Agentic AI in KYC is not a chatbot added to a compliance dashboard. It is software that can work towards a defined case objective, choose from authorised tools, complete a sequence of tasks, inspect the result and decide what to do next within controlled limits.

For a KYC case, that objective might be: establish who the applicant is, map the relevant ownership and control relationships, collect the required evidence, analyse screening results, identify unresolved risks and prepare the file for a human decision.

The agent does not need unrestricted autonomy to be useful. In regulated work, the strongest design is bounded agency: broad capability within narrow permissions, clear evidence for every action and mandatory escalation where judgement or accountability is required.

What is agentic AI in KYC?

An AI agent combines five elements:

1. A goal: the outcome it is working towards, such as a complete and review-ready KYC file.

2. Case state: a structured view of what is known, what is missing and what has already happened.

3. Tools: authorised ways to read documents, query records, run native screening, calculate ownership, create tasks or draft communications.

4. Policy and permissions: rules that determine what the agent may do, what evidence is acceptable and when it must stop.

5. Evaluation: checks that compare the current case with the required standard and choose the next action.

This loop is what makes the system agentic. It can move from one task to the next according to the state of the case, rather than waiting for a person to trigger every individual action.

The FCA's 2026 Supercharged Sandbox explicitly identifies agentic compliance agents as an area of interest. That does not remove the need for governance. It shows that the relevant question has moved from whether agents will enter financial services to how they can be tested and controlled.

Automation, generative AI and agentic AI compared

These terms are often used interchangeably, but they describe different behaviour.

All three can belong in the same workflow.

Rules should control deterministic requirements such as thresholds, permissions and approval gates. Generative models are useful where the input is unstructured, or the task requires interpretation. An agent coordinates the work, moving between AI-assisted analysis and deterministic controls.

The distinction matters because an agent is not simply a more powerful model. It is an operating design around a model.

The anatomy of a KYC agent

Consider a business investor onboarding into a fund.

The case arrives with an incorporation certificate, ownership chart, register of directors, identity documents and a completed questionnaire. The agent first classifies the applicant and loads the relevant policy. It reads the documents, extracts people and entities, then compares the ownership chart with the registry evidence. It identifies a holding company that has not been verified and creates the next task. Once the ownership chain is complete, it runs screening on the relevant subjects, analyses possible matches and assembles unresolved items for review.

Throughout the process, the agent updates the case state:

• Evidence received.

• Facts extracted.

• Sources consulted.

• Relationships established.

• Checks completed.

• Exceptions found.

• Actions taken.

• Human decisions required.

Without that structured state, an agent is merely producing isolated answers. With it, the agent can understand what remains to be done.

What can agentic AI do across a KYC case?

1. Open and organise the case

An agent can classify the applicant, identify the applicable workflow and create the initial case structure. It can distinguish a person from a company, trust, partnership or fund, then load the relevant evidence requirements and approval route.

This first classification controls everything that follows. A company cannot be treated as a person with extra documents. The practical distinction between KYB and KYC shows why entity verification depends on relationships, ownership and control.

Where the applicant type is ambiguous, the agent should ask for clarification or route the case to a reviewer. It should not silently choose the easiest workflow.

2. Read documents and identify gaps

An agent can classify documents, extract fields, compare values and check basic requirements:

• Is the document the expected type?

• Is it current?

• Are all required pages present?

• Do names, dates and identifiers agree across sources?

• Is a translation or certification required?

• Does the evidence support the stated fact?

The useful output is not a block of extracted text. It is a set of sourced facts plus explicit exceptions. A reviewer should be able to open any fact and see the document, page and extraction that support it.

3. Resolve entities and map ownership

Entity resolution is the work of deciding whether records from different sources describe the same person or organisation. An agent can compare names, identifiers, addresses, jurisdictions and dates, then propose matches or identify conflicts.

It can also turn unstructured ownership evidence into a relationship map, calculate holdings through several layers and identify the natural persons or control relationships that require review.

The agent should distinguish:

• A fact stated by the applicant.

• A fact extracted from a document.

• A fact obtained from an independent source.

• A calculated relationship.

• An inference that still needs confirmation.

That distinction prevents a plausible interpretation from being treated as verified evidence.

4. Gather and compare evidence

With permissioned access to approved sources, an agent can retrieve registry records, prior approved evidence and internal records. It can compare them with the current submission, highlight differences and determine which gaps remain.

Tool access must be narrow. The agent should know which sources are permitted for each task, how results may be used and what to do when a source is unavailable. It should record the query, timestamp and returned evidence rather than preserving only its own summary.

5. Analyse screening results

Native screening can produce sanctions, PEP and adverse-media results within the case. An agent can compare possible matches with the subject's secondary identifiers, retrieve relevant context and prepare a reasoned analysis.

For an apparent name match, it may compare date of birth, nationality, location, role and other identifiers. It can explain why the evidence points towards a false positive, a credible match or an unresolved result.

The final disposition should follow the firm's approval policy. Agentic analysis is valuable because it prepares the evidence. It should not hide uncertainty or bypass the controls described in how to reduce false positives without weakening AML screening.

6. Prepare outreach and follow-up

When evidence is missing, an agent can draft a targeted request that explains exactly what is needed and why. It can avoid asking for a document already held elsewhere in the approved record, schedule reminders and update the case when a response arrives.

The important change is contextual follow-up. Instead of sending a generic list, the workflow asks only for the unresolved evidence that applies to that applicant and relationship.

Permissions should determine when a message may be sent automatically and when a person must approve it. Higher-risk or sensitive communications may need review even if the underlying request is routine.

7. Assemble the risk file

Once the evidence and checks are complete, an agent can prepare a structured case summary that points back to sources. It is a navigation layer over the file, not a replacement for the file.

A human reviewer then accepts, changes or rejects the proposed analysis according to the firm's approval framework. That action, and the evidence available at the time, should become part of the audit history.

8. Monitor the relationship

After onboarding, an agent can continue to evaluate events against the relationship:

• A document expires.

• A company changes director or ownership.

• A new screening result appears.

• A risk-relevant data point changes.

• A periodic review becomes due.

It can assess which records and relationships are affected, gather the updated evidence and open a review when the policy requires one. This connects agentic work with the KYC periodic review process and event-driven monitoring.

The agent should not reopen every case for every change. Materiality rules and human-approved policy determine which events require action.

What an AI agent should not decide

An agent should not own:

• The firm's risk appetite.

• The definition of acceptable evidence.

• The policy for approving or rejecting a customer.

• A material exception to that policy.

• The final treatment of an ambiguous sanctions or PEP match.

• The decision to onboard, restrict or exit a relationship.

• Its own expansion into new tools, data or use cases.

These are accountable governance and compliance decisions.

Human oversight does not mean a person repeats every task. It means people define the policy, approve material decisions, review uncertainty, test performance and can intervene at any point.

Where Steward fits

Steward is an AI-first AML/KYC platform for investment services. Its AI agents can do document review, complex ownership analysis, screening, ongoing monitoring, and create an audit trail. Book a demo to see our agents in action.