Nov. 18, 2026
26 minutes read
Share this article
A compliance analyst at a mid-size bank spends a normal shift clearing transaction monitoring alerts, most of which turn out to be false positives. She pulls customer records from one system, transaction history from another, adverse media hits from a third, and writes up a narrative by hand before closing each case. Multiply that by a team of dozens and a queue that never empties, and the problem becomes clear: KYC and AML compliance is not short on rules. It is short on capacity to apply those rules consistently, at scale, with a defensible record of why each decision was made. That capacity gap is what makes agentic AI for KYC and AML worth taking seriously — and it is also where the pitch gets oversold.
Agentic AI is being pitched as the fix, and that pitch isn’t unreasonable. Anti-money laundering and know-your-customer work is exactly the kind of multistep, document-heavy, cross-system process autonomous agents are built to handle: pull data, apply judgment within defined limits, take an action, hand off what needs a human, leave a record behind. The scale of the problem is not in dispute either. The United Nations Office on Drugs and Crime puts the volume laundered globally at 2 to 5% of GDP a year, an order-of-magnitude estimate rather than a measured total. UNODC pairs that percentage with a dollar range, though the range is pegged to an older GDP base and reads best as illustrative; on the supply side, Gartner projects that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024. Neither figure says anything about banking compliance specifically. The narrower claim is the more useful one: this is repetitive, document-heavy work that is expensive to staff and getting more expensive, which is the profile that pulls agentic AI investment in first.
The gap is not enthusiasm. It is design discipline. Deploying an autonomous agent inside a regulated financial crime workflow without clear permissions, audit trails, and a defined human checkpoint is not innovation; it is a new source of regulatory exposure. This guide lays out where agentic AI genuinely fits inside KYC and AML workflows, where it does not belong, and what a defensible implementation looks like for a bank or fintech evaluating this shift in 2026.
Compliance leaders are also evaluating this shift against a cost tension that comes up repeatedly in Coderio’s own conversations with financial institutions: alert volume tends to grow as transaction velocity increases and regulators expand reporting expectations, while compliance headcount budgets do not always scale at the same pace. That is a pattern we observe in client engagements, not a citation to an industry-wide budget survey, and it is worth reading as directional rather than as a measured statistic. This is part of why agentic AI is landing differently in compliance than it has in, say, marketing or customer service: the case for it rests less on novelty and more on a queue that keeps growing faster than teams can staff to clear it.
The word agentic gets applied loosely, so it is worth being precise before going further. A rules engine flags a transaction because it crossed a static threshold. Robotic process automation clicks through a screen the same way every time, following a fixed script with no judgment. A generative AI copilot drafts a case narrative when a human asks for one, but it does not decide anything on its own. An agentic system is different because it can plan a sequence of steps toward a goal, call tools and APIs to gather what it needs, and adjust when the situation deviates from the expected path, all without a human prompting every step. Whether it is permitted to make a bounded decision on its own, versus surfacing a recommendation for a person to act on, is a governance choice each institution must make deliberately for each workflow, not a property built into the technology.
In a compliance context, that means an agent assigned to triage an alert can pull the customer’s KYC file, check transaction history against typical account behavior, screen counterparties against sanctions and adverse media lists, and apply the bank’s documented risk logic. Whether it is then authorized to close the alert as a false positive on its own, versus only ever escalating it with a drafted narrative or a request for additional information attached, is a specific, controlled implementation choice each institution configures, not a general compliance norm that applies by default. A rules engine cannot do any of this because it has no reasoning step. RPA cannot do it because it breaks the moment a screen layout changes. A copilot cannot do it because it waits for a person to ask. These four approaches are often grouped together as automation, but they behave very differently the moment a case stops matching the expected pattern.
| Approach | How it operates | Where it fits in AML/KYC | Key limitation |
|---|---|---|---|
| Rules engine | Fires on static thresholds and conditions | Initial alert generation | High false-positive rate, no context |
| RPA | Scripts a fixed sequence of screen or system actions | Data entry, record retrieval | Brittle, breaks on UI or system change |
| Generative AI copilot | Drafts text or answers on request | Narrative drafting, research summaries | Reactive, no autonomous action |
| Agentic AI | Plans steps, calls tools, decides within limits, adapts | Alert triage, EDD gathering, screening | Requires governance, permissioning, audit design |
Coderio’s broader look at how agentic AI reshapes business functions covers this distinction in more depth across departments beyond compliance, and the same core traits apply here: autonomy, goal orientation, context handling, adaptation, orchestration, and learning from feedback.
