Most “AI prospecting” content online is either a screenshot of ChatGPT or a demo of a $2,000/month tool nobody I know actually pays for. This post is neither. It’s the workflow I actually run every week in Clay, on top of a real client’s ICP, with the specific enrichment sources, prompts, and cost tradeoffs.
The starting point
The client sells a technical developer product to mid-market fintechs. Traditional prospecting was: an SDR pulled a LinkedIn list, guessed at fit, and sent a generic sequence. Reply rate hovered around 1.2%. Everyone knew it was broken; nobody had time to fix it.
What I automated
Three layers, in order:
- Sourcing — a Clay table pulls companies matching the ICP from four different providers, deduplicates by domain, and drops anything already in the CRM. This alone killed 40% of the wasted volume.
- Enrichment — for each surviving company, an AI step reads the last 6 months of their tech blog, extracts what they’re building, and scores fit against 5 signals we care about. Not “is this a fintech” — that we already know — but “are they hiring platform engineers right now”, “did they mention scaling pains in the last quarterly”, etc.
- Personalization — one paragraph per lead, written by an AI step with strict rules: reference something specific from their public output, no compliments, no “I saw you’re doing amazing things”.
What I kept manual
Everything after “should I actually send this”. The workflow generates ~200 leads/week with a personalized opener. A human SDR reviews and approves in ~30 minutes. That’s the shape of the leverage: AI does the research grunt work, the human keeps the judgment call.
Where the ROI actually shows up
Reply rate went from 1.2% to 6.8% over 8 weeks. That’s the headline metric, but the more interesting one is that the SDR now handles 3× the volume without more hours — because the manual step is review, not research.
More detailed teardown coming in a future post — including the specific Clay tables and the prompts I settled on after a lot of iterations.