Skip to content
Free guide: How to use AI in compliance
← Back

AML Vendor Due Diligence Checklist: Buying AI-Assisted Review Services as a US Bank

Michelangelo Frigo Michelangelo Frigo (Co-Founder at Zyphe) Published November 12, 2026Updated November 12, 2026
AML Vendor Due Diligence Checklist: Buying AI-Assisted Review Services as a US Bank: evidence, investigation and review workflow

An AML vendor due diligence checklist for US banks buying AI-assisted review services: scope, evidence, AI controls, pilot, contract, monitoring and exit.

Table of contents

An AML vendor due diligence checklist is the list of evidence a bank collects before a supplier performs part of its Bank Secrecy Act and anti-money laundering work. For AI-assisted review services it covers scope, people, data access, AI components, case quality, subcontractors, resilience and exit, with each claim tied to a document and an internal owner.

A vendor demo shows a well-written case memo. It rarely shows who obtained the documents, how the exceptions were resolved, or what happened when a registry was down. Those are the questions that decide whether outsourcing review work strengthens your BSA/AML program or weakens it.

The cost of getting monitoring wrong is not theoretical. On October 10, 2024, FinCEN assessed a record $1.3 billion penalty against TD Bank, which it said "allowed trillions of dollars in transactions annually to go unmonitored for potentially suspicious activity." Buying help with review work is a reasonable response to volume. Buying it without diligence moves the risk instead of reducing it.

This checklist is written for US bank BSA/AML officers working with procurement, information security and third-party risk management.

Who should use this AML vendor due diligence checklist?

This checklist is for US banking organizations subject to the federal banking agencies' third-party guidance. A non-bank fintech can use the operational questions, but its legal obligations differ: it should map its own BSA program requirements and the commitments in its bank-partner agreement rather than assume every bank supervisory expectation applies to it directly.

Volume explains why so many banks are looking at outside help. US financial institutions filed about 4.7 million suspicious activity reports in fiscal year 2024, an average of about 12,870 a day, according to FinCEN's Year in Review. Every one of those filings sits on top of alerts and reviews that someone had to work.

What activity are you actually outsourcing?

Define the queue, the case types, the permitted actions and the output before you assess anyone. For example: preparation of corporate periodic-review files using authorized sources, with ownership discrepancies and screening exceptions routed to the bank's designated reviewer.

Then list the decisions the bank keeps. Acceptance of a prepared file, the customer risk rating, restrictions on the relationship and the decision to file a SAR may sit with different people. Do not collapse them into one "approval" field in the contract or the system. A vendor that prepares the file and a bank officer who decides are two separate controls, and an examiner will want to see both.

What does US third-party guidance expect from banks?

The Federal Reserve, the FDIC and the OCC issued the Interagency Guidance on Third-Party Relationships: Risk Management in June 2023. It replaced each agency's earlier general guidance, including the Federal Reserve's 2013 guidance on managing outsourcing risk, and it remains the reference point for buying review services. On September 11, 2026 the agencies, joined by the NCUA, proposed replacing it with new principles-based guidance; the 2023 guidance applies until a final version is adopted.

Its central principle is stated plainly in the guidance text: "A banking organization's use of third parties does not diminish its responsibility to meet these requirements to the same extent as if its activities were performed by the banking organization in-house." The requirements are to operate in a safe and sound manner and in compliance with applicable laws. The guidance organizes risk management around a life cycle:

  1. Planning
  2. Due diligence and third-party selection
  3. Contract negotiation
  4. Ongoing monitoring
  5. Termination

Use those five stages to structure your file. Most programs are strong on stage 2 and thin on stages 4 and 5, where outsourced review work tends to go wrong.

What belongs on an AML vendor due diligence checklist?

Start with this evidence matrix and add institution-specific requirements. It is a procurement worksheet, not a regulatory approval.

