triox informatika

July 21, 2026 · Triox Informatika · 5 min read

Your boss calls and asks for a transfer. Except it isn't your boss.

For most companies, the most tangible effect of artificial intelligence isn't efficiency—it's that fraud has become indistinguishable from the real thing. What you can do about it—mostly without software.

In brief

  • Fraudsters use the same tools you do: artificial intelligence that writes text and imitates voices.
  • That hasn’t made them smarter. What made it dangerous is that looking credible has become cheap.
  • The old warning signs—poor Hungarian, an odd greeting, a botched logo—are disappearing one by one.
  • Defense, however, has stayed simple: a few rules that everyone in the company knows are worth more than any filter.

Picture a Monday afternoon. Someone in finance gets a phone call. On the other end is the managing director’s voice—with their own intonation, their usual half-finished sentences. A supplier invoice has to be paid ahead of the normal schedule, because a contract is stuck and has to be closed today. The employee makes the transfer. Two days later it turns out that the managing director was in negotiations all day and never made that call.

A few years ago this story belonged in a movie. Today a few minutes of public recording—a conference video, a radio interview, a greeting shared on social media—is enough for someone to convincingly imitate a person’s voice. And recordings of company executives are typically available online.

What used to give fraud away no longer does

Over the past ten years, we mostly recognized phishing emails because you could tell they had been made with a translation program. Odd word order, missing accent marks, sentences straining for an official tone yet still sounding foreign. Companies’ security training was built on this too: “check the spelling.”

That handhold is gone. Today’s language models write flawless Hungarian, know the phrasing of formal correspondence, and if they get a few samples of the company’s public communication, they get the style right too.

How we used to recognize fraud Why it doesn’t work today
Bad spelling, sentences that don’t sound like real Hungarian The text is written by a machine, flawlessly
Generic greeting (“Dear Customer,”) The name, the job title and the project name are all correct
Botched logo, off-brand look A fake site can be copied pixel for pixel
“I’d recognize their voice live, anyway” A voice can be imitated from a few minutes of recording
Unknown sender The request may come from a compromised, genuine partner account

Fraud hasn’t become more dangerous by getting smarter. It has become more dangerous because looking credible has become cheap.

Why smaller companies in particular?

It’s tempting to assume that attackers are interested in large corporations. Practice shows otherwise: at a twenty-person company, a transfer is often initiated by a single person, the decision chain is short, and everyone knows the boss—so “the boss asked for it” is reason enough on its own.

On top of that, preparing an attack costs almost nothing today. There’s no need to research targets one by one: the model writes the email, generates the voice, and copies the company details from the public company registry and the website. What used to be a targeted, expensive attack against a few large companies can now be sent out en masse.

The chain can always be broken at the same point

The deception is almost always the same three steps. And it can almost always be stopped at the same point: between the second and the third step, by verifying the request through a different channel.

The three steps of fraud: a credible approach, then pressure and a request for secrecy, and finally the transfer or the handover of data. The chain breaks before the third step if the request is verified through a different, known channel. 1. APPROACH Familiar voice & style call, email or voice message 2. PRESSURE “Do it right now” and “don't tell anyone” 3. LOSS Wire or password no way back from here CALLBACK TO KNOWN NUMBER the chain breaks here

The key to verification is not to use the contact details given in the request. If the email includes a phone number, we don’t call that number back. We call the one in the address book, the contract or the partner records.

Five rules that cost nothing

  1. Callback rule. Any request involving a transfer, a change of bank account number or a password is always verified—through a different channel, using contact details known from before.
  2. Four-eyes principle above a threshold. Above an amount agreed in advance, two people approve the transfer. This rule isn't about trust; it's about the fact that one person is easier to deceive than two.
  3. Urgency itself is the warning sign. Someone who asks for something right now and tells you not to mention it to anyone is trying to keep you from exactly that verification. A real executive won't be offended by a callback.
  4. Bank account numbers only change on paper. We never accept a change to a partner's bank account number by email—this is one of the most common types of fraud today, and it causes serious losses.
  5. Make it explicit that mistakes are allowed. The size of the loss almost always depends on how many hours pass between the suspicion and the report. If an employee is afraid of the consequences, they keep quiet—and that's what costs the most.

These five rules fit on a single page, and you don’t need any software for them. They work because everyone knows them, and the boss sticks to them too.

What’s worth doing on the systems side

It’s worth backing the rules with a technical layer—not because it protects you on its own, but because it keeps a good share of attacks out of the mailbox in the first place, and makes the rest traceable.

Technical details
  • Multi-factor authentication (MFA) on every account, especially email—most abuse starts with a stolen password
  • Email authentication set up (SPF, DKIM, DMARC), so that it isn’t easy to send email in the company’s name
  • Visual tagging of external senders in the email system—this breaks the illusion that a message “came from inside”
  • Anti-phishing and impersonation protection turned on in the email platform, with a maintained list of partners
  • Logging and alerts for unusual sign-ins and for the creation of mailbox rules—attackers typically create a forwarding rule to delay being found out
  • Regular review of permissions: who has access to the online banking interface and the partner records

If it happens anyway

At that point what matters is time, not explanations. Notify the bank first—in some cases a freshly initiated transfer can still be recalled. In parallel, change passwords and sign the accounts out of all sessions, because the deception often doesn’t end with the transfer. What happened has to be documented, and a police report filed; the insurer and the auditor will ask for this as well.

The question isn’t whether your employees will “fall for it.” It’s whether there’s a habitual move that stops the process before the money goes out.

Let’s see where your company stands now. It doesn’t take a big project: turning on MFA, getting email authentication in order and a half-page internal rule together handle the vast majority of attacks. If you’re curious which of these you already have in place, write to us—you can read more about our approach to security here.

← Back to the blog