Before AgentGov had a single paying customer, it had a dataset: New Jersey's own Open Data Portal, public and free for anyone to download.
We pointed our first fraud detection agent at 505,713 pension and payroll records — the kind of dataset a state fraud unit sees every day, except this one was already public. No special access, no data-sharing agreement, no pilot contract. Just records anyone could have looked at.
The agent came back with 25 top-ranked investigation leads.
The count wasn't the point
It would have been easy to stop at "we found 25 anomalies in half a million records" and call that the pitch. But a raw anomaly count is not useful to an investigator. Anomalies are cheap — every payroll system has thousands of them, and most are typos, timing quirks, or perfectly legal edge cases.
What we learned early, and learned the hard way, is that the hard part isn't finding statistical outliers. It's knowing which ones matter under the law that actually governs them.
One pattern in the data looked, on its face, like double-dipping — someone drawing income from two sources that shouldn't overlap. We flagged it. Then we learned some of it is legal. Certain overlaps are explicitly permitted under specific statutory exceptions. An agent that doesn't know that isn't saving an investigator time — it's handing them a stack of dead ends dressed up as leads.
That's the reason AgentGov's fraud detection agent doesn't just do anomaly detection. Before anything becomes a lead, it runs a legal cross-reference pass against applicable statute and precedent, and filters out what the law already excuses. What's left is a queue an investigator can actually work, not a queue they have to re-investigate from scratch.
Why we built it this way
We started with fraud detection because it's the sharpest version of a problem every government office has: too many records, too few people, and no time to read all of it. Public payroll and pension data was the fastest way to prove the approach against something real, with results anyone could check against the same public source.
Every lead in that first run came with a confidence score, the source record it was drawn from, and the statute or precedent that made it worth a second look. That's not a nice-to-have — it's the difference between a tool that displays data and one that hands back something a person can act on.
Fraud Detection is live today, and agency pilots are in progress. If your office is sitting on payroll, pension, or benefits data and wondering what's actually in it, we'd like to talk.