AreaEvidence to requestBank ownerOpen question to record
ScopeSample statement of work and exclusionsBSA/AML sponsorWhich case types need separate agreement?
PeopleRoles, supervision and escalation pathOperations ownerWho handles judgment-heavy exceptions?
Case qualitySource-linked sample files and correction historyCompliance QADoes the sample match the proposed workload?
AccessSystem permissions and source-access methodInformation securityCan the service write, or only read?
DataData flows, storage location, retention, subprocessorsPrivacy and securityWhich parties receive which information?
AI componentsComponent inventory and test evidenceModel risk and technologyWhich steps are statistical, generative or rule-based?
SecurityAudit reports or audit status, penetration test resultsInformation securityWhat is certified today, and what is in progress?
ContinuityFailure handling and a recovery exerciseOperational resilienceWhat happens to work in progress during an outage?
ExitExport format, deletion and transition planProcurement and business ownerCan the bank continue the queue after termination?

For each answer, record the supporting document, its date, the reviewer and any open question. "Available on request" is a follow-up item, not a completed assessment.

How do you evaluate the AI inside a review service?

Ask the supplier to name each component and its intended use. A numerical risk score, a language model that drafts the memo and an agent that chooses which source to query next need different evidence, and the bank has to map each to its own governance.

The model risk framework changed in 2026. SR 26-2, joint guidance from the Federal Reserve, the OCC and the FDIC issued April 17, 2026, supersedes SR 11-7 and SR 21-8, the 2021 statement on model risk management for BSA/AML systems. It covers traditional statistical models and places generative and agentic AI outside its scope, which leaves those steps to the bank's own AI risk governance rather than exempting them from control. Our guide to AI model validation in compliance sets out a component inventory you can hand to the vendor.

Request concrete evidence rather than an accuracy percentage:

  • examples of unsupported statements the system produced and how they were caught;
  • missed escalations found in testing, with the fix and the retest;
  • what the system does when a required source is unavailable;
  • how a change to a model, prompt, source or permission is tested and approved before it reaches your cases.

What should a pilot prove before you sign?

A pilot should prove that the service can handle your workload, not a curated sample. Build it this way:

  1. Use the bank's representative cases and an agreed reference assessment made independently of the vendor.
  2. Include a clean file, an ownership conflict, an unavailable source and an ambiguous screening result.
  3. Measure first-pass acceptance, material errors, missed escalation triggers, retained reviewer time and complete cost, with denominators and the observation period stated.
  4. Test the path for a case the service cannot finish, and check that it reaches the right bank owner with the evidence attached.
  5. Decide the pass and stop criteria before the first case runs.

The review pilot scorecard keeps quality, escalation and cost measures separate, so a high acceptance rate on simple cases cannot hide weak performance on complex cross-border ownership. If the service is aimed at an existing backlog, our KYC remediation plan shows how to segment the queue before you test on it.

What needs to be settled in the contract?

Translate the pilot's operating assumptions into the agreement. Specify:

  • acceptable inputs, quality standards and the rework process;
  • incident notification timelines and who can stop execution;
  • how changes to sources, components or subcontractors are notified and approved;
  • the bank's right to inspect work and obtain records, including in an examination;
  • what happens when source access expires or a dependency fails;
  • dependencies on the bank, such as a policy question that must be answered before work continues, so delays are attributed correctly.

If a responsibility affects a live customer file, it belongs in the contract or the operating procedure, not in a sales presentation.

What does ongoing monitoring look like after go-live?

Monitoring is the life-cycle stage that most often lapses, and it is the one an examiner will test. Reconcile case counts against delivered work every period. Sample accepted files as well as rejected ones, because errors can sit inside work that passed. Watch changes in source availability, case complexity, correction rates and the age of unresolved work.

Review significant workflow changes before you expand the scope, and keep the ability to suspend one problematic step without losing the evidence already collected.

Enforcement history puts a price on these failures. Zyphe's AML enforcement tracker records 32 enforcement actions and breaches from 2023 to 2026, about $7.6 billion in penalties across 10 jurisdictions, each linked to its source.

How do you exit a review vendor without losing the queue?

Plan the exit before you sign. For review services the risk is concrete: incomplete cases, evidence in a format only the vendor can read, and no one at the bank who knows the queue's state.

