The AI-written first line got 0.53% human replies and the plain template got 1.00%
Sales.co runs cold email campaigns for B2B clients. Since April 2025 we have sent those campaigns through a sending tool we built ourselves. Its main feature was the personalized first line: for each contact, a language model reads the person's public profile and company details and writes an opening sentence about them. We built it because we believed it worked, and we expected the numbers to show a clear lift.
They did not. We took every contact first-emailed between April 8, 2025 and September 25, 2026, 1,279,153 in all, and grouped them by how their first email was written. A reply counts as human when it is not an auto-reply, out-of-office notice, bounce or spam. It counts as positive when it was labeled interested, possibly interested, meeting booked or an ongoing conversation.
| First email | Contacts | Human reply rate | Positive reply rate |
|---|---|---|---|
| AI-personalized first line | 360,004 | 0.53% | 0.15% |
| Plain template | 28,896 | 1.00% | 0.56% |
| Randomized components | 747,340 | 0.83% | 0.22% |
| No method recorded | 142,913 | 1.54% | 0.57% |
The gap is widest on positive replies. A contact who got the template was almost four times as likely to reply with interest as one who got the personalized line: 0.56% against 0.15%.
The last row is mostly early 2025 sends from before the tool recorded how an email was written. Reply rates were higher across the board then (see Cold email is dying and there is nothing you can do about it), so we leave that row out of every comparison below.
The template matched or beat the personalized line in each of the 4 months where both ran
The overall table is unfair to personalization. The AI first line only ran on 500 or more contacts a month from May 2026 onward, and by then reply rates for every kind of email were lower than a year earlier. Sales.co's longer series, Cold Email Reply Rates by Year, 2010 to 2026, shows the same long decline. So the fair test is the same month, counting only groups with 500 or more contacts in that month.
| Month | Personalized: human / positive | Template: human / positive | Components: human / positive |
|---|---|---|---|
| May 2026 | 0.82% / 0.21% | 1.23% / 0.64% | 0.74% / 0.27% |
| June 2026 | 0.59% / 0.17% | under 500 contacts | 1.43% / 0.67% |
| July 2026 | 0.56% / 0.15% | 0.56% / 0.23% | 0.83% / 0.30% |
| August 2026 | 0.53% / 0.15% | 0.56% / 0.23% | under 500 contacts |
| September 2026 | 0.38% / 0.13% | 0.83% / 0.44% | under 500 contacts |
In May the template got 1.23% human replies and the personalized line 0.82%. In July they tied at 0.56%, but the template still got more positive replies, 0.23% against 0.15%. In August it was 0.56% against 0.53%. In September, which is not finished and has had the least time to collect replies, it was 0.83% against 0.38%. On positive replies the template won all four months.
Randomized components, where the tool mixes a value proposition, tone, length and call to action from a set of options but writes nothing about the person, beat the personalized line on positive replies in May, June and July. In June the gap was 0.67% against 0.17%.
The template group is small: 28,896 contacts against 360,004 for the personalized line. That is why the next check matters.
Inside the same client, the personalized line won 10 times out of 19
One obvious objection is client mix. If the clients who used templates happened to sell into easier markets, the template would look better for reasons that have nothing to do with the copy. So we compared the two inside each client.
19 clients ran both an AI-personalized first line and a plain template on 300 or more contacts each. The personalized line had the higher human reply rate in 10 of them and the template in 9. That is about what you would get by flipping a coin. We ran the same check for the personalized line against randomized components: 24 clients ran both, and the personalized line won in 9.
The two clients with the largest template samples show the biggest gaps:
| Client | First email | Contacts | Human reply rate | Positive reply rate |
|---|---|---|---|---|
| Client A | AI-personalized | 8,297 | 0.98% | 0.47% |
| Client A | Template | 1,097 | 2.28% | 1.73% |
| Client B | AI-personalized | 9,051 | 1.01% | 0.64% |
| Client B | Template | 5,042 | 2.98% | 2.24% |
For both clients the template's human reply rate was two to three times higher, and its positive reply rate more than three times higher.
Every sender's first line now comes from the same models and the same public profile data
We cannot prove the reason from this data, but here is our reading. A personalized first line used to be a signal that someone had looked at your profile before writing. Now any sender can get one written in seconds, and most of them are written by the same few models from the same public profile and company page. A buyer who gets several cold emails a week sees the same kind of opener about their latest post or their company's growth in each of them. After a few of those, the recipient reads the line as the template it is, and it may count against the email.
We sell cold email as a service, and this is our own tool and our own clients' campaigns. We would have preferred the personalized line to win, since we built it. It did not win in this data.
Methodology
Source. The Sales.co sending tool database: every email sent and every reply received between April 8, 2025 and September 25, 2026.
Unit. A contact is one recipient address within one client workspace. Each contact is dated by its first email (not a follow-up) and classed by how that first email was written. 1,279,153 contacts were first-emailed in the window.
Replies. A contact counts as replied if the same address replied to the same client workspace at any time. Auto-replies, out-of-office notices, error notifications and spam are excluded from human replies; in an earlier Sales.co study, Cold Email Response Rates in 2026: The Real Numbers From 1,288,605 Sent Emails, 53.7% of all replies were automated. Positive means the reply was labeled possibly interested, interested, meeting booked or ongoing conversation.
Classes. AI-personalized: the first email opened with a line written by a language model for that contact, or an opener assembled or rotated per contact. Plain template: fixed copy with no line about the contact, including emails where the AI line was dropped as weak. Randomized components: a value proposition, tone, length and call to action drawn at random from a component set, with no line about the contact. No method recorded: mostly early 2025 sends.
Checks. A same-month comparison counting only classes with 500 or more contacts in the month, and a per-client comparison counting only clients with 300 or more contacts in each class compared. Clients are anonymized in the published files.
Limitations
- The client mix changes from month to month. The per-client check addresses this for the personalization comparison, but not completely.
- Recent months have had less time to collect replies. September 2026 is partial, up to September 25.
- A contact is classed by its first email. Later follow-ups to the same contact may have been written another way.
- Replies are matched by address within a client, so a reply to a follow-up is credited to the contact, not to a specific email. Most replies come from the first message anyway: 69% in The Cold Email Reality Check.
- The personalized class is concentrated in May to September 2026, when reply rates were low for every class. The same-month table is the fair comparison.
- This is our own sending tool and our own clients only. Other senders, markets and tools may see different numbers.
Data
- cold-email-personalisation-2026.csv: contacts and reply rates by class
- cold-email-personalisation-by-month-2026.csv: by class and month
- cold-email-personalisation-by-client-2026.csv: by class and anonymized client
- cold-email-reply-rate-by-month-2025-2026.csv: all contacts by month of first email
How to cite
Sales.co. "Personalization no longer matters for cold email." September 27, 2026. https://sales.co/research/personalization-no-longer-matters