A small business stops AI voice and payment scams with one rule, not a detection tool: before anyone moves money or changes payment details, call back on a number you already held before the request arrived. A cloned voice can pass a listening test. It cannot answer a phone it does not control.
This article calls that rule the Known Number Rule, and gives you a one-page card you can print today. It also shows what the evidence says about the size of the problem, where that evidence runs out, and how a 25-person firm can put the rule in place this week without buying anything.
What does AI change about payment fraud?
AI makes the fake more convincing. It does not change what the fraudster wants. The goal is the one it was before cloned voices existed: get a trusted person to send money, or to change where money goes, because a familiar name, voice or face asked.
What has changed is the cost of the performance. A scammer once needed a good ear and a plausible email. In a 17 September 2026 alert, the FBI's Internet Crime Complaint Center says scammers now use AI to "appear as law enforcement and government officials on video calls with victims" (FBI IC3, alert I-091726-PSA). The FBI's April 2026 announcement of its annual report lists voice clones among the tools in use (FBI, 6 April 2026).
For a small firm, the consequence is plain. "It sounded like him" used to be a reasonable check. It is now no check at all. Recognition has stopped being evidence, so the protection has to come from a procedure that does not depend on anyone's ear.
How big is the problem, and what do the numbers miss?
Business payment fraud is large, and the slice labeled as AI is small, which is a reason to fix the procedure before shopping for anything clever.
Start with the United States. The FBI's 2025 Internet Crime Report, covering complaints received by the Internet Crime Complaint Center in calendar year 2025, records 24,768 business email compromise complaints with total reported losses of $3,046,598,558. The report's first-ever AI section counts 22,364 AI-related complaints across all fraud types, with reported losses of $893,346,472, most of it investment fraud. Inside business email compromise, it says businesses reported losses of over $30 million to scams involving AI (FBI IC3, 2025 Annual Report).
That $30 million is roughly 1 percent of the business email compromise total. Read that carefully. It does not mean AI is a minor threat. The label depends on victims noticing and reporting that AI was involved, so it is a floor, not a measurement. My reading is simpler: most of the money still moves through a fake email and a changed bank detail, and AI is a better costume on an old trick.
Now the United Kingdom. The government's Cyber Security Breaches Survey 2025/26, published on 30 April 2026 from fieldwork between August and December 2025 with 2,112 businesses, found that 43 percent of businesses had experienced a breach or attack in the previous 12 months. Phishing affected 38 percent and impersonation 12 percent, down from 17 percent in 2023. By size, 46 percent of small businesses and 65 percent of medium-sized businesses were affected. The report also found that 58 percent of businesses have an agreed process for staff to follow when they meet a fraudulent email or website (GOV.UK, Cyber security breaches survey 2025/26). The sections I read did not report any deepfake figure, so I am not offering one.
Two limits to hold in mind. Neither dataset measures how often a cloned voice is tried and fails, which is the number a small firm would most like to know. And both are reports of what was caught and logged. The honest summary is that the threat is real, the AI share is growing from a low and uncertain base, and the defense that works against the old version works against the new one.
What is the Known Number Rule?
The Known Number Rule is one rule for any request to move money or change payment details: confirm it by calling a number you already held before the request arrived, never one supplied in the request. It has no exceptions for seniority, urgency or a familiar voice, and nobody can waive it by phone.
Three properties make it work.
First, it checks on a channel the fraudster does not control. A cloned voice on an incoming call, a fake video meeting and a spoofed email all arrive on the attacker's terms. Your own outgoing call to a number from your own records does not.
Second, it binds the boss. Fraud of this kind works by borrowing authority, so a rule with an exception for senior people has an exception for the thing being imitated.
Third, it is cheap. A call takes two minutes. The fraud, if it lands, can take the business's cash.
The rule is not our invention alone. The FBI's September 2026 alert tells people to verify using publicly available contact information and not "contact information provided by the person contacting you." The UK's National Cyber Security Centre tells defrauded organizations to contact their bank using its official website or phone number, not details from the suspicious message (NCSC, Business payment fraud). The US Federal Trade Commission's guide for small businesses says "don't trust caller ID" and recommends clear approval procedures for invoices (FTC, Scams and Your Small Business, page modified 23 September 2025). Three agencies in two countries, one habit.
A known number is one that sat in your records before the request: the accounting system, a signed contract, the staff directory. It is not the number in an email signature, a "new number" in the message, the number that just called you, or anything else that arrived with the request.
What does the rule look like on a card?
Put the rule on a single card that the person making the payment fills in, one card per request. A blank line on the card is a stop sign.
The Payment Change Card (copy it into a document or print it)
| Line | Fill in |
|---|---|
| 1. What was asked | New bank details, urgent transfer, other payment request |
| 2. Who appears to be asking | Name and role |
| 3. How it arrived | Email, call, video meeting, message, voice note |
| 4. Number we already held | Where it came from: system, contract or directory. Never the request itself |
| 5. Called by, and when | Name, date, time |
| 6. What they said | Confirmed, denied or no answer |
| 7. Second person | Name of the colleague who approved after seeing lines 1 to 6 |
| 8. Money moved | Yes or no, with date |
Two rules go with it. If line 4 cannot be filled in, nothing is paid until you have established a number another way, for example through a person at that organization you have dealt with before. And line 7 applies to every change of payment details and to any transfer above a limit you choose for your own business. The limit is yours to set.
Keep each card with the invoice or payment record. If something does go wrong, the card also tells you exactly who did what, which matters when you call the bank.
Which requests need the card, and which do not?
