Agentic Screening: Mapping the AML/KYC AI Agent Market

2026 AI Agent Market Map

13 vendors

8 criteria

4 quadrants

Which vendors offer agentic AI for AML/KYC screening, and how do they compare?

We identified 13 vendors that offer agentic AI for AML/KYC screening as part of their core solution. Their capabilities vary widely. Some agents support individual tasks, while others screen, investigate and adjudicate alerts across the workflow. The vendors also differ in how their systems are tested, implemented, monitored and documented for regulatory use.

Screening provides a clear basis for comparing these solutions. Financial institutions must continuously screen customers and payments against sanctions, PEP and adverse media risk. The work is high-volume, but the tolerance for error is low. Missed risk can lead to regulatory action, financial penalties and reputational damage.

Using AI in this workflow requires more than faster processing or fewer alerts. Institutions need evidence that the agent can apply their policies, produce accurate and consistent decisions, explain each outcome and maintain a complete audit trail.

The market map evaluates vendors across two axes: agentic screening capabilities, and governance and reliability. Together, these axes show both the work each agent performs and the controls supporting its use in a regulated environment.

Explore the map, then use the criteria below to compare any agentic AI screening vendor on the same terms.

Agentic Screening: Mapping the AML/KYC AI Agent Market

Agentic Screening: Mapping the AML/KYC AI Agent Market

Screening is the universal control layer in financial crime compliance. Every bank, credit union and fintech is required to screen their customers and payments. Sanctions screening carries strict liability with fines in the billions. PEP and adverse media misses are increasingly attracting regulatory attention and fines in the hundreds of millions.

Agents offer unparalleled ability to improve screening speed and reduce costs, but without accuracy and validated models any reduction in alerts comes with high regulatory risk. For this reason, when compiling this map we focused on two axes: agentic screening capabilities, and governance and reliability.

Quadrant chart of 13 AI agent vendors in AML/KYC screening, plotting agentic screening functionality against reliability.

Methodology

Every vendor is scored 0–100 against eight criteria. Those eight scores are combined into the two axes: one measuring what the agents actually do in production, the other measuring whether their decisions can be governed, validated and defended.

The eight criteria, and what goes into each

1 · Screening data provenance & controlDoes the vendor own the sanctions, PEP, adverse media and watchlist data the decision rests on, or license it from someone else? How often is it refreshed? Do they own the matching engine, or only the workflow on top of it?
2 · Screening domain coverageHow many of the screening domains are covered: sanctions, PEPs, adverse media, other watchlists, and ongoing re-screening of existing customers.
3 · Screening adjudication depthLevel 1 alert clearing, Level 2 decisioning, and enhanced due diligence investigations, weighted by whether each is generally available or still in development.
4 · Independent validation & MRMIs there a named third-party validator? Is the work aligned to a recognized framework such as SR 11-7, the NIST AI Risk Management Framework or ISO/IEC 42001? Model risk management is the institution’s obligation, not the vendor’s, so this also asks whether the vendor supplies validation evidence and testing documentation the institution can actually use in its own review, or only self-attests.
5 · Policy & procedure training of agentsAre decisions tied to the institution’s own written policies and procedures, or to a generic model? How specific is the rationale attached to each match, and is the reasoning chain retained for review?
6 · Fits existing screening stackCan it deploy into the screening engine and case management the institution already runs, without replacing them? Integration mode, and documented time to go live.
7 · Regulated-FI screening production proofNamed customers and their regulatory tier, whether capabilities are generally available or on the roadmap, and how long the product has been running in production.
8 · Deterministic & reproducible outputDoes the same alert produce the same decision every time? Is there a deterministic retrieval and matching layer underneath, or is the entire chain probabilistic?

How the criteria are weighted

  • Agentic screening capabilities (vertical) leans hardest on data provenance and production proof, with coverage and adjudication depth carrying less. The reasoning: you cannot execute screening without the engine and data underneath it, and a capability only counts if it is running somewhere real.
  • Governance and reliability (horizontal) leans hardest on independent validation and MRM, then policy and procedure training, then determinism. The reasoning: an agent’s decision is only useful if a regulator and an internal model risk function will accept it.

