.2 |
Across the 40 ESL drafts the input contained exactly zero contractions — by construction, since the drafts were built from patterns that avoid them. What the run does show is the model putting them back: 2.3 per 100 words on output, in every draft.
Three examples
Unedited API output. Typographic apostrophes are the API's, not ours — preserved verbatim.
The real email — 126 → 59 words, −53%
"Dear Mr. Peterson, I hope this email finds you well. I am writing to you in order to inform you that we have unfortunately encountered some difficulties regarding the delivery of the second milestone, which was initially planned for the 20th of July. The reason of this delay is that our backend developer has fallen ill and therefore we were not able to finish the integration tests in due time. We would like to kindly ask you if it would be possible to postpone the deadline until the 27th of July. We apologize for the inconveniences that this situation may cause to you and we remain at your entire disposal for any further information you may need. Looking forward to your kind reply. Best regards, Mikhail"
"Hi Mr. Peterson, we’ve run into some issues with the second milestone delivery, originally set for July 20th. Our backend developer is ill, so we couldn't finish the integration tests on time. Can we push the deadline to July 27th? Sorry for any inconvenience this causes. Let me know if you need more info. Looking forward to your reply."
The deepest cut — esl-31, 68 → 31 words, −54%
"Hello Erik, I hope this email finds you well. I would really appreciate if we could find a time that works for both of us in order to discuss the open points of the contract before the holidays. My calendar is quite flexible next week, except Wednesday morning. Could you please confirm at your earliest convenience which slot would be the most convenient for you? Best regards, Vera"
"Hi Erik, can we find a time to discuss the contract's open points before the holidays? I'm flexible next week except for Wednesday morning. Let me know what works for you."
The shallowest cut — esl-23, 70 → 54 words, −23%
This draft was already fairly direct, so there was less padding to remove. The tells still went — "please find attached", "as previously discussed", "please do not hesitate".
"Hi Peter, Please find attached the weekly progress report for calendar week 28. As previously discussed, the report now contains a separate section for the open risks and their mitigation status. The integration with the payment provider is still the main bottleneck and we are two days behind our internal plan. Please do not hesitate to contact me before the steering meeting if you need more details. Best regards, Celine"
"Hi Peter, attached is the weekly progress report for week 28. As we discussed, it now includes a section for open risks and their mitigation status. The integration with the payment provider remains the main bottleneck, and we're two days behind our plan. Reach out if you need more details before the steering meeting."
What this data does not show
- How often these constructions appear in real non-native email. The corpus was constructed, not sampled — the frequencies here are ours by design.
- Anything about non-native writers by first language, region, or proficiency. The L1 labels describe which patterns each draft was built from, not measured populations of writers.
- Anything that generalises past this prompt version, these two models, and this date.
- Whether shorter email gets more replies. We measured text, not outcomes.
- Removal rates are largely compliance with an explicit ban list, as stated above.
Using this data
These numbers are free to cite with a link back to this page. The corpus, the raw outputs, and the measurement script are all in our public repo, so anyone can rerun the measurement or contradict it.
Limato, "What a Rewrite Model Actually Cuts From Non-Native English Email" (2026) — median 37% compression across 40 non-native business email drafts, n=40.
Frequently asked questions
How much shorter does a native-style rewrite make a non-native email?
Measured across 40 constructed non-native business email drafts run through Limato's native tone on 2026-08-09: a median of 37% shorter, with the middle half of drafts falling between 30% and 40%, the shallowest cut at 23% and the deepest at 54%.
Does an AI text humanizer make writing shorter or longer?
Usually shorter, not always. Across 20 AI-written business email drafts run through Limato's humanize tone, the median cut was 21%, with the middle half between 16% and 23%. One of the 20 drafts came back 3% longer than it went in.
Which non-native English phrases survive a rewrite?
Almost all of them go. 17 of the 18 constructions tracked in this run were removed every time they appeared — "kindly", "I am writing to", "in order to", "I hope this email finds you well", and 14 others. The one exception was "sorry for the inconvenience", which survived 2 of 5 times it was planted, both times where the apology was the actual content of the email rather than filler.
Which AI writing tells are hardest for a humanizer to remove?
"Fast-paced" was the most persistent: it survived 2 of the 3 times it was planted in this run, as "in a fast-paced retail setting" and "in a fast-paced engineering organisation". The prompt bans the narrower string "in today's fast-paced world", so the model honours the literal string and keeps the collocation. "Streamline" survived 2 of 6 times. The other 22 tracked AI tells — moreover, comprehensive, furthermore, robust, crucial, seamless, leverage, and 15 others — were removed every single time they appeared.