A KYC evidence agent, walked through one case
A governed KYC agent gathers and compares evidence; it never decides. It reads what the applicant typed, what they uploaded and what registries hold, then lists every conflict that matters with both sources attached. An analyst resolves each exception and records why. A compliance officer approves, rejects or sends it back.
A commercial applicant whose ownership records disagree
An operating company applies for a commercial account. Its application lists one shareholder at 25 percent. The corporate registry shows 22 percent; the tax-authority extract shows 28. The bank screens beneficial owners at 25 percent, so the gap decides whether this person is screened at all.
- Agent prepares
- Control layer refuses
- Named person decides
- Person decides · Onboarding operator
The evidence run is started by a person
The applicant submits the form and documents. The run does not fire on submission: an authorised operator triggers it, so every run has someone who started it.
- Agent prepares · KYC evidence agent
It fetches the three evidence categories
What the applicant typed, what they uploaded (incorporation record, share register, group chart) and what independent sources hold (corporate registry, tax extract, sanctions and PEP lists, adverse media). Each source is read through a tool it has been granted, and nothing else.
- Agent prepares · KYC evidence agent
It lists conflicts instead of resolving them
Three exceptions: ownership 22 vs 28 percent, either side of the screening threshold; a director’s name spelled differently on the incorporation record and on a sanctions-list entry, with a different date of birth; and a subsidiary in the annual return that the applicant’s group chart leaves out. Both values stay on the record. Nothing is averaged or overwritten.
- Control layer · Control layer
Approval is out of the agent’s reach
The approve, reject and clear-alert actions are not tools this agent holds, so a request to use them fails and is logged. Approval stays unavailable to anyone while a required exception is open.
- Person decides · KYC analyst
The analyst works each exception
The tax extract is the more recent filing, so the analyst takes 28 percent and writes down why; the shareholder is now screened. The name match is checked against the date of birth and escalated rather than cleared. The missing subsidiary goes back to the applicant as a request for an updated group chart.
- Person decides · Compliance officer
The officer decides
The officer sees the reconciled evidence, the analyst’s reasoning on every exception and the risk rating, and approves, rejects with a reason code, or sends the case back. Here it waits for the updated group chart, and the comparison runs again when it arrives.
What the agent cannot do
These are not instructions in a prompt. The actions are not tools the agent holds, so a request to use one fails and the attempt is recorded. A record with refusals in it shows the boundary held.
- Approve onboarding
- Reject a customer
- Clear a screening alert
- Overwrite a value that another source disputes
The control answers for this agent
- Who owns it?
- The compliance officer, who can pause or stop it.
- What risk tier?
- Elevated: personal data, sanctions data and a regulated decision downstream.
- What can it reach?
- Corporate registry, tax extracts, sanctions and PEP lists, uploaded documents and the keyed intake. Nothing broader.
- Where is the hard line?
- No approval, rejection or alert clearance, enforced outside the prompt.
- Who decides?
- The analyst resolves exceptions with a written reason; the compliance officer approves, rejects or asks for more.
- What is kept?
- Ownership lineage, screening evidence, document comparisons, the analyst’s rationale, the officer’s decision and reason code, and every tool call.
From this walkthrough to your workflow
- 1
Workflow diagnostic
A short call comparing your current process with this one: where the evidence comes from, who decides, and where the time goes. No demo, no deck.
Book a diagnostic - 2
Workflow-to-production workshop
With your process owner and your risk or compliance reviewer in the room, map the agent, its boundaries and the evidence your reviewers will need.
Banking workflow workshop - 3
Paid production validation
One agent on your data, in your cloud, measured against a baseline and success measures agreed before work starts. Results are reported as measured, not modelled.
KYC evidence, asked directly
Can an AI agent approve KYC onboarding?
Not in a governed design. The agent gathers evidence, compares sources and surfaces exceptions. Approving a customer, rejecting one and clearing a screening alert turn on risk appetite and a reading of regulation, so they belong to a named person, and the approve action is kept out of the agent’s reach rather than discouraged in a prompt.
What does a KYC agent do when ownership records disagree?
It keeps both values, each with its source and retrieval time, and raises the conflict as an exception. An analyst decides which source to trust and records why, or asks for a certified share register. Averaging or quietly picking one is how a file ends up unable to answer an examiner.
What does the audit trail of a KYC agent need to contain?
Enough to reconstruct one case months later without asking the model: who triggered the run, the agent and prompt version, every tool call and source, the exceptions raised, any blocked requests, the analyst’s rationale and the officer’s decision with its reason code.
Is this a recording of a live customer case?
No. The case on this page is invented to show the shape of the workflow; the names, percentages and documents are illustrative. A walkthrough on your own process starts with a short workflow diagnostic.
Read the detail
- The full KYC evidence agent design, including the audit trail and what is not built yet
- Which banking workflows suit governed AI agents
- SR 26-2 explained, and why agentic AI sits outside its scope
- Moring for banks and financial institutions
Other walkthroughs: all four workflows.