Every cold email guide of the last five years says the same thing: open with something about the prospect. Mention their podcast, their funding round, the LinkedIn post from last Tuesday. Once tools like lemlist made that cheap, and then GPT made it free, almost everyone started doing it.
We did it too. From May 2026 our sending tool wrote an AI first line for each contact on many campaigns, based on their company and profile. Other campaigns ran without it: either one plain template, or copy assembled at random from a set of value propositions, tones and calls to action, with nothing about the person. So we have both kinds of email going out through the same tool, to the same kind of B2B lists, often for the same client.
The personal line did not help. Across everything we sent, the contacts who got one replied at 0.53% and showed interest at 0.15%. The plain template got 1.00% and 0.56%. Per 1,000 people, that is 1.5 interested replies with the personal line against 5.6 without it.
Same month, same result
Reply rates have been falling all year (see reply rates by year), and the AI first line came in late, so we also compared the two inside the same month. The template matched or beat the personal line every month both ran:
| First emailed | AI first line: replied | Template: replied | AI first line: interested | Template: interested |
|---|---|---|---|---|
| May 2026 | 0.82% | 1.23% | 0.21% | 0.64% |
| Jul 2026 | 0.56% | 0.56% | 0.15% | 0.23% |
| Aug 2026 | 0.53% | 0.56% | 0.15% | 0.23% |
| Sep 2026 | 0.38% | 0.83% | 0.13% | 0.44% |
Same client, still no lift
Some clients are easier to sell for than others, so we also looked at the 19 clients that sent both an AI first line and a plain template. Weighted the same way, the personal line got 0.51% replies and 0.18% interested. The template got 0.60% and 0.23%. For 10 of those clients the personal line did a bit better, for 9 it did worse. Inside a client it is roughly a coin flip, which is the point: the AI line costs a model call per contact and a research step per list, and it buys nothing.
Why we think it stopped working
Our guess is the same thing that happened to banner ads. The first banner ad in 1994 was clicked by 44% of the people who saw it. By 1998 the rate was around 0.5% and researchers had a name for people skipping them: banner blindness. "Saw your recent post about scaling your team" worked when a human had clearly spent five minutes on you. Now every inbox gets several of those a day, all written by the same kind of model, and people read them as a sign that software sent the email. We did not test that directly. What we can say is that the line no longer moves the reply rate.
What still moves it, from our other studies: who you email (seniority and company size), whether the email lands in the inbox, and the first message itself, which brings 69% of all replies.
How we counted
The data is every email sent through the Sales.co sending tool from April 8, 2025 to September 25, 2026: 2,073,981 sent messages and 59,537 replies. A contact is one email address within one client. We dated and classed each contact by their first email, not the follow-ups, which left 1,279,153 contacts. A contact counts as replied if that address ever replied to that client; auto-replies, out-of-office messages, bounces and spam are left out. "Interested" means the reply was labelled interested, possibly interested, meeting booked or an ongoing conversation.
"AI first line" is a first email with a generated opening line about the recipient (360,004 contacts). "Plain template" is a first email with no personal line, including cases where the tool tried to write one and dropped it as too weak (28,896). "Randomised copy" is copy assembled from components with no personal line (747,340). 142,913 early contacts have no method recorded and are left out of the comparison. September 2026 is partial and recent contacts have had less time to reply, which lowers both groups in that month equally.
Data: totals by type (CSV), by month (CSV), all contacts by month (CSV).