Investigating suspicious payments with an AI that admits what it doesn't know
Banks have to look into every transaction that might be money laundering. Most turn out to be nothing, but each one takes a lot of digging. In Alert Desk, an AI agent does the digging and writes up what it found. A person still makes the call.
Most of the work is fetching, not deciding
An alert shows up with almost nothing: an amount, a time, the rule it tripped. Everything that explains it lives somewhere else. So the analyst goes and collects it, writes a note, and moves on to the next one. Industry estimates say only 1 to 5% of alerts ever become a report.
Fetching is a good job for an agent. Deciding isn't. Someone signs their name to that decision, so it stays with a person.
Here's the catch. When an AI hands you a neat answer, agreeing feels free and disagreeing feels like work. Get that wrong and you end up with a tidy record of decisions nobody really made. So a lot of this design is friction, on purpose.
It tells you what it couldn't check
Every sentence the agent writes points to a source. If it can't find one, the claim doesn't go in. It goes into "What the agent couldn't confirm" instead, a section that never collapses and looks just as important as the summary. The agent writes it in the first person ("I couldn't find this vessel"), because a gap with a name on it is harder to skim past.
Departure from baseline
This instruction gives its purpose as “charter hire, vessel MV Celestine” 4. No vessel of that name is on the customer’s 14-vessel register or in the two commercial registries I searched 5.
What the agent couldn't confirm
- I could not confirm or exclude the match to list entry ENT-30417. The payment message carries no registration number or address.
- I could not find MV Celestine in the two registries available to me. A specialist maritime registry was not searched.
Saying no is as easy as saying yes
Most approval screens make "Approve" the big button. Here, that would be the bug. All three choices look the same, nothing is picked for you, and the agent's suggestion is just a small label. The bold button only shows up once you've chosen, when you confirm.
The effort goes into the reasoning instead. Every decision needs a short note, at least 40 characters, because one-word closures are what auditors flag. Go against the agent and a note asks what it missed. That's logged as feedback on the agent, not on you.
Three ways to show the agent's opinion
If you see the AI's answer before you've looked yourself, it's hard not to lean on it. I tried three versions.
One click to agree, two to disagree. Fast, but reviewing turns into rubber-stamping.
Agreeing and disagreeing cost the same, but the suggestion is always staring at you.
The suggestion only shows up when you move to decide. The summary still says what the agent thinks, so nothing's hidden. The buttons just stay neutral until you need them.
The clock starts when a person suspects, not when the alert fires
A suspicious activity report (SAR) is due 30 days after someone decides the activity might be suspicious. Not 30 days after the alert. Teams mix these up, and examiners notice. So until a person records suspicion, the screen just says No deadline yet. Once they do, it shows both dates and who started the clock.
- 21 Sep, 14:23Alert raisedPayee may be on a sanctions list (rule SAN-PAY-02). No deadline yet.
- 24 Sep, 10:15Suspicion recorded by J. Whitfield (L1)The 30-day clock starts here, with the reason quoted in the record.
- 24 OctSAR dueThe queue ranks this first: 2 days left, payment on hold.
Flagging a case isn't filing a report
These are two decisions, made by two different people. The investigator decides whether a case is worth escalating. A senior compliance officer, the BSA Officer, decides whether to file a report with the government. Their screen opens with why it was escalated, whether the investigator disagreed with the agent, and how long they have. Filing opens a draft built from the agent's write-up, and it flags any line that guesses what the customer intended. Nothing gets filed until they file it.
- Main reason
- Structuring
- Agent
- Agreed · high confidence
- Time allowed
- 30 days (suspect known)
- Minimum
- $5,000 · $131,400 involved
The customer can never find out
This is the strictest rule in the product. You can't tell someone they're being investigated, or even hint at it. So the tool never contacts the customer. If the investigator needs more information, the request goes to the customer's relationship manager inside the bank, with wording that sounds like routine paperwork and a list of things they must not say. That warning can't be dismissed.
As part of routine account maintenance, we are refreshing the information we hold for Priya Raman. Could you ask them for an updated source-of-wealth declaration, with supporting documents?
RM must not mention- A review, alert, investigation or report.
- Which transactions led to the request.
- Any change to the account while the request is open.
Every decision leaves a trail anyone can follow
Each decision is written down the moment it's made: who decided, why in their own words, and a snapshot of what the case looked like right then. Two years later, someone who wasn't there can see what the analyst saw, not what the case looks like today. Nobody can edit it.
You can tell what's official by the font
Anything that becomes part of the official record is set in a serif. Everything else is just the interface. You don't have to read a word to know which is which.
Details
- Ranking shows its workEvery alert says why it's where it is in the queue. Open it to see the five things that decide the order.
- Unused evidenceThings the agent found but didn't use are marked, not hidden. In the case where the agent says "clear," the two earlier alerts it never looked at are sitting right there.
- Close matchesA name match always shows how close it is (88% alike, against an 85% cut-off), so you know whether to trust it.
- Help when you want itHow-it-works notes sit behind small (i) icons that open on hover, tap or keyboard. Evidence is never tucked away like that. The icons wiggle until you open one, then stop.
- MotionTooltips fade in over 150ms, buttons press in 120ms, and everything holds still if you've asked your device to reduce motion.
- KeyboardPress 1, 2 or 3 to decide. Clicking a citation jumps to the evidence and back.
- US rulesUS terms throughout. An early version mixed in UK terms and Nigerian branch names. Fixing it meant rebuilding the data, not a find-and-replace.
How I'd know it works
Nothing here has shipped, so I can't claim results. If it did, I'd watch four things: how often QA finds mistakes in closed cases, whether a stranger reading a closure note would reach the same decision, how often analysts disagree with the agent over time, and how much of their day moves from digging to deciding. I wouldn't chase speed. A tool tuned for alerts per hour would make the real problem worse.
Open questions
- Does hiding the suggestion actually reduce bias? Or does it just move it? That needs real analysts, not a design review.
- What's a healthy disagreement rate? Too low looks like rubber-stamping. Too high means the agent isn't helping.
- When should the clock start? Here, escalating starts it if nobody has. Some teams start it at the senior review. I'd ask someone who does this job.
- Does marking unused evidence scale? It works with nine items. With two hundred, it might just be noise.
Let's talk
If you work alerts for a living, or you're building tools like this, I'd love to hear what I got wrong.