How to detect fictitious suppliers with an AI agent

Fictitious suppliers are set up by insiders, so the check must run where insiders cannot see it. The agent compares the supplier master file with employee data and payment patterns, scores each supplier against known warning signs, and sends cases only to internal audit or the fraud lead. It never holds payments, blocks suppliers or contacts anyone. People decide whether to investigate and how.

What is a fictitious supplier, and who owns detection?

A fictitious supplier is a business set up, or borrowed, to receive payments for goods or services that were never provided. The usual author is someone inside: a person who can add suppliers, raise orders or approve invoices, alone or with help. Payments are often modest and regular, designed to sit below approval limits and blend into normal spend.

Internal audit or the fraud lead owns detection and every case it produces. The data protection owner must approve the use of employee data for matching, because comparing staff bank accounts and addresses with suppliers is a significant use of personal data. The accounts payable team is deliberately not the owner: the person being looked for may be in it. The wider finance position is in AI agents for finance: books you can audit.

How does the agent detect them, step by step?

1. Agent, depth 1: reads the supplier master file. Names, addresses, bank accounts, tax numbers, registration details, who created each record and when. Evidence: the snapshot read.
2. Agent, depth 1: compares suppliers with employee data. Bank accounts, addresses and phone numbers shared with current or former staff, using the data scope the data protection owner approved. Evidence: each match and the fields that matched.
3. Agent, depth 1: checks registration. Whether the company appears in the public register where it claims to be registered, when it was formed, and whether its address is residential or a mail drop. Register data is external. Evidence: the result and its source.
4. Agent, depth 1: reads payment patterns. Invoices only ever approved by one person; amounts just below approval limits; round amounts; invoice numbers that rise one at a time, suggesting you are the supplier's only customer; services described vaguely; no purchase orders. Evidence: the pattern and the transactions behind it.
5. Agent, depth 1: checks set-up controls. Was the record created and later paid through actions by the same person? Evidence: user identifiers on creation, approval and payment.
6. Agent, depth 2: scores and prepares cases. Each supplier above the threshold gets a case: the signals, the evidence, and innocent explanations that would also fit.
7. Person, internal audit or fraud lead: decides. Dismiss with reason, monitor, or investigate. Any hold or block on payments is their decision, timed so as not to alert a suspect. Evidence: the decision and reason.

Why does the agent not hold suspicious payments?

In most finance processes, a hold is the safe default. Here it can be the opposite. A hold on a supplier's payment is visible to the accounts payable team. If the person behind the supplier sees it, they may destroy evidence, stop the scheme, or create a cover story. Investigators often prefer to watch a few more payments while they gather evidence.

So the agent's only output is a case to a restricted queue. Whether and when to stop payments is an investigation decision. This is one of the few places where an agent should be kept away from a protective action, and the reason should be written in its brief.

What stays with a person, and why?

Every judgement about a colleague. A match between a supplier's bank account and an employee's might be fraud, or it might be an employee who also runs a legitimate small business that was properly declared. A residential address might be a sole trader working from home. The agent lists the innocent explanations alongside the signals so the investigator weighs both.

Contacting the supplier, interviewing staff and involving authorities are all for people under your investigation procedures.

Which controls matter, and what evidence do they leave?

Restricted output. Cases go only to internal audit or the fraud lead. Evidence: the access list on the case queue.

Approved data scope. Employee data is used only as the data protection owner approved. Evidence: the approval and the fields used.

Innocent explanations recorded. Evidence: the explanation field on each case.

No protective actions by the agent. Evidence: no hold or block verdicts from the detection agent.

Every dismissal reasoned. Evidence: the reason on each closed case.

How does this relate to onboarding and duplicate checks?

Detection is the second line. The first is onboarding: checking registration, verifying bank details through a known contact, and separating who requests a supplier from who approves it, as described in onboarding new suppliers to the supplier master file. Duplicate checks overlap too, because fictitious suppliers are sometimes cloned from real ones with small differences; see detecting duplicate supplier invoices and can an AI agent detect duplicate invoices.

What should you measure?

• Cases raised and their outcomes: dismissed, monitored, investigated, confirmed.
• Time from supplier creation to first case, for confirmed cases.
• Dismissal reasons, which refine the scoring.
• Suppliers with no registration result, as a sign of onboarding gaps.

Where does it usually go wrong?

Results go to the wrong queue. Cases are posted where the accounts payable team can see them.

Employee data is incomplete. Former staff are excluded, but many schemes continue after someone leaves. Include leavers for a period.

Thresholds catch every sole trader. Small legitimate suppliers share many signals. Weight signals in combination, not alone.

Where Pryme fits

In Pryme, the detection agent's brief gives it read capabilities and a single output: cases to a restricted queue. It has no hold or block capability at all, so it cannot alert a suspect by accident. Every case and every dismissal is filed for internal audit.