Quadrant boundaries sit at 60 on both axes.

On the inputs. All scores and the evidence behind them are drawn from publicly available information: vendor websites and product documentation, funding announcements, press coverage and independent analyst rankings. Castellum.AI’s own figures are confirmed internally. Where a capability could not be confirmed from a public source it is recorded as not disclosed, rather than as absent, because those are different claims. Determinism is inferred from architecture and governance evidence, since no vendor in this cohort publishes reproducibility test results.
Market map v1-1 · scores as of 3 August 2026

Top AI Vendors for AML/KYC Screening

Castellum.AI

Trusted

Castellum.AI's Arbiter agents run on top of a proprietary screening engine and data operation the company owns outright, rather than a workflow layered on someone else's list. Screening data draws on more than 200,000 sources with a five-minute refresh, and the matching engine sits inside Castellum's own architecture. Agents ingest alerts, apply an institution's own written policies and procedures, and produce a documented disposition with reasoning built to hold up in an examination.

Castellum.AI supplies independent third-party validation results and continuous testing documentation that institutions can use directly as evidence in their own model risk management review, addressing the governance question most agentic vendors leave to self-attestation. Clients include a top 4 US bank, a top 5 global payments company, Lead Bank and Persona.

  • Owns the underlying sanctions, PEP, adverse media and watchlist data, refreshed every five minutes, rather than licensing it from a third party.

  • Owns the matching and screening engine agents decision against, not just a workflow layer on top of someone else's engine.

  • Agents are trained on each institution's own policies and procedures and stay aligned as those policies change.

  • Supplies independent third-party validation and continuous testing documentation institutions can use in their own MRM review.

Bretton AI

Trusted

Bretton AI builds AI agents for KYC/KYB reviews, AML and sanctions investigations, and ongoing transaction monitoring. For some deployments, Bretton also operates as a managed service, with its own analysts reviewing AI output before it reaches the client, rather than shipping software the client's team runs independently.

Bretton relies on partners including LexisNexis and Middesk for underlying risk data and business verification rather than owning that data itself.

Silent Eight

Trusted

Silent Eight's Iris platform provides AI-driven adjudication for sanctions, AML, fraud, and due diligence workflows, built on its own matching and investigation engine, though it licenses the underlying watchlist data. Model validation is run internally.

Sphinx

Trusted

Sphinx offers browser-native AI compliance agents that log into a bank's existing systems, including native connections to Verafin and Jack Henry, rather than requiring a new integration. It also offers a managed-service version, Frontline, that sells cleared cases and filed SARs as an outcome rather than software alone.

WorkFusion

Specialists

WorkFusion sells pre-built AI Digital Workers that perform Level 1 and some Level 2 compliance tasks — sanctions and adverse media alert review, KYC, and transaction monitoring support. It also relies on integrations with screening and data providers.

User reviews describe agents requiring extensive training before go-live and an implementation process that typically requires external vendor support rather than self-service setup. Reviews also note limited integrations and functionality compared to some competitors.

Themis

Specialists

Owns proprietary sanctions, PEP and adverse media data with a six-hour refresh, the strongest domain coverage score scored, across a 30-plus module platform. Level 2 and Level 3 decisioning are human-delivered rather than agentic, no named independent validator was found, and the platform replaces an incumbent stack rather than layering onto it.

Themis provides AML and due-diligence software built on proprietary data drawn from regulators, law enforcement, and policy institutions, including detailed criminal-conviction records. Its AI Investigator product uses agents trained on that data to automate parts of the investigation process. Themis primarily serves small business and corporations, not regulated financial institutions.

spektr

Innovators

spektr is a no-code compliance automation platform for KYB and KYC onboarding, risk monitoring, and case management, built around modular AI agents that conduct document review, network discovery and source-of-funds checks through integrations with third-party data providers.

Roe AI

Innovators

Roe AI runs as a browser agent that operates inside a customer's existing consoles and case systems rather than replacing them, producing an investigation file with evidence cited to its source and an audit trail exportable to a GRC system. Roe is positioned for investigation automation and evidence traceability on top of an existing stack; it does not provide underlying screening data, screening, or ongoing monitoring natively.

