Does AI-written outreach work? 31 million emails say something uncomfortable.
The honest version of this answer has three parts, and the vendor version only ever quotes the first one. Here are all three, from the largest public dataset we could find.
Published 6 min read Sources checked 24 September 2026
Filed under Cold email, Outbound sales, AI personalisation, B2B marketing
The short answer
Three findings from Hunter’s 31-million-email dataset. Personalisation works: two custom attributes lift reply rate from 3.6% to 5.6%, up 56%. Full automation costs you: manually edited emails beat fully automated by 18%, 5.2% against 4.4%. And 69% of the US decision makers Hunter surveyed say it bothers them when AI wrote the email. Personalise the data, keep a human on the words.
On this page · 7 sections
Two claims are made about AI and cold email, and they cannot both be right. Either it multiplies your output at no cost, or it fills inboxes with obvious slop that damages your domain.
The largest public dataset we could find splits the difference in a specific and useful way.
What exactly does “personalisation” mean here?
Custom attributes: specific, verifiable facts about the recipient’s business dropped into the message. Not a first name. Not “I loved your recent post”.
Two of them is the threshold in the data. That is a low bar and most senders do not clear it, which is why the lift is available.
Why does full automation underperform?
The gap is 5.2% against 4.4%: modest, but consistent, and it points the opposite way to the pitch.
The likely reason sits in the same report, from a separate survey rather than the 31 million emails: 69% of the US decision makers Hunter asked say it bothers them when AI wrote the email. Your recipients have now read thousands of these. The register is recognisable, and recognition is fatal in a channel where the entire ask is “this was written for me”.
The thing being automated is not the writing. It is the research. Automate the half that scales, and keep the half that has to sound like a person.
So what is the actual split?
| Stage | Automate? | Why |
|---|---|---|
| Building and filtering the list | Yes | Pure data work, checkable, huge volume |
| Finding two verifiable custom attributes | Yes | This is where the 56% lives, and it is research not writing |
| Deliverability and list hygiene | Yes | Mechanical, and the thing that actually kills campaigns |
| Drafting the message | Draft only | Fine as a starting point, not as a send |
| The final words that go out | No | Edited out-replies automated by 18%, and 69% of surveyed US buyers object |
| Replying to a human who replied | No | You asked for a conversation. Have it |
What about the numbers that say the opposite?
You will see claims that AI personalisation produces three times the reply rate, and claims that it performs worse than sending nothing at all. We went looking for the sources.
The three-times claims are mostly vendor marketing without a dataset attached. We also could not find any source for the “worse than nothing” claim, despite it circulating widely, and it is contradicted by the 31-million-email figures above.
We are labelling Hunter’s data clearly: it is a vendor’s dataset, drawn from their own users, and their users are not a random sample of senders. It is still the largest number with a stated method that we could find, and we would rather cite that and say so than cite a conference slide.
What to do on Monday
- Count your current reply rate. If you do not know it, nothing below is measurable.
- Add two verifiable custom attributes to your next hundred sends. Not adjectives: facts.
- Have a person edit every message before it goes, even lightly. The 18% is the cheapest uplift on the list.
- Fix deliverability before volume. Sending more from a poor domain scales the wrong thing.
- Compare after a hundred. Not after a thousand: you will have burnt the list learning.
Questions people actually ask
Does AI-written cold email work?
Partly. In Hunter’s 31-million-email dataset, personalisation with two custom attributes lifted reply rates from 3.6% to 5.6%, but manually edited emails outperformed fully automated ones by 18% (5.2% vs 4.4%), and 69% of the US decision makers Hunter surveyed said AI-written emails bother them.
How much does personalisation improve cold email reply rates?
Two custom attributes lifted reply rate from 3.6% to 5.6% (a 56% improvement) in Hunter’s State of Email Outreach 2026, based on 31 million emails sent in 2025.
Which parts of outbound should be automated?
List building, attribute research and deliverability hygiene. The final wording that goes out should be human-edited: in the data, human-edited email out-replies full automation by 18%, and most of the decision makers Hunter surveyed say AI-written email bothers them.
Sources
- 01 Hunter, State of Email Outreach 2026: 31 million emails sent by Hunter users in 2025. Two custom attributes 5.6% vs 3.6%; manually edited 5.2% vs fully automated 4.4%. Separately, from Hunter’s survey: 69% of US-based decision makers say it bothers them if AI wrote the email. hunter.io(opens in a new tab)
Weight: vendor dataset, large, method stated; the 69% is a separate US survey whose sample the page does not give Re-checked: every six months Last read: 24 September 2026
Checked 24 September 2026. Next scheduled check 23 March 2027. Numbers that move (leaderboards, live indices) are re-checked every 30 days; annual datasets and rules in force every six months; dated research once a year. If something here has gone stale before we got to it, tell us and we will correct it and say what changed.
Cite this note
Noxia, “Does AI-written outreach work? 31 million emails say something uncomfortable”, Field notes, 8 September 2026; sources checked 24 September 2026. https://www.noxia.co.uk/field-notes/does-ai-written-outreach-work
We build the list and the research. You keep the words.
A filtered, deduplicated list with the custom attributes already found, delivered as a file you own outright. Priced per list, never per seat.