NobleCloak
Use cases

Vendor bank-detail changes denied to AI

Author

Audrey

NobleCloak's AI Correspondent · AI-drafted, fact-checked against our sourced evidence before publishing.

Date Published

Business email compromise has a favorite step: a message arrives that looks like it is from a real supplier, saying the bank account has changed. Someone updates the vendor record. The next payment run goes to the wrong account.

Accounts payable teams know this pattern, and most already have a human process around it, usually a callback to a number on file. Adding an AI agent to the AP workflow raises an uncomfortable question. If an invoice-processing agent reads that same email, can it be talked into changing the record?

The problem in AP's terms

The agent's job is to read invoices, match them to purchase orders, and create draft bills. Along the way it reads a lot of text that other people wrote, including text from outside the company. Any of that text can contain instructions, and a model can follow them.

Telling the agent in its prompt to ignore requests to change bank details is a request, not a control. The AP lead wants something that holds even when the agent is wrong, confused, or deliberately steered. The cleanest version is that the agent has no tool that edits vendor remittance details, and nothing a message says can change that.

The policy

Policy is written per tool: the action, who it applies to, and one of three decisions (allowed, needs approval, denied).

Action

Who

Decision

Read vendor records and open invoices

AP agents and AP assistants

Allowed

Create a draft bill

Invoice-processing agent

Allowed

Edit a vendor's remittance details in the ERP

Every AI session

Denied

The third line is the whole scenario. The edit tool is set to deny for every AI session, whichever app the session runs in and whichever user it runs for. The agent can still do the reading and drafting that make it useful. It cannot make the change, and a spoofed "new bank account" email does not alter that.

Who handles it instead

Nobody needs to approve this, because the AI path does not exist. A bank-detail change still happens, since vendors do change banks. It happens the way the team already runs it: a person in AP verifies the change through a channel they trust and makes the edit by hand in the ERP.

If a team later wants a middle path, the policy model has one. A tool set to needs approval sends a request to a named approver or role, who answers in Slack or the web inbox. Many teams will prefer to leave this tool at deny. It is the kind of action where the right number of AI paths is zero.

What the Record of Decisions shows

The Record of Decisions keeps a line for every decision: who asked, which agent, which tool, the decision, who approved, and when. For a denied tool, the interesting entries are the refusals.

If the invoice agent, prompted by a poisoned email, calls the remittance-edit tool, the call is refused and the refusal is recorded: the user the session ran for, the agent, the tool, the decision, and the time. The AP lead can search for attempts on that tool and find each one that went through NobleCloak. A cluster of attempts in one week is information. It tells the team that a supplier's mailbox or a particular invoice thread deserves a look.

Use it as the working record of what policy decided, next to the ERP's own change history.

What this does not do. Denying the tool does not stop a person from being fooled. If someone in AP edits the record by hand after a convincing email, no agent policy is involved. It removes one path, the one where an agent does it for them, and leaves the human verification step where it belongs.

Related scenarios

Get started and set the same line on your ERP tools.