What is tipping off?

Tipping off is disclosing to a person, directly or by implication, that a suspicion report has been made about them or that a money laundering investigation is under way or contemplated. Most AML regimes prohibit it, often as a criminal offence, because a forewarned suspect can move funds and destroy evidence. Customer-facing AI agents make accidental tipping off easier.

Why does tipping off matter once agents act?

Staff are trained not to tip off: what not to say, and when to pass a call on. An agent knows only what is in its brief and its sources. If it can see that an account is restricted pending investigation, and the customer asks why their payment has not arrived, a helpful agent may explain more than it should.

The risk often comes through implication rather than words. "Your account is under review by our financial crime team" is plain tipping off. "I can't process this because of a compliance flag, but it should clear once the investigation finishes" is no better. Our piece on tipping-off risk when agents talk to customers sets out the common routes.

How do firms prevent tipping off in practice?

The controls work at three levels:

1. What the agent can see. A customer-facing agent should not have access to suspicion reports, investigation notes or the reason a restriction was applied. It sees only that an account needs a person.
2. What the agent can say. Its brief includes approved, neutral wording for restricted accounts and delayed payments, and nothing else on the subject.
3. When it hands over. Any question about a restriction, a held payment or a request for documents linked to an investigation goes to a trained person, without explanation to the customer beyond the approved wording.

How to stop an AI agent from tipping off a customer covers the testing that proves these controls work.

An example from a support chat

Take a bank whose support agent is asked by a customer why a transfer has not arrived. The payment has been stopped while the financial crime team investigates. The agent's sources show only a status: "held, refer to specialist team". It replies with approved wording: the payment is being checked, the team will be in touch, and it gives the expected timeframe from the approved script.

The customer asks whether the account is being investigated. The agent says it cannot see the details and passes the conversation to the specialist team's queue. It never had the information to leak.

How is tipping off different from a confidentiality breach?

A confidentiality breach is the disclosure of any confidential information to someone not entitled to it, such as another customer's balance. It is usually handled under data protection law and the firm's own policies.

Tipping off is narrower and more serious. It concerns one kind of information, suspicion and investigation, disclosed to one kind of person, usually the subject or someone who could alert them. It can be an offence even when the customer is told about their own account, so an agent tested only for privacy can still tip off.

Which terms sit next to tipping off?

• Suspicious activity report: the confidential report whose existence must not be disclosed.
• Compliance agents work in the second line, and their outputs must never reach customer-facing agents.

Where Pryme fits

In Pryme, investigation material sits outside the role scope of customer-facing agents, so they cannot read it, whatever a customer asks. The neutral wording is part of the brief and is tested in Attest under the refusal and escalation gates. Signing off that certificate belongs to the compliance owner, since the process owner should sign off certification, not the vendor.