Not every business process benefits equally from autonomous agents. KYC and AML do, for specific structural reasons. First, the work is inherently multistep: onboarding a customer means verifying identity, screening against watchlists, assessing risk, and setting a monitoring profile, in that order, across several systems. Second, it is data federation heavy: a single alert review might touch a core banking platform, a sanctions screening tool, a case management system, and external adverse media feeds. Third, most of the volume is repetitive judgment applied to slightly different facts, which is precisely the pattern where an agent with clear decision boundaries adds the most value without displacing the harder judgment calls that still need a human.
Fourth, and this is often underweighted, the work is exception-driven. Most alerts are not suspicious; they are noise generated by legitimate but unusual behavior. Reviewing Coderio’s work on hyperautomation in banking shows how RPA and rules-based automation already reduced manual effort in areas like loan decisioning and back-office reconciliation. Agentic AI extends that gain specifically to the judgment layer that RPA could never reach, because triaging an alert requires weighing context, not just executing steps.
Fifth, KYC and AML workflows already produce a paper trail by regulatory requirement, which is unusual among business processes and a significant advantage for agentic deployment. Every KYC decision, alert disposition, and SAR filing is already expected to be documented and defensible to an examiner. That existing documentation discipline gives compliance teams a head start on the audit trail an agentic system needs, compared to a department building explainability requirements from scratch. The institutions that struggle are usually the ones that treat the agent’s output as a replacement for that documentation discipline rather than an extension of it.
Most banks will not build agent infrastructure from scratch. Established AML and fraud platforms such as ComplyAdvantage and Napier AI already sit deep in the transaction monitoring and screening stack. ComplyAdvantage has begun formally marketing agentic workflows as part of its platform, and other vendors in the category, including Napier AI, describe their own AI-driven capabilities. The scope and maturity of those capabilities vary considerably, and each should be confirmed against that vendor’s current product documentation rather than assumed to be equivalent. That shift matters for how a bank should think about build versus buy. A platform vendor’s agentic layer typically comes pre-integrated with its own screening and monitoring engine, which shortens time to value but limits how far the agent can reach into systems outside that vendor’s ecosystem. A custom-built orchestration layer, connected through standardized tool interfaces, can span a wider set of systems, core banking, case management, external data feeds, at the cost of more upfront integration work.
Neither path is universally correct. A bank with a single, modern core system and a narrow compliance stack may get more value from a vendor’s built-in agentic features. A bank running a mix of legacy and modern systems, which describes a large share of mid-size and regional institutions, usually needs a custom orchestration layer to get agents to see the full picture, because no single vendor tool has access to every system a compliance decision actually depends on. Evaluating that fit honestly, before committing budget to either path, is one of the more consequential decisions in this kind of program.
The most useful way to evaluate agentic AI for KYC and AML is stage by stage, because the right level of autonomy is different at each point in the customer lifecycle.
During onboarding, an agent can orchestrate document capture, run identity verification checks, cross-reference submitted information against public records and sanctions lists, and calculate an initial risk score, presenting a complete file to a human reviewer rather than a partial one. This does not replace the analyst’s decision on borderline cases. It removes the manual data-gathering that used to precede that decision, which in the onboarding workflows Coderio has mapped with clients is consistently the largest single block of analyst time in that step.
