Do AI SDRs actually work?

Yes, with a human in the loop. Mine have run every day since 2023. They do not work as autonomous set-and-forget replacements for people, and anybody selling you that version is selling the part of the job that was never the bottleneck.

I am Eduard Klein. I advise CEOs, founders and boards on AI Business Strategy, and I run several businesses on the agents described in the full AI SDR blueprint, published on 19 August 2026. This page is about what works and what broke.

What "works" actually means in my own system

Google's AI Overview in September 2026 says AI SDRs do not work as autonomous replacements for human labour but do work as efficiency multipliers under tight supervision. I agree with that and I would put it more precisely.

My agents do not run fully autonomously, and that is a choice rather than a limitation. I want control, and I want to adjust the process as I learn things. Within their own bounded areas the sub-agents are completely autonomous. They research, deliver data, enrich records, work on documents and proposals, document what they did, adapt, send email, check quality criteria, verify infrastructure, prepare drafts, and produce the data I make decisions on.

The number I watch is time per lead. I measured it every month through 2026 and it dropped every month since I put the second version into service, and it is the only metric here that compounds. Meetings booked is what everyone reports to their board, and it is the one that fools you in month 1.

Where mine failed

Plenty of my own agent projects failed across the 3 years from 2023 to 2026, and the failures were not the ones I predicted.

The current generation of my system took 8 months of 2026 to build. That was not 8 months of writing code. I built a version, I ran it, I watched it fail in ways I had not anticipated, and I rebuilt. The rebuild was the expensive part every time, and no amount of planning up front shortened it, because the failures were not in the plan.

What broke was rarely the model, whether that was Claude or GPT behind the API. It was the shape of the system around it. An agent that kept its own state produced a pipeline where I could not answer "what do we know about this account". A component with more access than its job required turned an ordinary bug into a question about what it had touched. Both are architecture mistakes, and I made both in 2024 before I wrote the 3 rules that now sit above every component.

I have also watched the third-party version of this repeatedly since 2023. A non-technical person gets 3 months of a production agent working, then the thing decides to restructure itself and breaks itself in the process. That is why I now follow a structured build procedure instead of improvising.

What are the biggest reasons AI SDR pilots fail?

Someone without development experience builds it, or nobody in the company understands the whole picture.

On the first: you barely write code any more, and you still have to know the concepts. Security models, architecture, project management, build management, versioning, plus compliance under the EU AI Act. That is 20 to 30 distinct topics to hold at once, and for a beginner the jargon alone is overwhelming before the work even starts. High cognitive load, and it does not announce itself until the thing breaks in month 3.

On the second: if nobody understands how sales actually hangs together, and departments build silos where information does not flow, you do not get a failed agent. You get 20 new construction sites at the same time, in IT security, in maintenance, in compliance. Individual agent tasks may work fine while the company slides into chaos around them.

There is a third reason that sours pilots without anybody calling it a failure. Weak AI gets introduced, promises a lot, delivers little, and management still expects the 10 times. That produces frustration and then attrition. The fear of replacement in the room is real and it deserves a straight answer rather than a reassuring one.

What are the biggest failure modes of AI SDRs?

No human in the loop, trying to build everything at once, and letting a beginner build it.

The first is underestimated. I have watched agents write genuinely silly things, and I do not mainly mean hallucinations. They talk too much, overwhelm the reader with information, and go off in directions no experienced person would. An agent cannot assess the consequences of its own actions. Only people can, in September 2026, which is why I stay in every loop that touches a human being.

Second, this has to be set up iteratively. Everything at once fails. AI needs longer test cycles than conventional software and constant correction, which is most of why 8 months in 2026 was the honest number for my own build rather than 8 weeks.

Third, and this is the one nobody says out loud, a beginner can get something impressive running and have no way to tell that it is structurally wrong. The demo works. The 6-month version does not.

When does building your own AI SDR stop making sense?

In an enterprise, when everyone starts building their own. That is the point where you need central infrastructure instead of 40 private agents.

Shared services, security monitoring, support. Small and mid-sized companies can be far more flexible, and rules like the EU AI Act still apply to them.

The honest threshold is not headcount. It is comprehension. The moment you can no longer manage the system because you no longer understand it, you need a professional on board. That is true at 10 people and at 10,000.

Where it does not work

I do not want to sell you the unbounded version, so here is the part the vendor pages leave out.

An AI SDR does not work if you plan to remove the human. It does not work if you expect the 10 times in month 1, because the first 3 months go into correction and context rather than output. And it does not work if nobody in the building can say how your sales process actually hangs together, because an agent will faithfully automate a process nobody understands.

It also does not replace judgement. I deliberately did not automate the parts where being wrong is expensive and hard to detect. Component 7 in my blueprint, the proposals, is not an agent at all. I made that call in 2026 and I would make it again.

If you want somebody who has already hit these failures to sit on your side of the build, that is what my AI coaching work is. The related question of whether your company is visible at all to the models your buyers now ask is a different piece: AI strategic visibility.

What I still cannot tell you

I do not have clean data on response rates.

Disclosure changes them, I expect a hit during the transition, and I would rather say that than quote a number I do not have. Component 4 in my own system introduces itself honestly rather than pretending to be me, which is a position I hold for reasons that are not measurement. I have run it for 6 months of 2026 without a comparison group.

The second thing I cannot answer is what happens when the person on the other end also has an agent. I saw the first of those threads in August 2026. I have no idea what the equilibrium looks like when 2 disclosed agents pre-qualify each other for 6 messages before either principal is involved, and I will write that up when I have watched enough of them to say something true.

What sits on either side of this: whether to build your own AI SDR or buy one and what an AI SDR actually is.

About the author

Eduard Klein is an AI Business Strategist for CEOs, founders & boards, and he knows what a working sales motion looks like from having run one: he carried a quota as an SDR up to 3.5 million in revenue, worked years in inside sales and key account management, and was a sales manager in the software industry before any of this was automated. He has trained thousands of sales people, which is where the pilot-failure patterns in this piece come from, and has published 170+ articles across 100+ publications including CIENCE and Callbox (the full list).