Enterprise fraud-detection software is priced for banks, and for a South African SME it is almost never the right purchase. What is worth having is much smaller and much cheaper: a system that flags the handful of patterns that actually take money out of small businesses here, and puts them in front of a person before the payment goes out.
What fraud actually hits South African SMEs?
Not the exotic kind. Four patterns account for most of it:
- Banking-detail changes on an invoice. Somebody intercepts or spoofs a supplier email and sends "our new account details". This is the single most damaging one for small businesses, and it works because the request looks routine.
- CEO impersonation. An urgent payment request that appears to come from a director, usually late on a Friday, usually with a reason not to phone and check.
- Proof-of-payment forgery. A convincing-looking EFT confirmation, goods released, money never arrives.
- Card-not-present chargebacks on online stores, where the goods have shipped before the dispute lands.
Every one of these is a process failure before it is a technology failure, which is why software alone does not fix it.
What can AI genuinely do here?
| Job | Automatable? | Note |
|---|---|---|
| Flag an email requesting changed banking details | Yes | High value, low complexity — the pattern is distinctive |
| Flag a payment request that breaks your normal pattern | Yes | Unusual amount, unusual supplier, unusual hour |
| Spot a lookalike sender domain | Yes | A character swap is trivial for software and easy for a person to miss |
| Verify a proof of payment | Partly | It can flag oddities; only the bank statement proves it |
| Decide whether to release goods | No | A person, against cleared funds |
| Accuse an employee | No | A flag is a question, never a conclusion |
What controls matter more than the software?
Three, and they cost nothing:
- Never change banking details off an email. Phone the supplier on the number you already had, not the one in the message. This single rule stops the most expensive pattern on the list.
- Two people on any payment over a threshold you set. Pick a number that is meaningful for your business and hold to it, including on Fridays.
- Release against cleared funds, not a proof of payment. A screenshot is not money.
Software makes these easier to follow. It does not make them optional, and any vendor implying otherwise is selling you comfort.
Where does an AI email assistant fit?
This is the realistic entry point for a small business. An assistant reading the shared inbox can flag a message that asks for a banking change, a message from a domain that resembles a known supplier's, or an urgent payment request that does not match your normal pattern — and route it for a second look instead of letting it sit in a queue with forty others. Our AI email assistant starts at R349 a month plus R500 setup, which is the scale this problem deserves for most SMEs.
What about false positives?
You will get them, and that is the correct trade. A flag costs somebody thirty seconds; a missed banking-change fraud costs an invoice. What matters is that a flag is designed as a prompt to check, not as an accusation — and that the person checking has a clear, quick way to confirm. A system that generates flags nobody has time to action is worse than no system, because it manufactures the feeling of safety.
The email asking you to update a supplier's banking details does not look suspicious. That is the entire design of it, and it is why the control has to be a phone call rather than a careful read.
What should a small business do first?
Write down the banking-detail rule and tell everyone who touches payments. That is free and it is the highest-value thing on this page. Then, if your inbox volume is genuinely too high for anyone to read carefully, add an assistant that flags the patterns above. Do not start with enterprise fraud software — the price is not the only reason it is the wrong tool.