Traditional KYC review runs on a fixed calendar, often annually for higher-risk customers, so material changes in a customer’s risk profile can go undetected for months. An agent monitoring for trigger events, an ownership change, a new adverse media hit, a shift in transaction pattern, can initiate a targeted review the moment a trigger fires instead of waiting for the scheduled cycle. This is often called perpetual KYC, and it depends on an agent that can watch multiple data sources continuously and decide when a review is actually warranted, rather than flooding the compliance team with noise.
This is the highest-volume use case and offers the clearest return. An agent assigned to alert triage gathers the same context a human analyst would gather and applies the bank’s documented risk criteria. Depending on how the institution has configured its authority, it can either resolve straightforward false positives with a logged rationale for later audit or escalate genuinely ambiguous cases with a pre-built case file attached. The goal is not to remove the analyst from the loop on real risk. It is to stop analysts from spending most of their day on alerts that a consistent, auditable process, operating within the authority it has been given, could resolve in seconds.
Sanctions screening already relies heavily on automated matching against lists such as OFAC’s Specially Designated Nationals list, but name-matching algorithms generate substantial false positives on common names and transliterations. An agent can apply secondary checks, cross-referencing date of birth, address, and transaction context, to reduce the volume that reaches a human screener, while never being the final word on a true sanctions match, which stays a human and typically a compliance officer decision by design.
Suspicious Activity Report and Suspicious Transaction Report investigations are research-intensive: pulling transaction history, related-party data, prior alerts, and open-source information, then synthesizing it into a coherent narrative. An agent can assemble that research package and draft a narrative for review, removing most of the manual retrieval work that precedes the write-up, while the decision to file, and the filing itself, remains with authorized and accountable compliance personnel at the institution. Under the Bank Secrecy Act, whose implementing regulations require financial institutions themselves to report suspicious activity, the filing obligation sits with the institution and not with any tool it uses to support that decision, so an agent’s role here is limited to preparation and drafting rather than the filing decision.
Agents can also maintain exam-ready documentation continuously, compiling the rationale, data sources, and human sign-offs behind each automated decision into a format examiners expect, rather than reconstructing that trail under pressure during an actual exam window. Across the six stages, a consistent pattern holds: the agent’s role expands with volume, while the human checkpoint stays anchored to legally weighted decisions.
| Workflow stage | Typical agent role | Human checkpoint |
|---|---|---|
| Onboarding and KYC | Assemble identity, risk, and screening data into one file | Approve or decline the account |
| Perpetual KYC | Detect trigger events and initiate targeted review | Confirm scope and outcome of the review |
| Transaction monitoring | Triage alerts, resolve clear false positives, draft escalations | Review and decide on escalated cases |
| Sanctions screening | Apply secondary matching to reduce false positives | Confirm or reject any true match |
| SAR/STR investigation | Assemble evidence and draft narrative | Approve and file the report |
| Regulatory reporting | Maintain continuous, exam-ready documentation | Sign off on submissions |
A single monolithic agent handling an entire KYC or AML workflow end to end is the wrong design for a regulated environment, because it collapses distinct decisions into one opaque process. The more defensible architecture uses an orchestrator agent that manages the workflow and hands specific, bounded tasks to specialist agents: one scoped to identity verification, one to sanctions screening, one to transaction pattern analysis, one to narrative drafting. Each specialist agent has a narrow tool scope, meaning it can only call the systems and data sources relevant to its task, which limits the blast radius if something goes wrong and makes the audit trail easier to reconstruct.
This pattern mirrors what Coderio has documented in its work on connecting AI systems through the Model Context Protocol: standardized, permissioned tool connections rather than ad hoc integrations, so every action an agent takes is logged, scoped, and traceable back to a specific tool call. Building this kind of architecture is a core reason compliance-focused engagements increasingly sit alongside core banking modernization work, because agents are only as reliable as the systems and data pipelines they connect to. A well-designed AI orchestration stack is what keeps six specialist agents behaving as one coherent, governed workflow instead of six unmonitored processes acting independently.
