In 2024, Andreessen Horowitz and Benchmark poured $76 million into a company called 11x.ai.
The pitch was simple and seductive: an AI agent named Alice that replaces your entire SDR team. Cold emails, follow-ups, meeting booking — autonomous, end-to-end, no humans required. The billboards went up around San Francisco. The valuation hit $350 million. Every GTM founder on LinkedIn had an opinion.
Then the numbers leaked.
According to TechCrunch’s investigation, 70 to 80% of 11x’s customers were churning within the first three months. Internal employees described the product as “not working as well as suggested.” The reported $14M ARR turned out to be closer to $3M in contracts that actually survived the trial period. LinkedIn restricted Artisan — 11x’s closest competitor — from automating outreach on their platform in early 2026.
I want to be fair: 11x is still operating. So is Artisan. So is AiSDR. They have real customers, they ship real product.
But the broader story the AI SDR category tells is brutal. According to independent industry reports, 50 to 70% of AI SDR tools churn within 12 months. Only 2% of implementations survive past the first year.
Here’s what almost nobody is saying out loud:
Autonomous AI SDRs don’t work. AI SDR systems do. And the difference is one weekend of setup.
The story nobody tells
I hired my first SDR in 2022. She was great. She got poached by a Series B in her eleventh month.
I hired my second in early 2023. He quit after seven months — told me outbound was “crushing his soul.”
I hired my third last year. She was also great, and also gone within a year.
Each hire cost me roughly $58,000 in base pay, plus commissions, plus LinkedIn Navigator seats I kept paying for during the two-month vacancy between reps, plus the six weeks of ramp time before they were productive. All in, I was spending close to $90,000 per year to get about six months of real output.
In February, I fired my last outbound contractor.
Not because he was bad. Because I had spent the previous weekend building an agent that was already doing 80% of his job — and the 80% that mattered most.
The agent costs me $180 per month. Not $180 per seat. $180 total.
In the eight weeks since I flipped it on, it has booked 34 qualified meetings. My last human contractor booked 19 in the same window, and I was paying him roughly $8,500 per month to do it.
I’m not writing this to sell you on firing anyone. I’ll say what I think plainly at the end. I’m writing this because the AI SDR story most people are hearing — the one sold by the $74M-funded platforms — is wrong. And the real story, the one that actually works, has never been fully written down in one place.
Until now.
Why the “autonomous AI SDR” pitch broke
The mistake 11x and its peers made was philosophical, not technical.
They sold the fantasy that you could plug in an AI, remove the human, and watch meetings pile up. That’s not how outbound works. Outbound requires judgment at four specific moments — and every single one of them breaks when you fully remove the human from the loop.
Those four moments are:
Defining who qualifies as a “real” prospect (ICP)
Deciding what’s worth saying to them (positioning)
Reading the nuance of a reply (intent classification)
Knowing when to escalate to a real conversation (handoff)
Get these four right, and an AI agent can handle everything else — research, drafting, sending, warming, following up, classifying, routing. That “everything else” represents roughly 80% of an SDR’s working hours. A recent Gartner survey confirms it: top-performing sales teams in 2026 now run 80% of their outbound volume through AI, focusing human time only on high-intent replies.
Get those four wrong — or try to automate them — and you’re Artisan circa January 2026, with LinkedIn banning your IP and customers churning because prospects recognize your emails as machine-generated from the first line.
The system I built respects this line. It gives the agent everything mechanical, and keeps the human exactly where judgment is required. That’s why it works.
What I actually built
Before I tell you what’s behind the paywall, here’s what it is at 30,000 feet. The system has four agents working in sequence, plus one human checkpoint:
Agent 1 — Research. It ingests a daily list of 200 companies matching my ICP, pulls funding rounds, job changes, hiring signals, and tech stack changes, then scores each account 0-100 against my criteria. Only accounts scoring 70+ advance.
Agent 2 — Draft. For each qualified account, it writes three personalization angles based on the research, drafts three email variants, and stores them in a queue for review.
Human gate. I spend 15 minutes a day reviewing the queue. I kill maybe 5-10% of drafts. The rest ship.
Agent 3 — Send & sequence. Warmed domains, rotated inboxes, spaced sends, automatic follow-ups at days 3, 7, and 14.
Agent 4 — Reply handler. Every reply gets classified into one of five buckets: positive, objection, question, unsubscribe, out-of-office. Positive replies and qualified questions get routed to my calendar. Objections get a human-drafted response template. Everything else gets handled silently.
That’s the architecture. What makes it actually work is the detail — the specific prompts, the exact tool stack, the warmup protocol, the ICP scoring rubric, and critically, the three kill switches I set up after losing $3,400 in domain reputation in my second week.
What’s behind the paywall
The full premium edition gives you everything you need to build this yourself in a weekend:
The complete architecture with the four-agent flow, the single human checkpoint, and the data that flows between each stage
The tool stack with real prices — what I actually pay per month, including the two tools most people overspend on and the one free tool that replaced a $300/month platform
The Research Agent system prompt — copy-paste, with my ICP scoring rubric and the five signals I track (and the three signals that sound important but I ignore)
The Draft Agent system prompt — with the anti-generic ruleset that keeps replies from recognizing it’s AI (the “no AI tells” protocol)
The Reply Handler — how it classifies replies with 94% accuracy, and the escalation rules when confidence drops below 90%
The warmup protocol — the exact 21-day domain warmup schedule, because skipping this is how I burned $3,400 in my second week
The three kill switches that prevent catastrophic failure: volume cap, reputation monitor, and the human approval gate
The real numbers — week-by-week metrics from my first 8 weeks: emails sent, open rate, reply rate, meetings booked, pipeline generated
What it can’t do — a full honesty section on the five things the agent still fails at, and why you should never try to automate them
The $3,400 mistake I made in week 2, so you don’t make the same one
Then at the end, the thesis: why I believe the SDR role as we know it is dead — and what it’s being replaced by, which isn’t what most people think.
Why subscribe now
Premium subscribers currently pay $100 / year for permanent access — meaning the price locks in for you forever, even as we raise rates for new subscribers (which we do regularly). For the cost of roughly one month of Clay + Apollo + Instantly, you get the full playbook plus every other edition in the AI Playbook series — including the next one, dropping in two weeks:
“I Replaced My Executive Assistant With an AI Agent.”
The SDR edition is the most tactical thing I’ve ever written. Every prompt, every config, every number — all in one place. If you run outbound, or you’re about to hire someone to, this is a one-weekend project that will save you $70,000+ this year.