Neil Katz on Why Investigators Sample, and What Breaks When They Stop
Investigative teams examine a fraction of what they hold, not because a fraction is the right amount of scrutiny, but because it is all the attention available. Valantor Chief Product Officer Neil Katz expects AI to remove that limit within five years. The harder question is what a full review is worth if nobody can trace where its conclusions came from.
Insurance fraud steals at least $308.6 billion a year from American consumers, by the Coalition Against Insurance Fraud's own estimate. Proving any single case means working through the claim file, and there are more claim files than hours. So investigative teams sample: a slice of claims, a fraction of transactions, whatever the available time covers. The sample size gets defended as methodology. It is a budget.
Valantor Chief Product Officer Neil Katz, in an interview with CyberTech Intelligence, argues that AI removes the budget. He says organizations sample a small percentage today because of the volume, and expects them to move toward full investigative coverage inside five years. His blunt version of it: audit every transaction rather than 5% of them.
What that breaks matters more than what it fixes. Coverage is only worth having if the conclusions survive an appeal, a regulator, or opposing counsel. Examine all of the evidence with a system that cannot show where a finding came from, and you have replaced a small defensible review with a large indefensible one.
“AI should expand the investigator's field of vision, not assume the investigator's authority.”
Katz's answer is to stop the system short of the conclusion. What it surfaces is evidence to be challenged, not a finding to be accepted, and the investigator decides what is justified and what action follows. FraudX, Valantor's packaged insurance-fraud product on the GroundX platform, is built that way. Every finding links back to the document behind it: the medical record, the deposition, the bill, the photograph. Several independent methods work the same evidence separately and their conclusions get compared rather than merged, so agreement raises confidence and disagreement stays visible.
Throughput is the wrong thing to watch, and Katz names four better items: accuracy against known scenarios, the quality of the evidence retrieved, how often the independent methods disagree, and how often investigators overturn what the system surfaced. Decision quality can fall while speed holds steady, and no productivity dashboard will show it.
Which comes back to the sample. Going from 5% to 100% is a gain only if the 100% holds line by line. Katz's future is machine-scale review with human judgment at the end of it, and the judgment half is the part that has to be engineered rather than assumed.