Financial institutions already operate under model risk management frameworks that require validation, monitoring, and documented limitations for any analytical model used in a business decision. That framework was refreshed in April 2026: SR 26-2, issued jointly by the Federal Reserve, OCC and FDIC, supersedes both the long-standing SR 11-7 guidance and SR 21-8, the interagency statement on model risk management for bank systems supporting BSA/AML compliance. The OCC rescinded its companion Bulletin 2011-12 and replaced it with Bulletin 2026-13. Autonomous agents fall inside that same expectation, even when the underlying technology is new. The governance principles align with the NIST AI Risk Management Framework, which is voluntary and non-prescriptive rather than binding, and organizes AI risk work around four functions: govern, map, measure, and manage. The govern function, often skipped in summaries of the framework, is most directly relevant to agent accountability. This aligns with the Financial Action Task Force’s 2021 report on the opportunities and challenges of new technologies for AML/CFT, which is a landscape study rather than a binding standard.
For institutions operating under other regimes, the UK’s Financial Conduct Authority has set out its approach to AI in financial services, relying on existing frameworks such as the Consumer Duty and the Senior Managers and Certification Regime rather than a new AI-specific rulebook. European institutions face a materially different picture. The EU AI Act is binding, and its timeline moved in 2026: under the AI Omnibus, in force since 27 July 2026, the obligations for stand-alone high-risk systems listed in Annex III — the tier covering creditworthiness assessment — now apply from 2 December 2027 rather than August 2026. The obligations themselves are unchanged in substance: risk management, data governance, record-keeping, human oversight, and transparency. The deferral buys time, not exemption, so in-scope institutions should still map their agentic compliance use cases against the Act’s risk tiers before deployment rather than assume general supervisory principles are sufficient.
Beyond the EU AI Act’s horizontal obligations, no supervisor has published AML-specific rules for agentic AI as of this writing, which is precisely why the general model-governance principles above, rather than a single AI-specific rulebook, are the safest baseline to build against. Six controls do most of the work in making an agentic deployment defensible, and each one answers a question an examiner is likely to ask directly. The technical mechanics behind them — permission models, tool scoping, policy-as-code — are covered in Coderio’s guide to agent guardrails; what matters here is how each one changes when the decision being governed is a regulated one.
| Control | Why it matters | Practical anchor |
|---|---|---|
| Documented tool scope per agent | Limits what each agent can access or execute | Permission map filed with the model inventory, to SR 11-7 documentation expectations |
| Full audit trail per decision | Reconstructs the rationale behind every automated action | Immutable log of data pulled, logic applied, and outcome |
| Human checkpoint at accountable decisions | Keeps authorized, accountable personnel responsible for filings and closures | SAR filing, account closure, true sanctions match |
| Independent validation of agent logic | Confirms the agent behaves as documented, not just as intended | Validation by a function independent of the model owner, as model risk management already requires |
| Explainability of each escalation or resolution | Lets examiners and internal audit understand, not just trust, the outcome | Plain-language rationale attached to every case |
| Change management for prompts and tool access | Prevents undocumented drift in agent behavior over time | Configuration changes logged to the same standard as a change to a monitoring rule |
Institutions weighing this work often start with a broader security audit or a formal compliance testing engagement before extending automation into judgment-bearing workflows, precisely because governance gaps are easier and cheaper to close before an agent is live than after an examiner finds one.
A meaningful difference exists between an agent that gathers evidence and drafts a recommendation and an agent that makes the final call on a regulated action. Filing a SAR, closing an account for suspected financial crime, and confirming a true sanctions match are decisions that carry legal and reputational consequences that should sit with an accountable human, not a system, regardless of how strong its underlying model is. The same logic applies to enhanced due diligence sign-off on high-risk customers and to any decision that could be construed as denying a customer access to financial services. This is what compliance and risk teams generally mean by human-in-the-loop design: the agent prepares the ground, but a named person makes the call and owns the outcome. Coderio’s work on governance frameworks for agentic systems in production sets out how those decision gates get designed and enforced outside a regulated context; inside one, they stop being a design preference and become an accountability requirement.