Agree in advance:

  1. the export format for completed files, evidence and decision records;
  2. how in-progress cases are handed back, with their status and open questions;
  3. how and when access is removed and data is deleted or returned;
  4. the transition assistance the vendor will provide, and its cost.

How should a bank assess Zyphe?

Zyphe operates review desks for periodic KYC review, UBO and EDD, sanctions and PEP alerts, adverse media and transaction monitoring triage. Agents prepare each case against the bank's own policy with the evidence and written rationale attached, Zyphe adds human quality assurance on managed workflows, and the bank's reviewers approve. Decisions such as the customer's risk profile and suspicious activity reporting stay with the bank.

The items a diligence team usually asks about first are published on the security page:

  • Data storage: identity data is encrypted and split into shards across distributed storage nodes, with no central store of customer PII. US end users' data stays in US regions, and EEA and UK data stays on EEA and UK nodes.
  • Audit trail: each output logs the sources reviewed, the reasoning applied and the action taken.
  • Retention: Zyphe retains verification results, audit logs and cryptographic proofs for the term of the agreement or as required by law.
  • Certifications: Zyphe is not yet SOC 2 or ISO/IEC 27001 certified. As of October 2026, the SOC 2 Type I report was targeted for October 2026 with the Type II observation period starting in Q1 2027, and ISO/IEC 27001 Stage 2 was targeted for November to December 2026. Check the security page for the current status.
  • Testing: penetration test reports are available to customers under NDA.

Rates for automated work are public on the pricing page; managed operations are scoped and priced around the workflow and volume. To see what a finished file looks like before a pilot, open the synthetic EDD review, and compare operating models across providers on the AI compliance agents comparison.

General information, not legal advice.

Michelangelo Frigo Michelangelo Frigo (Co-Founder at Zyphe) Michelangelo Frigo is a privacy and identity infrastructure expert and co-founder of Zyphe.

Frequently Asked Questions

An AML vendor due diligence checklist should include the outsourced scope and retained decisions, the vendor's people and supervision, sample case files with correction history, system access and data flows, an inventory of AI components with test evidence, security audit status, continuity arrangements and an exit plan. Each item needs a supporting document, a date, a bank owner and any open question.

A bank remains responsible for its BSA/AML obligations when it outsources review work. The 2023 interagency third-party guidance states that a banking organization's use of third parties does not diminish its responsibility to meet its requirements. The vendor can prepare files and perform agreed tasks, but the bank must oversee the work and keep the decisions reserved to it.

US banks buying AI-assisted AML services should start with the June 2023 Interagency Guidance on Third-Party Relationships and SR 26-2, the April 2026 joint model risk guidance from the Federal Reserve, OCC and FDIC. SR 26-2 covers statistical models and leaves generative and agentic AI to the bank's own governance, so both documents shape the diligence file.

A fintech can use the operational questions in a bank's vendor due diligence checklist, such as scope, evidence, access, pilot design and exit. It should not assume bank supervisory guidance applies to it directly. A fintech should map its own BSA obligations and the requirements in its bank-partner agreement, which may be stricter than the general guidance.

A vendor's SOC 2 report does not replace AML due diligence. Security assurance and case-quality evidence answer different questions: one tests controls over systems and data, the other tests whether the review work meets the bank's standard. A diligence file needs both, plus the vendor's honest status where an audit is still in progress.

A pilot with an AML review vendor should test representative cases against an independent reference, including a clean file, an ownership conflict, an unavailable source and an ambiguous screening hit. It should measure acceptance, errors, missed escalations, reviewer time and full cost, and it should prove how unfinished cases reach the right bank owner.

The exit terms that matter most for an outsourced AML review service are the export format for completed files and evidence, the handback of in-progress cases with their status, the removal of access and deletion of data, and the transition assistance the vendor will provide. Agree them before signing, because they are hard to negotiate during a dispute.

A bank should assign every unresolved diligence finding to an internal owner with the authority to decide whether it changes the scope, adds conditions or stops the purchase. Record the finding, the decision and the date. Unresolved items left in an email thread tend to resurface during an examination, without a record of who accepted the risk.

AML compliance without the PII liability

Screening, monitoring and reporting built on a privacy-first identity layer.

See Zyphe AML