In KYC, fraud and dispute operations, AI agents assemble the case: they gather the transactions and documents, check them against scheme rules and the bank’s policy, triage alerts, and draft the chargeback file, investigation note, KYC refresh request or suspicious transaction report. Investigators and officers still decide liability, freeze or credit accounts, clear sanctions hits and file reports. The result is shorter queues and more consistent files, with every step on the record.
Why these queues suit agents
Fraud, disputes and KYC share three traits. The volume is high. The steps are well defined, by card scheme rules, regulator directions and the bank’s own procedures. And most of the effort goes into collecting and organising evidence before anyone can decide. That is exactly the shape of work where agents help most, and where their output is easy to check against a known standard.
They also share a risk: a wrong action is costly and sometimes hard to reverse. So the design principle is the same across all three. Agents prepare, check and propose. People decide.
Card disputes and chargebacks
A dispute file needs the transaction, the customer’s claim, the merchant’s response where there is one, the scheme’s reason code and its deadlines. The Card Dispute and Chargeback Preparer assembles that file and drafts the submission. The Fraud and Disputes Case Orchestrator keeps every case moving, routes each part to the right specialist, and makes sure deadlines are tracked rather than remembered.
Fraud claims and mule accounts
- Fraud Claim Investigation Copilot builds the investigation: what happened, in what order, which signals point which way, and what is still missing.
- Mule Account Alert Triage works through mule-account alerts, groups related accounts and ranks them for an investigator.
- Fraud Classification File Preparer prepares the file a bank needs when it classifies an account as fraud and reports it.
The investigator decides liability, and whether to freeze, credit or close an account. The agent’s job is to make that decision quicker and better documented.
KYC and onboarding
Tuvel’s KYC and onboarding team runs customer due diligence for individuals and businesses, from the first document to the KYC officer’s decision:
- KYC Onboarding Orchestrator owns each case: it picks the identity route and onboarding mode the market’s rules allow, assigns the specialists and tracks the clocks.
- Individual KYC Verifier and Entity KYB Verifier check each customer against the rules and the document list for its type, and list the gaps.
- Beneficial Ownership Tracer walks the ownership and control tree to the natural persons behind a business, and shows the arithmetic.
- KYC Document and Evidence Checker records which source verified which field, with its reference and time, and flags signs of tampering.
- Customer Risk and EDD Assessor proposes the initial risk tier by the bank’s method and applies enhanced due diligence triggers.
- CKYC and Customer ID Steward tracks registry uploads as they fall due, and duplicate customer IDs, and prepares each record for a maker.
Each bank sets its own KYC rulebook in its knowledge base, and the regulation is the floor. Market rules cover India, the United States, the United Kingdom, the United Arab Emirates, Saudi Arabia and Singapore. Documents are verified at their source through the bank’s connectors, so a file shows where each field was confirmed.
Periodic KYC refresh is a large, steady workload. The Periodic KYC Refresh agent identifies which customers are due, checks what the bank already holds, and drafts the request for what is missing. For business clients of the transaction bank, Corporate Onboarding and KYB takes the application and mandate and sends the KYB work to the KYC team. A person accepts every customer and sets every risk tier.
AML alerts, STRs and sanctions
From the compliance desk, three agents join the work:
- AML Alert Triage and STR Drafter reviews transaction-monitoring alerts, sets out the reasoning for each, and drafts a suspicious transaction report where the facts point that way.
- Sanctions Alert Reviewer compares each screening hit against the customer’s data and prepares the case for clearing or escalation.
- Dynamic Customer Risk Re-scoring recomputes a customer’s AML and KYC risk rating when an event arrives, such as an alert pattern, adverse news or a dormant account waking up, and proposes the new rating.
Closing an alert, filing a report, clearing a sanctions hit and changing a risk rating stay with people.
The decisions that stay with people
On Tuvel’s fraud and disputes desk and its KYC and onboarding desk, these always wait for a named person, approved on a card in the desk’s room:
- Crediting, freezing or closing an account.
- Deciding liability on a fraud claim.
- Classifying an account as fraud, and reporting it.
- Closing an alert or filing a suspicious transaction report.
- Clearing a sanctions hit.
- Accepting a customer, setting a risk tier or changing a risk rating.
What to check before you deploy
- Evidence with sources. Every claim in a file should point to the transaction, document or rule it came from.
- Untrusted content handled as data. Customer messages and merchant documents can contain instructions. Agents should treat them as evidence, never as commands.
- Scoped access. Each agent sees only the customer details and systems it was granted. Personal identifiers are masked before text reaches a model.
- Approvals that cannot be faked. Only the button on the approval card approves. Typing “approve” does nothing, and high-risk requests expire after 24 hours.
- A replayable record. Every turn, tool call and decision is recorded, so an examiner can see how a case was handled.
- A kill switch. Any agent can be paused in one click.
We cover these controls in full in how banks govern AI agents.
How it connects
Agents connect to your card management, case management, transaction monitoring, screening and KYC systems, and to identity, tax and registry sources for verification, through OpenAPI or MCP, granted to one agent at a time. Those connectors are built for each bank in its pilot Per deployment. See integrations.
The market
Specialists such as Casap (disputes and first-party fraud), Bretton AI (AML and KYC investigations) and Gradient Labs (customer operations, disputes and KYB) go deep on single queues. Platform vendors such as FIS and Temenos have announced agents for financial crime and sanctions alerts. Our buyer’s list describes each. Tuvel’s approach is a full desk of agents under the same control plane as the bank’s lending, compliance and service agents, so one set of levels, approvals and records covers them all.
A first pilot
A disputes queue is a natural place to start: clear rules, clear deadlines and clear measures. Run the agents on recent cases beside your team, compare files and turnaround, and review the record every week. See how a pilot works.