We don't sell AI SDRs, and the last year shows why
The pitch was software that finds leads, writes emails and books meetings while you sleep. It hasn't gone well. What went wrong, and how we run outbound instead.
For about two years the loudest idea in outbound was the AI SDR. You describe your ideal customer, the software finds leads, writes the emails, handles replies and drops meetings into your calendar. No sales team needed. A few of these companies raised a lot of money on that promise.
It hasn't gone well. Back in March 2025 TechCrunch reported that 11x, one of the best funded of them, was showing logos on its website from companies that weren't paying customers, and that plenty of clients left after short trials. That story set the tone for a lot of what followed.
The writing was rarely the weak point. Current models write a decent email. What broke was handing the whole process over and letting nobody look at it.
Why hands-off outbound burns out
Cold email has very thin margins for error. Since early 2024, Google and Yahoo have required bulk senders to set up SPF, DKIM and DMARC, offer one-click unsubscribe and keep spam complaints under 0.3%. Microsoft started enforcing similar rules for Outlook and Hotmail in May 2025. The rules themselves didn't change this year. Enforcement did, and mail that fails the checks is now more likely to be refused outright than quietly sent to spam.
0.3% works out to about one complaint per 330 emails. An agent sending a thousand emails a day to a loosely matched list, with nobody reading them, can get there in an afternoon. And it's your domain that takes the hit.
LinkedIn has been tightening up too. This spring it took down HeyReach's company page and the tool dropped its LinkedIn features soon after. Detection seems to be about behaviour now (identical send times, low acceptance rates, data centre IPs) more than which tool you happen to use.
What the numbers say works
Instantly's 2026 cold email benchmark is the most useful data set I've read this year. The average reply rate across their platform is 3.43%. The top 10% of senders get more than 10.7%. My read is that most of that gap comes down to who you email and why, which happens to be the part fully automated tools are worst at.
Two other findings stuck with me. The first email in a sequence gets 58% of all replies, so your opener matters far more than follow-up number five. And emails under 80 words perform best. Short, specific, one question at the end.
How we actually run it
We use AI a lot, just not as a substitute for judgement. Agents do the research: what the company sells, who makes the decision, whether something changed recently that gives us a real reason to write. They draft the emails. Then a person on our side reads what's going out before a campaign starts, and a person handles every reply from someone who's interested. The dull infrastructure (separate sending domains, warm-up, authentication, bounce checks) is done before a single email leaves.
It's slower than "10,000 emails on day one." It's also how you still have a working domain six months later.
If someone is pitching you an SDR that needs no human at all, ask them who reads the emails before they go out. If the answer is nobody, you have your answer too.