Fraud teams sit at an awkward point in GDPR Article 22. An AI agent that blocks a card, files a fraud marker or recommends closing an account can significantly affect a genuine customer who did nothing wrong. Fraud analysts must be trained to review those decisions properly, under time pressure, without revealing the detection logic a fraudster could exploit.
What does Article 22 require of fraud decisions?
Article 22 gives individuals the right not to be subject to a decision based solely on automated processing, including profiling, with legal or similarly significant effects. The decision may be made that way if it is necessary for a contract, authorised by law with suitable safeguards, or based on explicit consent. Under contract or consent, the controller must at least allow human intervention, the person's point of view, and a way to contest.
Fraud work involves profiling by design: the agent compares behaviour with patterns and scores the risk. Most of its outputs are alerts for analysts, which have no effect on the customer until someone acts. The significant decisions are the ones that reach the customer:
- Blocking a card or account, especially when the customer is abroad or needs the funds.
- Filing a fraud marker to a shared industry database, which can affect the person's access to banking elsewhere.
- Exiting a customer because of suspected fraud.
- Refusing a refund for a disputed transaction on the grounds that the customer was complicit.
Some fraud decisions rest on legal duties to prevent financial crime, and the authorised-by-law exception may apply, with safeguards set by that law. Even then, firms should expect to show that the people involved made real decisions.
Who is in scope?
Banks, card issuers, payment firms, insurers, lenders, retailers and marketplaces running fraud agents on individuals' activity. The reviewers:
- Fraud alert analysts who work agent-triaged queues.
- Investigators handling cases referred for deeper review.
- Fraud marker decision makers who approve filings to shared databases.
- Customer-facing fraud teams who talk to blocked customers and lift blocks.
- Disputes staff deciding refunds where complicity is alleged.
What makes fraud review meaningful?
A firm's review is meaningful only if the analyst has authority to change the outcome, competence to judge, and the information needed; what meaningful human review is explains the test. Fraud adds three complications.
Volume and fatigue. Alert queues are long, and most alerts are false. Analysts who clear hundreds an hour stop reading. The agent may close obvious false positives at depth 3 if the firm has certified it to, which leaves analysts with fewer, harder cases. Fewer cases only help if analysts use the time.
Speed. A blocked card at a hotel desk needs resolving in minutes. Review for blocks often happens after the agent acts, through the customer's request. Customer-facing staff must have authority to lift the block once identity is confirmed, not just log a ticket.
Secrecy. Analysts cannot tell a suspected fraudster which signals tripped the alert. But a genuine customer contesting a block or marker is entitled to put their side and have it considered. Training must cover how to give a meaningful outcome and a fair hearing without disclosing detection logic.
Fraud markers deserve the most care. Filing one is held for a named decision maker, every time, because its effect travels beyond the firm. The decision maker must apply the standard of evidence the database requires and must know that a pattern match is not proof.
What must fraud analysts be trained to do?
- Separate the agent's pattern match from evidence of fraud.
- Weigh provenance: device and transaction data from the firm's own systems are sealed; a call from someone claiming to be the customer is declared; an emailed document is external.
- Recognise genuine customers caught by the pattern: travellers, people with new devices, older customers whose behaviour shifts.
- Lift blocks promptly when the evidence clears the customer.
- Apply the evidence standard for fraud markers, and refuse to file when it is not met.
- Hear contests and change the outcome when warranted.
- Sample agent-closed alerts for missed fraud and for wrongful closures.
The process side is set out in how to review card fraud alerts with an AI agent.
How do you meet it and show it?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Identify significant fraud decisions | Blocks, markers, exits and complicity refusals mapped and held where needed | Decision map in the Article 35 DPIA |
| Authority to change outcomes | Customer-facing staff can lift blocks; marker decisions by named staff | Role definitions; block-lift records |
| Analyst competence | Training and assessment on fraud cases before release rights | Assessment results per analyst |
| Review is real | Seeded cases in queues; review times monitored | Seeded case detection; review metrics |
| Right to contest | Contested blocks and markers re-decided by a trained person | Contest log with outcomes |
| Record | Each verdict filed with analyst identity and reason | Append-only decision record |
What will a supervisor or auditor ask?
- Which fraud decisions reach customers without a person deciding?
- Who files fraud markers, and what evidence standard do they apply?
- How quickly can a genuine customer get a block lifted?
- Show us agent-closed alerts that analysts sampled. What did they find?
- How do analysts handle a contest without revealing detection logic?
Auditors will test analysts, not just agents; how auditors test that AI agent reviewers are qualified explains how.
What does Article 22 not cover?
It does not cover alerts that never reach a customer. It does not set the evidence standard for shared fraud databases, which their own rules do. It does not override anti-money laundering obligations, including the bar on tipping off. And it does not replace PSD2 strong customer authentication, which governs how payments are authenticated in the first place. Compare the payment-time pressure in payments review and the complicity questions in insurance claims review. The guide to AI literacy and certification covers reviewer training generally.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Is a fraud alert itself an automated decision?
Not usually. An alert sent to an analyst has no effect on the customer. The decision is the block, marker or exit that follows.
Can we tell a customer why their card was blocked?
You can tell them it was blocked for security checks and what they need to do. You need not reveal the detection signals, but you must consider what they say.
Should agents ever file fraud markers?
An agent can draft the filing at depth 2. Filing should be held for a named, trained person every time, because the effect reaches beyond your firm.