Diligent AI

Newcomers

Diligent AI builds AI agents for KYC/AML screening alert review, merchant risk investigation, and customer onboarding workflows. Rather than offering a platform, Diligent AI integrates with existing AML/KYC screening or case management software.

Arva AI

Newcomers

Arva AI provides AI agents for AML, KYB, and screening alert review. The company has an independent validation partnership with FairPlay, an AI assurance firm, covering AML and KYB use cases specifically.

Axle

Newcomers

Axle offers named AI agents for onboarding, transaction screening, and SAR narrative drafting, integrating with a customer's existing alert sources rather than supplying its own underlying data.

Variance

Newcomers

Variance provides AI investigative agents for fraud, KYC, KYB, AML and transaction-monitoring workflows that integrate with an SOP enforcement layer driven by client’s internal documentation. Variance integrates with external data sources to source risk signals for its AI agents.

Tangos AI

Newcomers

Tangos AI, founded in 2025, is a recent entrant building autonomous agents for financial-crime investigations. Its agents operate after an alert has already been generated and escalated internally, conducting evidence-gathering and producing case files for the investigation stages rather than initial detection and alerting phases.

Buyers Guide

How to Evaluate an AI Agent for AML/KYC Screening

Whichever AI agents vendor you’re evaluating to improve your AML/KYC workflows, ask about capabiltieis and governance separately. A strong demo may answer the first question, but only public evidence and documentation answers the second.

Capabilities: What the Agents Actually Do

Does the vendor own the underlying data, or license it?
Vendors relying only on customer-provided context and third-party data take longer to integrate and require ongoing validation to trust. A vendor that sources its own risk data directly from issuing authorities and pairs it with a pre-trained agent can trace every decision back to a controlled input. Ask how often the data refreshes and what happens when it's incomplete or conflicting.

How many screening domains does it actually cover?
Sanctions, PEPs, adverse media, other watchlists, and ongoing re-screening are not the same capability. A vendor strong in one is often weak or silent in another. Evaluate coverage and agent effectiveness during proof of concept testing domain-by-domain, not as a single yes/no.

Does it fit your stack, or does it replace it?
Ask for the integration timeline into your specific screening engine and case management system, and whether that timeline has been hit for a reference customer at your size.

How does the vendor define agentic?
Do the vendor’s agents incorporate a harness around an LLM, or is it a wrapper? How does the archetecture ensure recall and repeatability of decisions? Having structured guardrails around an agent that ensures decisions always follow your policies and procedures is essential for being explainable and auditable.

Governance: Where Decisions Hold Up

Is there a third-party validator, or does the vendor self-attest?
Self-attestation is not independent validation. Ask for the framework they tested against (SR 11-7, NIST AI RMF, ISO/IEC 42001), and what resources are provided to facilitate your own internal validation. Remember model risk management is your institution's obligation, not the vendor's: the right answer is "here's testing documentation you can put in your own MRM file," not "we handle that for you."

What does a decision journal actually look like?
Ask to see one. It should show the input data, the tools and datapoints the agent consulted, the reasoning chain, and a confidence score or rationale, in language a reviewer can read without pulling in an engineer. If the vendor can only show you a disposition code and a timestamp, that's not an audit trail.

How does it handle edge cases instead of just escalating everything?
A system that escalates every deviation isn't reasoning, it's routing. Ask how it distinguishes an acceptable deviation from a real anomaly, and test it against your hardest cases: fuzzy name and attribute alerts, transliterated PEP aliases, renamed sanctioned entities. Also ask about how the vendor tunes the escalation thresholds to align with your internal policies and procedures before going live.

How are decisions retained and reused across time?
If the same subject triggers another alert later, does the system draw on memory of the prior adjudication, or start fresh? Reliance on prior adjudications without a check for changed risk is a control gap worth surfacing.

