I've spent months reading about AI tools with the same scepticism most of us have - most are vapourware dressed up as a landing page. But three months ago, circumstances forced my hand. Our one SDR, who was booking maybe one call a week on a good week (most weeks zero), quit. My co-founder wanted to just 'use AI'. I wanted to punch him. We were broke, though, so I tried.
First attempt was disastrous. Gave Claude our ICP and asked for 50 cold emails. Got fifty variations of 'I hope this email finds you well...' Straight into spam. Reply rate was 0.3%. Almost told my co-founder where to stick it.
Then I stumbled on a framework about splitting the SDR job into separate 'agents' - signal detection, lead scoring, message writing, reply handling. Each gets its own context. The signal detection part changed everything. Instead of blasting lists, I set up monitoring for actual buying signals: someone posts about struggling with our problem, a company raises a round, a competitor's customer leaves a bad review. Built janky Python scrapers, piped everything into Claude with a scoring prompt.
That prompt took two weeks to get right. First versions scored everyone 'high intent' because I wrote it like a wishful thinker. Had to add negative examples - someone posting thought leadership is usually a vendor, not a buyer. Once I called that out explicitly, it worked.
For message writing, I stopped trying to generate 'good cold emails'. Instead, I fed Claude the actual signal and told it to write a reply to that signal. Not an email. A genuine engagement. If someone posts about team struggling with data quality, the message shares a specific take and casually mentions we solved it for a similar company. Often no CTA - just a perspective ending in a question. Reply rate went from 0.3% to about 4.5%. Doesn't sound huge, but when you're sending 30-40 targeted messages a week instead of 500 spray-and-pray, it adds up. Booking 6-7 calls a week now, compared to the previous 1. Closed two small deals last month directly from this.
I don't fully automate sending - maybe 15% of messages need editing, another 5% I kill. Spend about an hour a day reviewing signals, tweaking messages, handling replies. The reply-handling agent was hardest: Claude wants to over-explain and sell. Feeding it 30 of my actual replies and saying 'match this tone exactly' finally worked.
Total cost: about $300/month between Claude API and scrapers, versus the $4,500 we were paying an SDR who booked one call a week. Output is tied to actual buying signals, not just ICP matches.
Things that still suck: LinkedIn DMs are hard to automate without account restrictions - one account already flagged. Claude sometimes hallucinates company details - got a reply asking 'what are you talking about?' Enterprise deals need real account research that AI can't do yet. About once a week Claude refuses to write a sales email because it decides it's 'manipulative'. And some weeks signals dry up and I'm sitting on 8 leads instead of 35 - inconsistency I haven't solved.
I'm not sharing exact prompts because they're too specific, but the architecture - splitting the SDR function into specialised agents - is what matters. A good senior SDR would book 15-20 calls a week and handle enterprises. But for $300/month when you're a tiny startup trying to survive? Seven calls a week keeps the lights on. In this macro environment, that's not just a hack - it's survival.
Happy to answer questions. Tyler, if you're reading this, you know it's true.
Tools I use: LinkedIn Sales Nav (manual), Google Alerts, a G2 review scraper from GitHub, Trigify for job changes, and ConnectSafely AI for sending LinkedIn messages without getting your account nuked. The magic isn't the data collection - it's the scoring step where Claude decides what's worth acting on. And no, I won't build this as a SaaS. I like my duct‑tape Python scripts only I understand.
