Autonomous outbound is the AI sales bet that ramped fastest and broke first. The pitch was plausible and exciting. Point an AI SDR agent at a list, let it research, write, and send at a volume no person could match, and watch pipeline show up. On a Powerpoint slide the math is unbeatable. Infinite volume, almost no cost per touch, no ramp time, no comp plan, no PTO. No brainer!
BUT! It scaled exactly like it promised, and scaling it is what broke it. When you let an AI write and send thousands of messages with nobody reading them, the quality drifts, the structure becomes recognizable, and the same calculus that made it cheap to produce made it easy to ignore. Buyers learned to spot an AI-written email faster than vendors learned to vary it. Deliverability took the hit next, because mailbox providers are very good at noticing when one domain starts behaving like a machine, and programs that ramp from zero to several thousand sends a day inside a 30-day pilot trip that detection almost immediately. So the cost per touch stayed near zero and the value per touch went to zero with it, obviously not a trade anyone signed up for.
In scenarios where companies ran ten or a dozen of these tools side by side, the results speak for themselves. The autonomous setups underperformed. The setups that worked kept a person in the loop. Sucks if you bought the infinite-scale story, but it matches what I’ve seen up close. I’ve spent the last year and a half building and fixing the kind of GTM systems that were supposed to make the human SDR obsolete, and instead of agents taking over, humans continue to outperform the agents brought in to replace them.
For a while the failures were kept quiet, buried in client CRMs where you only saw the residue of an abandoned installation. A pipeline full of replies that look like meetings but aren’t, a sender reputation quietly underwater, a list that got blasted so hard the good accounts now auto-route you to spam. Then the proverbial AI SDR sh!t hit the fan when 11x became the poster child for what can, and does, go wrong. They were backed by Benchmark and a16z, and TechCrunch reported they were listing companies as customers that weren’t customers. ZoomInfo ran a one-month trial and said the product “performed significantly worse than our SDR employees,” then spent months demanding 11x take their logo off the site. A former employee told TechCrunch the company was claiming around $14 million in ARR when only about $3 million of contracts had cleared the 90-day window.
It wasn’t one company either. Artisan ran the “stop hiring humans” billboards, and by early 2026 LinkedIn had rate-limited Ava’s automated outreach and banned accounts tied to suspected automation, which quietly removed an entire channel from a multichannel product. If you open LinkedIn on any given day right now the operator stories are everywhere, the burned domains, the cancelled pilots, the pipeline that evaporated the second someone read the replies. The pattern under all of them is the same. The AI did exactly what it was told. It wasn’t told anything worth doing, because the premise was that volume would cover for judgment. And scaling a bad process only breaks it faster.
The fix is rarely “turn off the AI.” It’s putting the decision back where it belongs. When I rebuild one of these, I keep the AI focused on the parts it’s good at, which is research, signal monitoring, first-draft writing, boring enrichment work that a person hates and does at 4pm. What changes is that a human owns the part the machine gets wrong… the judgment. Is this account worth a touch right now, is this a real trigger or a coincidence, does this draft sound like me or does it sound like a template with the company name pasted in? That’s the approval step, and your differentiator.
You read it here: humans still have a role in sales outreach. It’s going to be a while until the AI gets good enough to cut them out. However, the research, the drafting, the monitoring, those are the parts that a person doesn’t have to spend the time sourcing anymore. Judgment was never going to automate cleanly, because it depends on context the model doesn’t have and consequences the model doesn’t feel. A bad send costs AI nothing. It costs you domain reputation and tracks against quota.
So the right answer is neither “AI does outbound” or “people do outbound,” it’s AI does the volume work and a person owns the decisions that carry risk. That’s less exciting than a fleet of autonomous agents manufacturing pipeline while you sleep, and it’s also the version that’s still standing after a year of everyone testing the other one in public. This aligns with everything I believe about building in general. The goal is to remove friction and not to remove the human, and any system that makes someone’s job worse, or makes your sender reputation worse, is a bad system.
The interesting question now isn’t whether to keep a person in the loop, it’s which decisions need one. Approve every send and you’ve rebuilt a slow manual process with extra steps. Approve nothing and you’re back where this started, watching another logo come off another website. The teams that figure out where exactly that line sits, which accounts and which triggers are worth a human look and which can run on rails, are going to win.
Back to Blog