The next fraudster is already inside

In short: The fraudster used to need skill, time and nerve. AI removes all three. In KPMG’s two-part Week of Integrity workshop, Muel Kaptein, Nicole van Klooster and Charlie Groen ran the room through an insider AI fraud risk assessment: first think like a fraudster, then like a manager. The uncomfortable takeaway is that the dangerous insider is no longer a profile, it is almost anyone with a login and a chatbot. The fuller account is below.

KPMG’s workshop came with a title designed to unsettle, “Insider AI Fraud: The Next Fraudster is Already Inside,” and it kept its promise. Run twice during the week, as an early session led by Muel Kaptein and a lunchtime session led by Nicole van Klooster with Charlie Groen, it was not a lecture but a live risk assessment, with the audience doing the work in the chat. The framing was deliberately narrow and repeated at the top of both sessions. This is not about whether AI itself is trustworthy, and not about hackers breaking in from outside. It is about insiders, employees, managers, even board members, using AI to commit or hide fraud. As van Klooster put it, “the traditional way of working by fraudsters is changing rapidly.”

The ground has shifted

A few numbers set the scene. Around half of employees already use AI at work and at home, and roughly 35 per cent use AI tools without their employer’s knowledge or permission, the quiet spread known as shadow AI. Boards are in it too, using AI to digest an 800-page board pack or to generate the questions they will ask in a meeting, which raises van Klooster’s sharpest governance worry: the blurring line between AI that assists a decision and AI that effectively makes a decision the board remains accountable for.

Kaptein made the human point that gives the workshop its title. KPMG’s long-running profile of the typical fraudster, historically “a male in his 40s or 50s, very loyal,” no longer holds. “Everyone can now become a fraudster,” he said. The reason is a change to the oldest model in the forensic toolkit. The classic fraud triangle of pressure, opportunity and rationalisation has become a fraud diamond, with a fourth corner added: capability. AI supplies the capability. Someone with little expertise, no time and no resources can now run a scheme that once needed all three. One prompt changes a figure on an invoice, and the whole chain of supporting documents can be regenerated to match.

AI pushes on the other three corners too. It raises the pressure, by lifting everyone’s expectations of speed and output. It widens the opportunity, helping people find routes to fraud at greater scale and then cover their tracks. And it eases the rationalisation, because delegating the dirty work to a machine lets people hold it at arm’s length. Kaptein cited a study in Nature in which people told to have an AI agent commit fraud on their behalf were markedly less hesitant. “They don’t make their hands themselves dirty.”

Think like a fraudster

The heart of the session was an invitation to brainstorm, out loud, how you would actually do it. The room obliged: fake and modified invoices with a matching paper trail, falsified expense claims, CEO impersonation using a voice cloned from a voicemail, because “three seconds of voice recording is enough,” deepfakes, fake diplomas, ghost employees on the payroll, insider trading, cooking the books, and the subtlest category, asking an AI agent to reach a target that can only be hit by cutting a corner, so the instruction to cheat is never actually given.

One case recurred in both sessions and lingered. An employee at a research-heavy company coded a hidden backdoor into a critical server, resigned, and then used simple AI prompts to reach in from the outside, steal the intellectual property and set up in competition. It was traced only through a brief, telltale change in server access rights. The lesson the presenters drew was that the insider starts, by definition, already past the perimeter. There are no external defences left to circumvent.

The room did not simply nod along. One participant argued that fraud is “systemic from the nature of the individual,” not a function of the tools, and was sceptical that easier capability really produces more fraud. Van Klooster conceded the force of it, agreeing that rationalisation remains “the key in the lock.” Others pressed the practical risks, in particular the discipline of labelling and classifying data so that confidential material is never casually fed to an AI tool or agent in the first place.

Think like a manager

The second half flipped the exercise from offence to defence, organised around three words: prevention, detection, response.

Prevention is mostly about not relying on trust. Strengthen approval processes that lean too heavily on a familiar name, so that a sensitive payment or a change of bank details requires a phone call or a face-to-face confirmation, not just an email that looks right. Set a clear AI policy and acceptable-use rules, train people in AI literacy, run phishing tests, and classify data properly. One concrete KPMG example: staff barred from USB sticks and external devices altogether.

Detection is about verification and patterns. Check claims against the original source and the real person, monitor transactions for abnormal behaviour, and watch for the quiet signals, such as an unexplained change in someone’s access rights. Make reporting easy, including anonymously, because roughly a third of fraud still surfaces through a tip from an employee rather than any system.

Response turns on containment and clear heads. Preserve the logs and mailboxes within privacy limits, restrict access, bring in legal, compliance and security early, and work out whether you are looking at a mistake, negligence or deliberate fraud. A recurring tip was to go physical when the digital evidence is forgeable: to confirm a ghost employee is real, do not ask for another document, go and check that the person exists.

A nice thread running through both sessions was cultural. Positive guidance, telling people to “use AI responsibly,” works better than a flat “do not commit fraud,” and employees are best understood as gatekeepers rather than suspects. There was also a running joke that made the point better than any slide: Microsoft Copilot kept flagging the workshop’s own chat as harmful, because a room full of people brainstorming fraud techniques trips its filters. It would only help once the request was reframed as preparing a fraud risk assessment, which is, of course, exactly the reframing a real fraudster would make.

What to do tomorrow

Both sessions ended with the same small, pointed ask: name one process you will go back and review tomorrow. That is the spirit of the whole workshop. The tools are new and the capability is now universal, but the defences are recognisable and mostly human: verify, segregate duties, keep a person in the loop on anything irreversible, and watch what your people are feeding into AI as closely as you watch what they take out. Or, as van Klooster put it with a half-smile that was not entirely a joke, “most of us could also be a very good fraudster.” The best option remains the oldest one: “keep the fraudster outside.”

The Week of Integrity continues year-round, with a spotlight on International Anti-Corruption Day on 9 December.