Jason Lemkin's SaaStr team got a 6.7% response rate out of AI SDRs. That is a real number and a good one.
It came with a receipt attached. They got there only after fixing their RevOps, defining messaging they already knew converted, and committing 15 to 20 hours a week to running five agents, on the thing that was sold as not needing any. SaaStr's write-up is blunt about which number matters: "the bigger investment is your time."
The phrase for this comes from Death to Cold Emails, whose post-mortem on the category concludes that what happened was "not a technology failure. It is a naming failure." That is the right diagnosis and it is worth taking further than the post-mortem does, because you can trace what the word does.
"Autonomous AI SDR" tells a buyer they are purchasing a replacement for a headcount. So they price it against a salary — $50,000 to $100,000 a year looks like a bargain against a fully loaded rep — and they staff it like a replacement, which is to say not at all. Then the oversight arrives anyway, unassigned and nowhere in the business case, and the tool underperforms for a reason nobody costed. SaaStr's own readiness checklist puts the question before you buy, not after: can you commit 10 to 20 hours weekly for 90 days? If no, wait.
The contracts already knew
Here is the part that should bother you more than the churn.
TechCrunch reported in March 2025 that 11x — backed by a16z and Benchmark — sold one-year enterprise contracts carrying a break clause at three months. Former employees and a prospective customer described that clause as functioning like a trial period. TechCrunch reported in its own voice that even after prospects used the break clause to end their trial and their payments, the company kept counting the ARR as if they were completing the full year.
11x contests the characterisation. It says it reports contracted ARR, that its investors knew it used that metric, and that as of March 2025 it put retention at 79%. Benchmark says it received transparent updates including the break clauses. One employee quoted by TechCrunch described the early picture differently: "We were losing 70-80% of customers that came through the door."
Set the dispute aside and look at the structure both sides agree on. A three-month escape hatch was written into a twelve-month contract. You do not put that clause in a deal for a product you expect to hold. You put it in because you need the customer to sign now and you know the first quarter is where they find out.
The word promised no oversight. The contract quietly conceded that the buyer would want out by month three. Both were in the same document.
Artisan sold the word better than anyone, and got a different bill
Artisan's "Ava" is, by one widely-cited count, the most copied persona template in outbound — a digital employee with a name and a face, sold under billboards reading "Stop Hiring Humans." It is genuinely excellent positioning. It also commits you to the strongest possible version of the claim: not a tool, a hire.
Artisan raised repeatedly on that positioning — $11.5M in seed money by late 2024, then a $25M Series A led by Glade Brook Capital in April 2025.
Then, in December 2025, the company vanished from LinkedIn. Not for spamming, which is what the viral posts assumed. According to CEO Jaspar Carmichael-Jack, LinkedIn objected to Artisan using LinkedIn's name on its website, and alleged that Artisan was working with data brokers who had scraped the platform without permission. LinkedIn did not publicly confirm its reasoning. Artisan removed the references, tightened its vendor checks, worked with LinkedIn's enforcement team for about two weeks, and was reinstated.
So this is not a company undone. Carmichael-Jack downplayed the damage, noted that little of Artisan's data comes from LinkedIn, and observed that the ban drove more inbound than usual. His line on it: "Every startup inevitably has some kind of thing that comes back to bite them [from things] that they do early on."
But notice where the bill landed. The positioning promised an employee. What LinkedIn actually objected to — its own name in Artisan's marketing, and where the data had come from — sat in the layers underneath the persona, not in the agent itself. You can rent those layers. You cannot rent knowing what is in them.
What this costs you, specifically
Dru Riley puts autonomous sales tools at 50 to 70% annual churn — roughly twice the human role they were sold as replacing. Gartner expects AI agents to outnumber sellers 10 to 1 by 2028, while fewer than 40% of sellers will say agents improved their productivity. Gartner is forecasting both curves diverging, not reporting it.
Meanwhile the underlying economics are real and improving. Dru Riley reports Explee billing roughly 3 cents an email with no subscription, which puts a thousand prospects at about $30. That job used to require a hire. The sourcing half genuinely got automated.
So the category is not dead. It was misnamed, and the misnaming picked the wrong buyer, set the wrong budget line, and skipped the staffing conversation that would have made it work.
The rule this leaves you with
If you are buying: make the vendor name the single job the thing does, then ask whether one week proves it. Whole-loop replacement is the promise that keeps failing. Buy the sourcing — software is better than you at finding people and getting mail delivered. Keep the writing and the calls, because that half teaches you which objection kills your deal, and outsourcing it buys meetings and skips the education. And ask where the data came from, in writing, because that question is now a platform-risk question.
If you are selling: the word you choose sets the budget line your buyer prices you against, and the budget line sets whether they staff you. "Autonomous" got 11x priced against a salary and compared to one. A narrower promise would have been worth less per contract and survived past month three.
Either way, export your lists. Seven sales-tech vendors changed hands in five months — Zoom agreed to buy Common Room on 2 July 2026, HubSpot announced Warmly days earlier. Your reply history compounds on someone else's server, and you will migrate on their schedule, not yours.
The category told buyers the agent was the variable. It was never the agent.