How is model drift monitored and controlled?
Ask how analyst feedback (tags, overrides, corrections) flows back into the model, whether that process is version-controlled and reversible, and whether they can roll back to a prior model state if performance degrades. "The model keeps learning" is not itself an assurance without a rollback and audit trail behind it.

Is the same input guaranteed to produce the same decision?
Ask whether there's a deterministic layer underneath the reasoning, or whether the full chain is probabilistic end to end. No vendor in the current market publishes reproducibility benchmarks, so this is usually answered from architecture, not a published number. Request architectural documentation rather than marketing claims around "consistency" or “recall.”

For more, download our Guide to Agentic Alert Resolution.

How the map is built

Methodology

Every vendor is scored 0–100 against eight criteria. Those scores combine into two axes: one measuring what the agents actually do in production, the other measuring whether their decisions can be governed, validated and defended. Scope is screening only: sanctions, PEP, adverse media and other watchlist screening, adjudication of those alerts, ongoing re-screening, and the data and governance underneath them. Transaction monitoring, fraud detection, SAR/STR filing, case management as a product, credit and underwriting, and disputes are out of scope.

On the inputs. All scores and the evidence behind them are drawn from publicly available information: vendor websites and product documentation, funding announcements, press coverage and independent analyst rankings. Where a capability could not be confirmed from a public source it is recorded as not confirmed in public sources, rather than as absent.

  1. Screening data provenance & control: Owned vs. licensed data, refresh latency, and whether the vendor owns the matching engine.

  2. Screening domain coverage: Sanctions, PEPs, adverse media, other watchlists, and ongoing re-screening.

  3. Screening adjudication depth: L1, L2 and EDD, weighted by whether each is generally available or still in development.

  4. Independent validation & MRM: A named third-party validator, alignment to SR 11-7 / NIST AI RMF / ISO 42001, and whether the vendor supplies evidence the institution can use in its own MRM review.

  5. Policy & procedure training: Decisions tied to the institution's own written policies, match rationale specificity, and reasoning-chain retention.

  6. Fits existing screening stack: Deploys into an incumbent engine and case system without replacing it. Scored and shown per vendor, but not currently weighted into either axis.

  7. Regulated-FI production proof: Named customers and regulatory tier, available vs. roadmap, and production tenure.

  8. Deterministic & reproducible output: Same input, same decision. Inferred from architecture and governance evidence, since no vendor in this cohort publishes reproducibility test results.

Agentic screening capabilities (vertical): 40% data provenance, 30% production proof, 15% domain coverage, 15% adjudication depth — you cannot execute screening without the data and engine underneath it, and a capability only counts if it is running somewhere real.
Governance and reliability (horizontal): 50% independent validation & MRM, 30% policy training, 20% determinism — an agent's decision is only useful if a regulator and an internal model risk function will accept it. Quadrant boundaries sit at 60 on both axes.

FAQ

Common Questions

  • Screening only: sanctions, PEP, adverse media and other watchlist screening, adjudication of those alerts (L1, L2, EDD), ongoing and perpetual re-screening, and the data and governance underneath them. Transaction monitoring, fraud detection, SAR/STR filing, case management as a standalone product, credit and underwriting, and disputes are out of scope and not scored.

  • Each company’s core product is an AI agent, or agents, for financial crime compliance. Data providers, such as Moody’s or Dow Jones, are excluded, as are other traditional platforms or case management systems where AI agents are not the primary solution on offer.

  • Scores are based on analyst judgment against publicly available evidence: Vendor websites and documentation, funding announcements, press coverage and independent analyst reviews.

  • The market map is current as of August 2026.

  • You are welcome to use the market map only after you securing explicit written permission from Castellum.AI. Reach out at contact@castellum.ai to request permission. Use without written permission and public citations is a violation of our terms of service.

Ready to get started?

Schedule a demo to see AI agents in action.

“Castellum.AI is one of the most nimble vendors I’ve ever worked with, and they care about your ideas. Unlike other providers where you submit a ticket and wait weeks for a response, the team is always readily available for assistance and technical support. They put out a great product that allows us to truly own the risk and the process.”
— Daniel Schneider, Director of Financial Crimes, BSA Officer, SVP, Lead Bank