The challenge
A majority of Pennsylvania’s licensed CPAs, roughly 16,000 people or about 62%, weren’t PICPA members. The opportunity was obvious. The prospect list wasn’t much help: names and license numbers only, without the firm, role, or seniority you’d need to reach anyone thoughtfully. Researching one person properly takes 20 to 30 minutes across multiple sources, which works out to an estimated three to four person-years of manual work. Commercial data vendors tested poorly, and they raised the kind of compliance and brand questions a professional association is obligated to ask.
The approach
I built a four-stage multi-agent AI pipeline:
- Anchor each person’s identity, rejecting the look-alikes
- Corroborate facts across five independent public sources
- Weigh the evidence by source reliability
- Verify every value against the page it was cited from
Two design choices mattered more than any model choice. The pipeline runs locally, so sensitive member data never leaves the building. And it declines rather than guesses: every field carries a source and a confidence level, and when the evidence isn’t there, the field stays empty.
Head-to-head, zero errors
In a bake-off against a commercial data tool, the pipeline scored 7 out of 10 with zero errors. The vendor managed 2–3 out of 10 and was confidently wrong.
The results
Thousands of records enriched, and an accuracy audit that found zero wrong entries. The output is a segmented, source-cited dataset behind a seven-figure membership opportunity, built solo on a flat ~$200/month AI subscription, without per-record metering or a data-vendor contract.
What I learned
The speed matters less than the trust. A pipeline that guesses confidently poisons its own results: the first wrong answer someone acts on discredits every row that was right. Provenance, confidence levels, and the discipline to leave a field blank are what make AI output something a team will build a campaign on.