How do I spot phishing when the language is flawless?

Not from the language but from the request. A scam asks for something the normal process never asks for: an urgent payment, a change of bank account, a login through a link, gift cards or a password confirmation. It adds time pressure and a reason to skip the usual check. A familiar name, the right logo and fluent language prove nothing, because they are the easiest part to fake.

What do I do when I suspect a message?

You do not reply, open the attachment or click the link. You verify the request through another channel: you call the sender on a number you already have, or ask them in person. You report the suspicion to whoever is responsible for security, even if the message turns out to be genuine. If you already clicked or entered your credentials, you report immediately: a fast report limits the damage, silence grows it.

Does the law require staff to be trained to recognise scams?

Within the scope of the Finnish Cybersecurity Act, yes. Section 9 lists the risk management measures an entity must keep up to date, and among them are personnel security and cybersecurity training, and basic information security practices. Article 21 of the NIS2 Directive says the same: basic cyber hygiene practices and cybersecurity training are a minimum measure. Other organisations do not get the obligation from this act, but they get the same risk.

What does AI change about scams?

It removes the old telltale signs. The message is grammatically perfect and written in the sender's style, a phone call can carry your manager's voice and a video meeting can show their face. Recognition therefore moves from content to procedure: an unusual request is always verified through another channel, however convincing it is.

Phishing and AI-assisted scams: how an employee recognises one

Phishing no longer gives itself away by its grammar. What employees must recognise, why unusual requests are verified through another channel, and what the law requires.

In brief

Phishing is a scam in which a message, a call or a website poses as a trusted party to make the recipient hand over credentials, pay an invoice or install a program. It is still the most common way an attacker gets inside an organisation, because it does not break a system but a person's ordinary day. AI has taken away its old telltale signs: language errors, clumsy style and odd sender addresses. Recognition therefore no longer rests on what a message looks like but on what it asks for and how the request is verified. The Finnish Cybersecurity Act makes this skill part of statutory risk management.

Why grammar no longer tells you anything

A few years ago workplace security training taught people to look for mistakes: poor language, the wrong logo, an address with one letter out of place. That advice is not wrong, but it is no longer enough. A language model writes flawless text and imitates the sender's style when given a few of their earlier messages. A short recording is enough to clone a voice, and the other side of a video meeting can be your manager's face and voice without your manager.

Our skills catalogue calls this skill recognising phishing and manipulation, and its content is email phishing, AI-assisted scams and deepfakes, phone scams and CEO fraud. What they share is that they are first of all social engineering, not technology. An attacker does not need a vulnerability if they can get a person to do the thing themselves.

What a scam asks for

When appearance tells you nothing, you look at the request. Scams share a structure regardless of channel.

Sign	What it looks like at work
An unusual request	A new bank account on an invoice, an urgent payment, buying gift cards, confirming a login through a link, opening an attachment "for review"
Time pressure	"Today", "before the meeting", "the director is on a plane and cannot reply"
A reason to skip the process	"Do not mention this to anyone yet", "the normal route is too slow", "this is a confidential project"
Familiarity	The right name, the right title, a reference to a real project or a recent event
A change of channel	An email asks to continue on WhatsApp, a call asks for confirmation by text message

One sign proves nothing. Urgent and unusual requests happen in real work too. What matters is not that the request is unusual but what happens next.

The second channel decides

The rule that survives AI is simple: an unusual request is always verified through another channel that does not come from the message itself. If an email asks to change a supplier's bank account, you call the supplier on the number already in the contract, not the one in the message signature. If your manager asks for an urgent payment on Teams, you walk to their office or call them. If the bank calls, you hang up and call the bank's own number.

This is level 2 of our skills catalogue word for word: recognises the signs of a phishing message, verifies an unusual request through another channel and reports the suspicion. It is a level every employee can be expected to reach, and it requires no technical skill. It requires that the organisation has said out loud that checking is allowed and wanted, also when the person asking is the chief executive.

Keep the reporting threshold low

The other half of the skill is reporting. An employee who suspects a message reports it to whoever is responsible for security, and an employee who already clicked reports at once. The second case decides the size of the damage: a stolen credential is useful to the attacker only until the password is changed and sessions are closed. Every hour of silence is the attacker's hour.

That is why at level 3, the supervisor's level, the catalogue requires that reports are handled so that the reporting threshold does not rise. An employee who was scolded for a mistake does not report the next one. An organisation that thanks people for a false alarm also gets the real ones.

What the law requires of training

The Finnish Cybersecurity Act (124/2025) brought the NIS2 Directive into Finnish law. Section 9 obliges an entity to implement the control measures of its risk management model and lists what the model must at least cover and keep up to date. The list includes personnel security and cybersecurity training, and basic information security practices for operations, communications, hardware, software and data. Article 21(2) of the NIS2 Directive names basic cyber hygiene practices and cybersecurity training as a minimum measure.

The law does not say how many hours or how often. It says that training is part of risk management, and risk management measures must be proportionate to the nature of the operations and the risks. Phishing is a risk that concerns everyone with an email address, so recognising it belongs in everyone's training. Who the act itself applies to is explained in Does NIS2 and the Cybersecurity Act apply to your organisation?, and the training obligation as a whole in What does the Cybersecurity Act require for staff training?.

What this does not mean

Recognition does not replace technical controls. Multi-factor authentication, spam filtering and an approval chain for payments catch a large share of attempts before anyone sees them, and section 9 of the Cybersecurity Act requires those too. Human recognition is the last layer, not the first.

Article 50 of the AI Act obliges the deployer of a system that produces a deepfake to disclose that the content is artificially generated. A scammer does not comply. The article helps to distinguish honest AI content, not to expose a criminal, so a workplace instruction cannot rely on labelling. What the transparency obligation requires of your own organisation is explained in Article 50 of the AI Act.

How the skill is taught so that it lasts

One annual training session does not maintain the ability to recognise scams, because scams change faster than the calendar. The skill lasts when it is revisited briefly and regularly, and when it is practised with real examples: a message where everything looks right and only the request is wrong. Why short and repeated beats a training day is explained in Microlearning at work.