False positives are the real cost of screening
By Krisan Nichani, Chief Compliance Officer
Summarize inChatGPTClaudePerplexityGemini
Key takeaways
- Sanctions screening matches wide on purpose (transliterations, aliases, partial names), and that same width is what generates false positives.
- Alert volume scales with transaction volume long before true matches do, which is why compliance headcount tracks alert volume rather than payment volume.
- The obvious fixes fail: narrowing match logic cuts true hits along with noise, and re-reviewing at every institution multiplies the same coincidental match.
- On Infinite, an alert adjudicated once carries its resolution to every network participant; screening still runs on every instruction, but redundant re-review disappears.
- AI assembles each alert into a reviewable case, but the model never clears one on its own: human-approved policy governs dispositions, and judgment calls go to named analysts.
Matching wide catches more than it should
Sanctions and watchlist screening cannot afford to match only exact strings: a name transliterated from another alphabet, a legal entity with several trading names, a common name shared by thousands of unrelated people all have to clear the same check. So screening engines match wide on purpose: fuzzy logic, phonetic matching, partial-name scoring. That is what makes the check work. It is also, unavoidably, what generates volume no rule set can separate from noise ahead of time.
The result is a review queue with an odd shape: the overwhelming majority of alerts resolve as false positives (a shared surname, an abbreviated entity name, a coincidental match to a list entry), and only a small fraction are true hits. An analyst still has to open every one, because the ones that matter are indistinguishable from the ones that don't until someone looks.
Alert volume, not transaction volume, sets the cost
This is why compliance headcount tracks alert volume rather than payment volume. Add a corridor with a common naming pattern, or a counterparty type with high name-collision risk, and the queue grows faster than the business that produced it. A team sized for last quarter's alert rate is undersized the moment the mix of countries or counterparties shifts, and the fix, more reviewers, scales cost linearly with a problem that is mostly noise, not risk. Sanctions screening and AML transaction monitoring both work this way: high recall, a high false-positive rate, and review work on every alert regardless of how it resolves.
The alert queue is not a measure of risk. It's a measure of how many names looked like other names this week.
Two responses do not work. Narrowing the matching logic to cut noise also cuts recall: the same tuning that clears more false positives clears some true ones with it, which is not a trade a compliance program can make. And re-running the same review at every institution a counterparty touches multiplies the noise instead of dividing it: the same coincidental match gets adjudicated, separately, everywhere that counterparty transacts.
What shared vetting changes
A compliance network changes where duplication lives, not how matching works. On Infinite, a counterparty screened and cleared once carries that outcome to every participant on the network. The compliance network thesis makes the underlying argument for onboarding; the same logic applies to ongoing screening. A name already adjudicated as a false positive does not get re-flagged and re-reviewed at the next institution; the resolution travels with the counterparty, the same way the initial vetting does.
That does not touch the matching logic itself: screening still runs wide, on every instruction, on every rail, because that is what keeps the check honest. What the network removes is redundant adjudication: the same coincidental name match, resolved once, instead of resolved from scratch by every participant it happens to reach.
Where AI fits, and where it stops
Assembling an alert into a reviewable case is exactly the kind of work worth automating: pulling the counterparty file, the match details, the transaction history, and any prior resolution for the same name into one packet an analyst can act on in minutes instead of an hour. Infinite Agents does that assembly for every alert on the network.
Clearing the alert on its own authority is not part of that job. Dispositions follow human-approved policy, and a match that requires judgment goes to a named analyst, who makes and documents the call; a strict-liability check leaves no room for a model's discretion. Faster assembly changes how long the decision takes to reach. It does not change who is accountable for it.
Frequently asked questions
Why does sanctions screening produce so many false positives?
Because screening matches wide on purpose: transliterations, aliases, and partial-name scoring catch real risk that exact-string matching would miss, and that same width produces coincidental matches to common names and entities. High recall and a high false-positive rate come from the same design choice.
Does shared vetting reduce false-positive alert volume?
It reduces redundant review, not the match itself. Once a counterparty's alert has been adjudicated, that resolution carries to every participant on the network instead of being re-reviewed from scratch at each one. The queue shrinks even though screening still runs on every instruction.
Can AI clear false-positive alerts without a human decision?
Not on its own authority. AI can assemble the case (match details, transaction history, prior resolutions) into a packet an analyst reviews in minutes. Dispositions follow human-approved policy, and a match that requires judgment goes to a named analyst, who makes and documents the call; the model never exercises that discretion itself.
See it on your own flows.
A 30-minute walkthrough of a live account and the screening record behind each payment.