Unit21: fraud, AML operations and governed AI investigations

Investigators work with linked customer, instrument, transaction and alert records. Compliance teams define detection logic and escalation procedures; risk leaders oversee queues, quality control and reporting. Case Management supports research requests, external enrichment, configurable queues and user permissions, with operational measures such as closures, investigation time and movement from alerts to cases and SARs.[3]

The Agentic Task Builder extends investigation agents through two published task types: online research and analysis of the customer’s own data. Teams describe the task, choose identifying fields and review a preview before saving it. Search tasks can separate verified findings from uncertain matches and retain source links; data-analysis tasks examine activity associated with the alert. The institution supplies the question and the acceptance criteria.[12]

The same page’s FAQ separately describes human review before closure, so its wording is not fully consistent. The applicable queue-level autonomy and approval settings need explicit confirmation.[11]

Case Agent, released in June 2026, assembles the attached alerts, entity records, adverse-media findings and network connections into a case narrative before review. The separate AI Agent API Suite lets customers register their own agents, assign supported tasks, launch runs and follow their progress without using the browser interface. These are published production features; access and activation remain part of the customer implementation.[16]

Reporting includes prefilled information, templates and deadline management. Continuing-activity tracking supports follow-up SAR work. The October 2026 update adds configurable SAR field mappings, including entity identity data, instrument account numbers and tags. A supported form or export is a defined reporting route, rather than evidence that every jurisdiction or filing obligation is covered.[13][18]

The July 2026 TRM integration writes wallet risk levels, indicators, categories, volumes and entity associations into Unit21 records and sends TRM monitoring alerts into case management. That announcement explicitly described post-transaction monitoring as live and real-time wallet screening on transactions as planned. The reviewed later releases do not substantiate general availability of that planned extension.[15]

The remote MCP server allows an external assistant to access authorised Unit21 context and reporting. An administrator must enable it and grant MCP read access; the assistant inherits the user’s existing permissions, and revocation cuts access. Custom webhooks can carry selected alert and case fields into other systems. In September, MCP added tools for discovering valid report filters and values.[17][18]

A more specific operational integration is Corelation KeyStone. Its dedicated extract loads historical records once, then sends changes about every five minutes. KeyBridge API calls let authorised analysts block or activate cards, place stop payments and place check holds from a case. Requests, responses, user identity and failures are recorded. Unit21’s implementation team, the customer’s KeyStone administrator and data team configure the extract, map fields and validate write-back credentials in a sandbox.[19]

Unit21 does not publish a general numeric tariff in the reviewed sales material; commercial enquiries go through its contact and demonstration process. Fees and the initial term follow the statement of work.[26][20]

The 19 December 2024 Terms of Service provide for US-dollar invoices payable within 30 days. If the statement of work specifies no renewal period, successive 24-month renewals apply, with written non-renewal at least 90 days beforehand. Negotiated terms may differ.[20]

The separate Transaction Screening Terms dated 15 January 2026 restrict licensed use to the agreed users, location and internal purposes. Contracted limits can be based on seats, assets, accounts or another specified measure; additional rights may be required to exceed them. Certain data also requires a separate Unit21 or third-party agreement. Its headings and text refer to both transaction screening and the Watchlists feature, so feature applicability belongs in the signed commercial scope.[21]

The rule lifecycle offers distinct test environments: historical validation measures a proposed rule against past activity, shadow mode runs it alongside live activity without acting on customers, and the sandbox supports experimentation. Promotion into an active decision flow is a separate step. This makes incoming data, chosen variables and the customer’s own dispositions central to implementation and later tuning.[5]

Optimization results describe the historical sample used, not prospective detection accuracy. The published safeguard preserves sampled entities associated with filed SARs. Its recommendations still pass through the customer’s validation and shadow testing. This also makes consistent disposition labelling an operating requirement: poor labels directly weaken the evidence available for tuning.[23]

The December 2024 terms require migration-assistance requests at least 15 days before termination. Reasonable assistance normally runs for up to two weeks unless agreed otherwise, with continued service fees and chargeable migration work.[20]

