AI Security for Small Business: Stop Deepfake Fraud

The accounts payable clerk gets a call. It is the CEO's voice, same cadence, same impatience, asking to push an urgent wire before a deal closes. Minutes later a video call appears to confirm it, the CEO's face on screen, asking for updated vendor bank details.

Nobody hacked a server. Nobody was careless. The attacker simply sounded and looked credible enough to skip one verification step.

Here is the rule that stops this: a voice, a face, and an email are claims of identity, not proof. The confirmation has to travel a different path than the request.

The process is the tell, not the voice

A scam can be entirely believable in content while being abnormal in process. Right vendor, right amount, right reason, wrong channel. The invoice looks correct. The person sounds correct. The only thing wrong is the path the request took to reach the person who can act on it.

That gives your team a test that does not depend on catching a fake. When the content seems fine but the process feels off, the process is the warning. Your finance person does not need to detect a deepfake. They need to notice a request that skipped a step.

Right vendor, right amount, right reason, wrong channel.

The attacker is buying three shortcuts. A familiar voice feels like identity, so the caller must be who they say. Video feels like stronger proof than it is, so a face on screen ends the question. Urgency feels like authorization, so speed replaces approval. All three are assumptions, and all three are exactly what the impersonation is designed to exploit.

Out-of-band, in plain English

The defense is confirming the request through a different channel than the one it arrived on. In plain terms: if the request came in by phone, the confirmation cannot happen on that phone call. Hang up. Use the number you already had on file.

That only works if the trusted channels exist before anything goes wrong. Pre-registered phone numbers, the internal chat tool, face-to-face for local teams. A policy invented mid-incident is not a policy, it is a guess made under pressure.

Name the high-risk actions once, so nobody has to judge in the moment. Moving funds, changing vendor bank details, editing payroll, granting data access, resetting credentials, releasing sensitive files. Any request to do one of these gets confirmed on a separate channel, every time, no exceptions for people who sound senior.

Vendor invoice fraud is the clearest example. A message arrives that sounds routine: a supplier updating their banking details ahead of the next payment run. The amount is normal, the vendor is real, the timing fits. One callback to the number already on file, not the number in the message, ends it in ninety seconds.

Executive emergency transfers work the same way, a rushed wire request from a senior voice on an unfamiliar number. So do deepfake meetings, where a fabricated executive on video asks for files or credentials. Same defense, same callback, every time.

The checklist

Paste this into your finance channel today. Every line is an instruction.

1. Stop if the request is urgent, unusual, or outside normal workflow.
2. Verify through a channel already on file, never one supplied in the request.
3. Call back for any wire, vendor banking change, payroll edit, or sensitive release.
4. Require two approvers above a set dollar threshold.
5. Confirm the request in writing on a second channel before acting.
6. Escalate when pressure is applied: "do this now," "don't tell anyone," "I'm in a meeting."
7. Report and log every flagged request so leadership can spot patterns.

The checklist only works if the culture backs it. An employee who pauses a legitimate request loses two minutes. An employee who guesses wrong loses the wire. Make the safe choice the expected one, and make escalation blame-free. A finance person who calls to double-check the CEO should be thanked, not embarrassed.

This is the part that does not require a budget. Small businesses do not need expensive detection tools to shut this down. They need one rule that money and identity requests confirm on a second channel, and a team that treats a two-minute pause as normal.

Get the policy in place

If your team moves money, updates vendor details, or releases sensitive data, book a Webspenser AI security review. We map your high-risk workflows and hand you a ready-to-publish verification policy your team can use the same day. This post is the operational half of a larger picture. The reasoning behind why identity claims need independent confirmation, and where verification fits in a broader accountability approach, lives in our guide to building AI systems you can hold accountable.