This is not a limitation to design around. It is the design. The value of agentic AI in this domain comes from removing the manual research and data-gathering burden that precedes a decision, not from removing the decision itself. Institutions that blur that line are not automating compliance; they are outsourcing accountability to a system that cannot hold it, which is precisely the kind of gap a zero trust approach to security and least-privilege access design are meant to prevent at the infrastructure level, and the same principle should govern decision rights at the workflow level.
A useful test for where the line sits: if a regulator would ask a named individual to explain why a specific action was taken, that action needs a named individual behind it, not just an agent’s output. Drafting a narrative passes that test because a human reviews and signs it before it goes anywhere. Filing the SAR does not pass that test if no accountable person actually made the filing decision. Applying that test consistently across every workflow stage is a more reliable guide than a fixed list of banned actions, because new use cases will keep appearing faster than any static list can anticipate them.
The realistic case for agentic AI in KYC and AML is operational, not magical. Analysts spend less time on manual data assembly and more time on the judgment calls that actually require a trained professional. Alert backlogs shrink because straightforward false positives are resolved consistently, rather than queued behind more urgent cases. Onboarding moves faster because a complete risk file is available at the point of decision rather than assembled piecemeal. The practical differences center on five operational measures, each comparing how the work is done manually today with how it is done with an agent in the loop.
| Metric | Traditional manual approach | Agentic-assisted approach |
|---|---|---|
| Alert research time | Analyst gathers data across systems manually | Agent assembles case file automatically |
| False-positive handling | Queued and reviewed individually | Consistently resolved with logged rationale |
| Onboarding risk file | Compiled across multiple manual steps | Assembled and scored before human review |
| Periodic review timing | Fixed annual or semiannual calendar | Triggered by detected risk events |
| Audit trail reconstruction | Manually rebuilt for exams | Continuously maintained and exam-ready |
These gains compound with the broader modernization work many banks are already pursuing. Coderio’s banking modernization studio and open banking studio both address the underlying system fragmentation that limits how much value an agent can add on top of legacy infrastructure, and our work with financial services clients, documented in this fintech partnership, reflects the same pattern: automation delivers the most value once the surrounding systems and data are ready for it.
Most agentic AI compliance pilots stall for predictable reasons. Data fragmentation is the most common: an agent asked to assemble a risk file cannot do so reliably if customer data lives in inconsistent formats across five disconnected systems. Unclear ownership is close behind: no single team is accountable for validating agent behavior, so issues surface only after they have already caused a problem. Internal reviews often find a missing audit trail, usually only after a pilot has run for months. Over-automation, extending autonomy into decisions that should stay human, tends to draw regulatory attention quickly and can set a program back further than a slower, more disciplined rollout would have. Vendor tools purchased without integration into the bank’s actual case management and data infrastructure are another common failure pattern; the tool works in a demo and stalls in production because it was never connected to the systems it needed to reach.
These failure patterns are not unique to compliance. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, attributing that to escalating costs, unclear business value, and inadequate risk controls. McKinsey’s 2026 State of AI survey shows the same tension from the other side: 40% of respondents at large organizations report scaling AI agents, up from 27% the year before, while the share attributing any EBIT impact to AI held flat at 37%. Deployment is accelerating and measured financial return is not, which is the gap this section is about. Both findings describe enterprise AI broadly rather than compliance specifically. The specific causes named above, weak process mapping, unclear ownership, and underestimated integration complexity, come from Coderio’s own client engagements rather than from either analyst, though they sit alongside the cost and control concerns those firms do identify.
Coderio’s broader review of AI security risks and our guide for business leaders evaluating AI investments cover several of these failure patterns in more general terms, and they apply with extra weight in a regulated compliance context where the cost of an ungoverned failure is measured in examiner findings, not just wasted budget.