The Support Center provides ticket submission, documentation links and separate status pages for US-Prod1, US-Prod2, US-Prod3 and EU-Prod. The public material does not establish a universal support timetable or response target. The services agreement points to a separate SLA, but its detailed text was not publicly readable in this review. Existing customers are directed to their customer success manager to activate newer capabilities.[24][20][16]

The feature-specific screening terms offer remote support to one designated contact, 06:00–18:00 Mountain Time on weekdays excluding national banking holidays. The published telephone is a placeholder, and the document mixes Transaction Screening and Watchlists terminology. These hours need confirmation for the contracted feature and do not establish platform-wide support.[21]

Status reporting covers individual components including ingestion, APIs, real-time monitoring, filings, webhooks and AI agents. On 12 October 2026, the US-Prod1 history recorded failed or delayed AI reviews associated with a Databricks serverless-endpoint outage; it was marked resolved that afternoon. This is a documented dependency incident, not an uptime statistic for other environments or the whole platform.[25]

Unit21’s current AI page states that customer data is not used to train its models or other customers’ models. It describes analyst overrides as review evidence rather than an automatic model-training loop. Agents retain the data consulted, checks, sources and assumptions behind their outputs. These are published product commitments.[11]

The December 2024 services terms separately permit service-improvement and consortium uses and allocate derived/usage-data rights. Customers own supplied data and must secure necessary permissions. These provisions sit alongside the narrower AI training statement.[20]

The April 2026 privacy policy expressly excludes personal information processed or hosted on behalf of customers. It covers Unit21’s own website and business interactions, so it does not establish production-case retention, residency or a customer’s complete data-processing terms.[28]

Unit21 reports SOC 2 Type I and Type II examinations, encryption in transit and at rest, segmented networks, production multifactor authentication, access auditing, regular penetration testing and encrypted daily backups. Reports are requested through its trust process. These are supplier security disclosures, separate from financial-services permissions or a contractual recovery-time commitment.[27]

Unit21 supplies the tools with which an institution operates its controls. The institution defines its risk rules, review policy and filing decisions. Direct reporting connections and workflow automation do not transfer the institution’s regulatory obligations or establish that Unit21 executes trades, holds client money or provides regulated brokerage services.[1][13]

The company’s March 2026 relaunch described a shift from separate fraud and AML tools toward an operating platform organised around AI agents. The older Fraud DAO name relates to its shared fraud-intelligence activity; the current product is presented as the Fraud Consortium.[31][10][9]

Operator

Unit21, Inc.

Service scope

Fraud and AML monitoring, customer risk ratings, sanctions and payment screening, investigations, AI-assisted rule development, regulatory filings and shared fraud intelligence.

Customer types

Banks; Credit unions; Fintechs; Payment businesses; Digital-asset companies; Financial intermediaries

Delivery

SaaS; API

Pricing publication

Not disclosed

Platforms

Unit21; AI Agents; Case Management; Customer Risk Rating; Fraud Consortium; Rule Writer Agent; Rule Recommendations

Integrations

Corelation KeyStone; KeyBridge; Jack Henry Symitar; PortX; Fiserv DNA; Persona; Middesk; Trulioo; LexisNexis Bridger; Fingerprint; Chainalysis; TRM Labs; MCP; Webhooks

Pricing scope

No general numeric tariff published; fees and licensed scope are set in the customer agreement.

Implementation

Data integration, mapped customer and transaction records, rule configuration, historical validation, sandbox and shadow testing, followed by controlled production deployment.

Support

Support Center tickets, customer success contacts and environment-specific status pages; general platform response targets were not established; separate screening terms specify limited weekday support.

Permissions and regulatory scope

Financial-crime software and reporting workflows. Institutions retain risk-policy, review and filing responsibilities; product integration is not a brokerage or custody authorisation.

Provider metadata checked 21 Oct 2026Source 1Source 20Source 2Source 3Source 4Source 5Source 7Source 6Source 10Source 9Source 11Source 13Source 14Source 15Source 19Source 17Source 18Source 24Source 21Source 26Source 22Source 23

Recent operational changes are specific enough to affect workflow design. September’s release added automatic Symitar account warnings through PortX and visibility into why NACHA or X9 enrichment records matched, remained pending or failed to match. Those functions connect investigation outcomes and data-quality checks back to the banking records on which detection depends.[18]