Prove your AI agents and models are safe
Your AI investigator writes a confident case note. Is it right?
Teams want AI agents to work fraud and AML alerts because queues are long and most alerts are nothing. But real case files make a poor test: a closed alert is not proof of innocence, and the rings you know about are the ones you already caught. Score an agent on past decisions and you mostly learn how well it copies them, blind spots included.
Built for Fraud and AML leaders piloting AI agents on alerts or case dispositions; Head of Fraud Ops, MLRO, Director of Financial Crime
The problem in numbers
What the public record shows
Every figure links to the page it came from, with its date.
2% to 4%
of alerts serious enough to escalate and file a suspicious activity report, per a 2025 benchmarking study by an industry analyst firm
sanctionscanner.com, 2026-04
15 percent
of daily work decisions Gartner expects AI agents to make on their own by 2028
theregister.com, 2025-06
130
vendors Gartner estimates are real (an "about" figure), out of thousands claiming agentic AI products
theregister.com, 2025-06
What you get
The deliverables
You get a synthetic back office (customers, transactions, alerts, KYC files, disputes) with planted fraud rings and mule networks and a hidden answer key, and a scored run of your investigator agent against it.
A synthetic back office
Customers, transactions, alerts, KYC files and disputes that behave like a real operation, with planted fraud rings and mule networks.
A hidden answer key
Every member of every planted scheme, including accounts that never raised an alert. It is never shown to the agent or whoever tunes it.
A scored run
What your agent found, what it missed, how many innocent customers it escalated, and whether its stated evidence supports its conclusion.
How it runs
Agreed on a short call, set by the agent and the case types, start to finish
Agreed on a short call, set by the agent and the case types
Scope the pilot
We agree which alert types and schemes matter to you and how your agent takes in cases.
Your agent works the cases
The agent runs on the synthetic back office the way it would on real alerts.
Score and readout
Results are checked against the answer key and walked through with your fraud and AML leads.
Method
How the answer is produced
Truth known in advance
Because the schemes were planted, the right answer is known for every case, including the misses a normal pilot never sees.
Connected cases
Rings are spread across several alerts, so you see whether the agent links them or treats each alert alone.
No customer data leaves
Everything is synthetic, so the test does not need data-sharing approvals before it starts. The key covers known typologies, not every possible scheme.
Why Amir
Designed and led by someone who has built fraud models in production
The patents closest to this problem:
Amir is a named inventor on US 12,541,721 B2 (Method for extreme class imbalance within fraud detection, filed at NICE Actimize; granted).
Fraud is a tiny share of cases, which is why rare planted schemes are needed to test an investigator, and this patent covers building models under that extreme imbalance.
View on Google PatentsAmir is a named inventor on US 2023/0267468 A1 (Unsupervised clustered explanation-based feature selection using transfer learning for low fraud scenario, filed at NICE Actimize; pending).
Amir is a named inventor on this patent application for detecting fraud when there are almost no confirmed examples, the same missing-labels problem that makes real case files a weak test.
View on Google PatentsThese patents were filed at the companies named. They show the problems Amir has worked on; they are not Riskernel products.
Is this you
You probably need this if
- You are piloting an AI agent to triage alerts or draft case dispositions.
- A vendor is selling you an AI analyst, shown on cases of its choosing.
- Your alert backlog is growing faster than your team.
- You approved the pilot because the agent's write-ups read well.
- You have no way to count the bad actors the agent cleared.
Questions
Before you reach out
Do you need our customer data?
No. The back office is synthetic, so no customer data leaves your organisation.
Can we test a vendor's agent?
Yes. Any agent that can read cases and return a disposition can be run against it, including one you are still evaluating.
Does it catch every kind of fraud?
No. The answer key reflects typologies Amir knows how to plant. It is a strong test of known patterns, not proof against every possible scheme.
Can we rerun it later?
Yes. A new version of the agent or its prompts runs on a new back office with the same scheme types planted at the same difficulty. The agent cannot have memorised the cases, and scores still compare across versions.
Scope AI Investigator Proving Ground with Amir
Tell Amir what you run and what worries you. He replies with a short scope outline.