A related pattern worth naming separately: teams that pilot an agent successfully in transaction monitoring sometimes assume the same design will transfer directly to sanctions screening or SAR drafting without adjustment. It usually does not, because each workflow stage carries a different risk profile and a different regulatory expectation for human involvement. What worked for resolving low-risk false positives in transaction monitoring can be the wrong level of autonomy for a sanctions match, where the cost of an incorrect automated decision is far higher. Treating each stage of the compliance lifecycle as its own governance decision, rather than assuming one architecture fits every stage equally, is what keeps an expanding program from outrunning its own controls. Banks that pair this work with ongoing application security testing tend to catch these gaps in staging rather than in production, where mistakes cost considerably more.
Yes, with the same governance obligations that apply to any analytical system used in a regulated decision. No rule prohibits using agentic AI in compliance workflows, but institutions remain accountable for outcomes, so documentation, validation, and human accountability at key decision points are not optional.
From a sound governance perspective, no institution should let it. The Bank Secrecy Act places the filing obligation on the institution itself, and an agent is not a legal filer. An agent can assemble the evidence and draft the narrative, but the decision to file, and the filing itself, should remain with authorized and accountable compliance personnel who can be held responsible for the judgment involved. This is a governance recommendation grounded in that accountability structure, not a claim that a specific rule names AI agents by name.
Most existing automation, rules engines, and RPA execute fixed logic without judgment. Agentic AI can plan a sequence of steps and gather context across systems, and, where an institution has deliberately configured and authorized it to do so, make a bounded decision within defined limits, which lets it handle the ambiguous cases that fixed automation cannot.
Extending autonomy into decisions that should remain human, such as final SAR filings or account closures, without a documented audit trail to support the agent’s reasoning. That combination is what turns an efficiency project into a regulatory finding.
Timelines vary widely based on data readiness, so treat any single number with caution. In Coderio’s engagements, a narrow, well-scoped pilot, such as alert triage for one transaction type, tends to reach measurable results in a matter of months rather than weeks. The deciding variable is almost always whether the underlying data is accessible and whether the audit trail was built in from the start rather than added later.
Not typically in the near term. Most institutions redirect analyst time from manual data-gathering to the judgment calls and investigations that still require a trained person, rather than reducing the team, since alert volume and regulatory expectations continue to grow.
Yes, proportionally. A smaller institution will not have the same alert volume or internal audit resources as a global bank, but the accountability standard for a SAR filing or a sanctions decision does not scale down with institution size. A community bank or a fintech partnering with a bank sponsor should expect the same documentation, audit trail, and human checkpoint requirements, sized to its own volume rather than skipped because the team is smaller.
Agentic AI is a genuine fit for KYC and AML compliance because the work is structurally suited to it: multistep, cross-system, judgment-heavy, and drowning in repetitive volume that a well-governed agent can absorb. The institutions that get real value from it are the ones that treat autonomy as something to be scoped deliberately, stage by stage, with a documented audit trail and a clear human checkpoint at every decision that carries legal weight. The ones that struggle usually skip that discipline in the rush to deploy. Getting the architecture, governance, and human escalation design right from the start is what separates a defensible compliance program from a new source of regulatory risk, and it is the difference between an agent that makes a compliance team faster and one that makes an examiner’s next visit considerably longer.
Institutions ready to move past the pilot stage can start with a focused scoping engagement to identify the strongest candidate KYC or AML workflow for a first, governed deployment. Coderio’s AI services team works with banks and fintechs on exactly this kind of architecture, from workflow mapping through pilot design and governance.
Leandro is a Subject Matter Expert in Backend at Coderio, where he evaluates and implements the language and framework choices behind modern AI and enterprise systems, rom Python-based ML pipelines to the C++ and Rust layers that handle production inference. He contributes technical thought leadership on backend architecture, AI-assisted modernization, and how engineering teams choose the right stack for scale, performance, and long-term maintainability.
Leandro is a Subject Matter Expert in Backend at Coderio, where he evaluates and implements the language and framework choices behind modern AI and enterprise systems, rom Python-based ML pipelines to the C++ and Rust layers that handle production inference. He contributes technical thought leadership on backend architecture, AI-assisted modernization, and how engineering teams choose the right stack for scale, performance, and long-term maintainability.
Accelerate your software development with our on-demand nearshore engineering teams.