Do AI SDRs work? Not with the data you have.

Every CEO has asked the question this year, usually right after a board meeting: should we be using AI SDRs? I have a bias here and you should know it before reading on. We spent years as executives at an outsourced SDR agency we bootstrapped to $20M ARR, built on human sellers. AI SDR vendors carry the opposite bias. Neither side’s word is worth much on its own, so what follows runs on evidence, and where the evidence is fake, it says so.

How well do AI SDRs work?

About as well as the data you feed them, which for most companies means poorly. We could not find an independent benchmark showing AI SDRs outperform human SDRs, the most quoted churn statistic is fabricated, and the best-documented company story is contested. What has been measured is readiness: a Harvard Business Review Analytic Services survey of 1,574 enterprise IT leaders, sponsored by Cloudera and published in March 2026, found just 7% say their data is completely ready for AI.

The data is the whole game. An AI SDR executes your records at scale: reachable or not, right person or wrong one, current title or two jobs stale. An AI SDR built on anything short of ready data does not underperform quietly. It underperforms in your prospects’ inboxes, in public, at volume.

What is an AI SDR?

An AI SDR is software that does the job of a sales development rep without the rep. It picks prospects from your CRM or a purchased list, researches them, writes and sends personalized outreach, follows up, handles replies and books meetings, with little or no human review between those steps. Vendors also sell them as AI sales agents or autonomous SDRs. The defining trait is autonomy: an SDR using AI tools still has a human deciding what goes out, and an AI SDR does not.

That autonomy is why the input data decides the outcome. Nothing sits between the record and the send to notice that the record is wrong.

The Dave problem

Picture the failure at unit scale. Your AI SDR emails Dave, a purchasing manager at Ford Automotive East Peoria, with a confident, personalized note, when the deal you wanted lives with procurement at Ford headquarters. Wrong entity, right company name, perfect grammar. Nothing in the system knows it happened. Now scale it by 25,000 sends and imagine the results, because that is what an autonomous SDR does with an entity-resolution problem: it commits to it, politely, everywhere.

Humans make the same mistake. The difference is that a human makes it once, hears something off in the reply, and stops. We watched this for years running SDR teams: the rep is the error-correction layer nobody prices in. They skip the record with the suspicious title, and they catch the subsidiary mix-up before it becomes a sequence. An AI SDR has no such layer. It reads the record as truth, because the record is all it has.

What does the evidence say?

Start with what it does not say. A statistic circulates on LinkedIn claiming 50 to 70% of AI SDR deployments churn within 90 days, usually attributed to UserGems. We traced it. The UserGems article it supposedly comes from does not contain the number, and no primary source for it exists anywhere we could find. Whatever the truth about AI SDR churn, that stat is manufactured, and this category deserves better evidence hygiene.

What is on the record is narrower and more interesting. In March 2025, TechCrunch reported on 11x, one of the category’s most funded startups. An employee told reporters the company was “losing 70-80% of customers that came through the door,” while 11x put its own retention at 79% in response. The same reporting found 11x had told investors it had $14 million in ARR while roughly $3 million in contracts had survived the standard three-month trial clause, and ZoomInfo, after a one-month pilot, said on record that the product “performed significantly worse than our SDR employees.” One company, contested numbers, anonymous sources. Do not read it as an industry churn rate. Read it as what happens when a category’s demos outrun its data.

The buyer side has been measured, and it points the same direction. A Gartner survey of 645 B2B buyers, fielded in late 2025, found buyers were 28 percentage points more likely to say a human sales rep, rather than GenAI, helped them advance to the next step of a purchase. Gartner also predicts that AI agents will outnumber sellers 10 to 1 by 2028 while fewer than 40% of sellers report the agents improved their productivity, a gap Gartner has since given a name: agent sprawl, more AI agents without more seller impact. The supply is arriving faster than the results.

What has to be true before an AI SDR works

The bar is knowable, and it is the same one we published as the definition of AI-ready data: every email validated past the accept-all layer, every phone number tested for whether it answers, titles and companies verified rather than inferred, and location data present and trusted, because an AI that emails an EU contact your system did not know was in the EU has converted a data gap into a compliance problem. If the AI SDR dials as well as emails, the phone adds a legal gate of its own: in the United States an AI voice call needs prior consent, which we cover in whether AI cold calling works and where it is legal.

If your CRM clears that bar, an AI SDR is worth piloting on a contained segment, with a human reviewing what it sends until its measured error rate earns more rope. If it does not clear the bar, the AI SDR will distribute your data quality to your market faster than any team you could hire, and your sellers’ remaining hours will go to apologizing for it.

Are AI SDRs worth it?

Not yet, for most companies, and the reason is the data rather than the software. AI SDRs work exactly as well as the substrate you stand them on, and the readiness data says that for most companies today the answer is no, not yet. Which puts the real question back on your desk: not whether to buy an AI SDR, but whether your data could support one.

Run that test before the pilot. The vendors will not ask you to.