Selling KYC or AML software to a bank takes seven stages. The pivot is a back-test on the bank's own past alerts and suspicious activity reports. A parallel live pilot follows. Model validation and outsourcing review then decide when you go live on one product line or one screening type.
What starts a KYC or AML buying cycle?
Four events start most cycles. Each one tells you what the pilot must prove:
- A regulator finding. On 10 October 2024 FinCEN assessed a record $1.3 billion penalty against TD Bank (FinCEN, October 2024). The bank agreed to an independent AML review and a historical transaction review.
- An alert backlog. Analysts cannot clear the queue, so the MLRO asks for headcount or fewer false positives.
- The EU AML package. The single AML rulebook applies from 10 July 2027 (Arendt, June 2024).
- A new product or market. Instant payments, crypto or a new country bring new customer types.
AMLA will select 40 high-risk entities active in at least six member states during 2027. It supervises them directly from 2028 (AMLA FAQs, checked November 2026).
Find these events in enforcement notices, consent orders, annual reports and financial crime job posts. Write to the MLRO about that event, and name the metric it puts at risk.
Who attends each stage, and what ends it?
The MLRO or Head of Financial Crime owns the deal up to the MOU. After that, model risk, IT and outsourcing teams set the pace. Up to the MOU:
- Discovery with the MLRO and the operations lead. Exit: the champion agrees to share past data for a back-test.
- Data and integration scoping with IT and data engineering. Exit: data protection and security approve the data sharing.
- Back-test, then a parallel pilot, with a model risk observer. Exit: agreed targets met on live data.
- MOU or LOI with the MLRO and the economic buyer. Exit: a signed non-binding MOU.
After the MOU:
- Commercials with the economic buyer and procurement. Exit: an agreed term sheet.
- Validation, legal and outsourcing review. Exit: all reviewers sign off.
- Go-live on one product line or screening type. Exit: the team books a post-launch review.
In discovery, ask for alerts per month, alerts closed per analyst per day, backlog age, and the share of alerts that become a report. Use them as the pilot targets.
How should the back-test work?
The back-test replays the bank's own history through your system and compares results with the analysts' decisions. Agree the design in writing before any data moves:
- Data: a fixed period of past alerts, analyst decisions and the cases that led to a report. Plan to run the test inside the bank.
- True positive retention: flag every case that led to a past report, or explain each miss. Set this target first.
- False positive reduction: measure the drop in alerts at that retention level. The MLRO rejects a reduction without a retention figure.
- Analyst handling time: time analysts on a sample of alerts in both systems.
- Explainability: each alert shows reasons an examiner can read in the bank's case screen.
Name who scores the results. Invite a model risk observer, so validation at stage 6 can reuse the evidence.
Why run the live pilot in parallel?
The parallel run protects the MLRO. On 3 December 2018, US regulators covered pilots run alongside existing BSA/AML processes. A failed pilot "should not subject banks to supervisory criticism" (joint statement). Banks must still meet their obligations, so the old system keeps running.
Send that statement to a US champion before stage 3. Write the pilot targets into the MOU or LOI, so the result forces a decision.
What do model validation and outsourcing review need?
Validation often sets the go-live date. The US rules changed in 2026. On 17 April 2026 the OCC, the Federal Reserve and the FDIC replaced the 2011 model risk guidance, SR 11-7 (Orrick, April 2026). They also withdrew the 2021 statement on BSA/AML models.
The new guidance applies to third-party models, and asks banks to document any vendor customisation. It is aimed most at banks above $30 billion in assets. Generative and agentic AI models sit outside its scope.
Prepare four documents: a model design paper, a training data description, the back-test and pilot evidence, and a change log for every customisation.
In the EU, software delivered as an ICT service falls under DORA, in force since 17 January 2025 (EIOPA). Managed alert review by your staff falls under the EBA's non-ICT third-party guidelines of 18 September 2026 (Arthur Cox, September 2026).
For the people who sign each review, see who signs off on a KYC or AML vendor. The stablecoin payments version of these stages has a different pilot and different reviewers.
What goes live first, and how does the deal expand?
The first slice is sanctions screening, transaction monitoring or onboarding KYC. Pick the one the back-test proved best. Keep the scope small but real, such as retail current accounts in one country.
Review production numbers against the pilot targets after the first quarter. Each next slice reuses the validation file, but any model change goes back to model risk. Price each slice as its own phase.
To see which banks are hiring for financial crime change or answering a regulator now, start with five named buyers, free.