Only requests that move money to a new place, or move it in a hurry, need the card. Ordinary repeat payments do not, or the rule will be dropped within a month. Here is the decision table.
| Request | Card? | What to do |
|---|---|---|
| New or changed bank details for a supplier | Always | Call the number in your records; second person approves |
| Urgent transfer "from the boss" by call, video or voice note | Always, including for the boss | Call the boss on the number in the directory; no exceptions |
| Payroll bank detail change from an employee | Always | Confirm in person or on the mobile number on file |
| Invoice from an existing supplier, same details as last time | No | Normal approval; match it to the order or delivery |
| Request to pay by gift cards, cryptocurrency or courier | Never pay | The FTC says that if someone demands these, "it's a scam" |
| Caller claims to be your bank, a tax authority or police and asks for payment or details | Hang up | Call the organization on its published number |
| Anyone asks you to keep it confidential or skip the usual approval | Treat as a warning | Run the card, and tell a second person |
The last row is my judgment, not a quotation from any agency. My reading of how these requests are built is that secrecy and speed are there to stop the second person and the phone call from happening.
What I See When I Train Business Teams
Across the business teams I train, I see the same pattern, and I describe it here as a composite, not a single story. People know the rule in theory and break it for the person they like or fear. The assistant who would never pay a stranger will pay a managing director's voice at ten to five on a Friday, because checking feels like doubting the boss.
The fix is cultural, not technical. The owner or managing director says out loud, to everyone with payment authority, that the rule applies to them and that they would rather be called back ten times than be right once. Staff then have permission, which is what they were missing. If you want the wider picture of how a team uses AI day to day, the Friday Four is a fifteen-minute weekly conversation that gives odd requests a place to be mentioned.
Does "a human checks it" protect you here?
Not by itself. This is where the Loop Test earns its place. The Loop Test is three questions I ask before accepting "a human checks it" as an answer: Would that person notice a wrong result without being told to look for it? Would they notice in time to stop it mattering? And if they missed it, could the work be put back? Oversight that survives all three is a safeguard. Oversight that fails any of them is a signature. (I use it elsewhere, for example in AI for retail and hospitality.)
Apply it to a cloned voice. Would the person notice? No, because the voice is the thing that makes it look right. In time? A same-day transfer is gone before doubt arrives. Could it be put back? Sometimes, if the bank is called at once, but often not. A human who only listens fails the Loop Test on all three counts. A human who makes a call to a known number passes the first, because the call itself does the noticing, and passes the second because it happens before the payment.
Here is a worked example, and it is hypothetical. Take a 25-person architecture practice. A partner is traveling, and the finance manager receives a voice note, apparently from him: a supplier needs paying today to a new account, and he will explain later. Under the old habit she pays, because it is his voice. Under the Known Number Rule she fills in the card, calls his mobile from the staff directory, and either he answers and says he sent nothing, or he answers and says he did, and she has lost five minutes. Both outcomes are fine. The only bad outcome is the one with no call.
What about the boss who really is unreachable?
Then the payment waits, or it was never an emergency. The rule does cost a little time on real requests, and I would rather say that than pretend otherwise.
You can reduce the friction without weakening the rule. Agree standing instructions in advance, in writing, for payments you know are coming: named payees, named accounts, a limit. A payment inside that agreement needs no card. A payment outside it does. This is the same idea as the Small Room Test: if the one person who knows the list of known numbers is away for two weeks, would the rule still work? If not, write the list down where two people can reach it.
What can you do this week?
Five steps, none of which needs a data team or a budget line.
- Print the Payment Change Card and put it where payments are approved. Fill in a practice one together so people have seen it done.
- Build the known-numbers list for your ten most important suppliers, your bank, payroll and your senior people. Take each number from a contract or your accounting system, not from email.
- Tell everyone with payment authority that the rule exists, and have the owner say it applies to the owner.
- Run a ten-minute drill. The owner, from a personal phone, asks the finance person to change a supplier's bank details in a dummy request and watches whether the card gets filled in. Nothing needs cloning or faking beyond a message.
- Agree what happens if it goes wrong. Call your bank first, using the number on your card or its official website. In the UK, report through gov.uk/report-cyber, as the NCSC advises. In the US, file a complaint at ic3.gov, as the FBI advises.
Then put the rule in writing. If you do not yet have an AI and security policy, does a small business need an AI policy is the place to start, and the rule fits on its first page. Add it to the first week for new starters as well: AI training for new hires covers what a new person should be shown, and finance staff should see the card on day one. Invoice handling and approvals sit in AI for finance and admin.
Where to go next
If you want to know where your business stands on AI risk, rules and skills before you decide what to fix first, take the free Business AI Readiness self-assessment: twelve questions, about five minutes, and a 90-day plan at the end. If your team needs to practice the habit rather than read about it, that is what AI training for teams is built for. And if you run a chamber, association or business event and want this in front of a room, I give business keynotes on practical AI risk; I have delivered more than 500 keynotes in 30 countries across 5 continents.
Sources and further reading
- FBI Internet Crime Complaint Center, 2025 IC3 Annual Report, calendar year 2025 (United States; complaints to IC3).
- FBI, Cryptocurrency and AI Scams Bilk Americans of Billions, 6 April 2026.
- FBI IC3, Scammers Impersonating Law Enforcement and Government Officials in Fraud Schemes, alert I-091726-PSA, 17 September 2026.
- UK Government, Cyber security breaches survey 2025/26, 30 April 2026 (United Kingdom; 2,112 businesses; fieldwork August to December 2025).
- National Cyber Security Centre, Business payment fraud, accessed 28 October 2026 (the page shows no date).
- US Federal Trade Commission, Scams and Your Small Business: A Guide for Business, page modified 23 September 2025.
Dan Fitzpatrick, The AI Educator, is a keynote speaker, Forbes contributor and five-times bestselling author who has trained more than 150,000 educators in 30+ countries. About Dan.
