I needed to research 16,000 people.
Done properly, one person takes 20 to 30 minutes: confirm the identity, find the firm, figure out the role, check that you haven’t mixed them up with someone who shares the name. Multiply that across the list and it comes to an estimated three to four person-years of full-time work. I didn’t have a team. I had a flat AI subscription that costs about $200 a month.
Here’s how the list got that big. At PICPA, the Pennsylvania Institute of CPAs, a majority of the state’s licensed CPAs, about 62%, weren’t members. Leadership wanted to change that thoughtfully, but the prospect list held names and license numbers and nothing else. You can’t write a thoughtful letter to a license number.
The obvious answer was a commercial data vendor. We tested that first, and the test settled it: scored head to head, the vendor managed 2–3 out of 10, and it delivered its wrong answers with total confidence. For a professional association, confident wrong answers are a compliance problem and a brand problem at the same time.
So I built the researcher I needed out of AI agents. The pipeline treats each person as a small research project: it pins down identity before anything else, gathers facts from five independent public sources, weighs the evidence by how reliable each source has proven, and then checks every value against the page it was cited from. In the same head-to-head test, it scored 7 out of 10 with zero errors.
Two rules shaped the design more than any technical decision. Member data stays local, on machines we control. And when the evidence is thin, a field stays blank; a blank cell is honest, and a plausible guess is a small landmine for whoever acts on it.
The result: thousands of records enriched, an audit that found zero wrong entries, and a segmented, source-cited dataset behind a seven-figure membership opportunity. The monthly bill never changed.
The lesson
Accuracy is the product. The first time a colleague acts on a wrong answer, they stop trusting the whole dataset, and no amount of speed buys that trust back. What made this one hold up is that it’s checkable: every value carries its source, so nobody has to take the machine’s word for anything.