AI was supposed to take scammers’ jobs, but it gave them superpowers instead

For anyone worried that AI would eliminate every profession except the one sending messages about your frozen crypto account, there's unwelcome news for that last bastion of human enterprise. Fraud has found an automation budget.

Scamming has always required a surprising amount of manual labor. Romance scams need someone to keep a conversation going for weeks, investment fraud needs someone to answer questions and maintain a believable identity, and impersonation scams need someone to keep the act together long enough to transfer money. AI can now do more of that work, meaning the same criminal operation can reach far more people without adding more humans.

That's what FATF president Giles Thomson meant when he told the Financial Times this month that AI could let “one or two people in a basement with a very big server” do work that once required a much larger scam operation. It's a grimly efficient version of the productivity boom every other industry has been promised.

The important change isn't that AI is about to make scammers unemployed. It's that every scam becomes cheaper to run. One person can maintain more fake identities, carry on more conversations, and attempt more fraud at the same time, which means criminals don't need a giant operation to create the appearance of one.

For an industry built around stealing other people's money, that's a fairly compelling productivity gain.

Fraud has discovered scale with AI

Chainalysis found average on-chain revenue of $3.2 million per scam operation with observed links to AI vendors, compared with $719,000 for operations without those links, or roughly 4.5 times as much revenue per operation.

That doesn't mean buying an AI subscription automatically quadruples a criminal's income. Larger operations may simply be more likely to buy sophisticated tools, and Chainalysis can't see every use of AI through blockchain data.

What the numbers do show is why criminals have an incentive to automate.

Traditional confidence fraud is labor-intensive because somebody has to maintain the illusion. A fake investment adviser needs to answer questions, a romance scammer needs to remember previous conversations, and an impersonator needs to sound enough like a colleague or executive that the target doesn't become suspicious.

AI makes that attention cheap. Criminals don't need software to feel empathy, only software that can imitate empathy well enough to keep someone talking.

That allows one operator to maintain more conversations, in more languages, for longer periods of time, while generating convincing documents, images, voices, and identities around them. The expensive part of the scam used to be maintaining the human performance. Increasingly, parts of that performance can be rented from a model.

The losses are already substantial. The FBI's summary of its 2025 Internet Crime Complaint Center report recorded 22,364 complaints containing AI-related information and approximately $893.3 million in adjusted losses.

Those cases include fake romantic identities, business impersonation, and other forms of fraud built on convincing someone they're dealing with a real person.

An AI provider has documented an even more direct example of the staffing change. Anthropic's August 2025 misuse report described an actor using Claude Code in an extortion campaign targeting at least 17 organizations.

According to Anthropic, the model assisted with parts of the technical work, analyzed information, and helped prepare extortion demands that sometimes exceeded $500,000. The company subsequently banned the accounts and shared information with authorities.

The important part isn't the particular model or the number of victims. It's that work that might have previously required different people with different skills can now be coordinated by one operator, with software helping across several parts of the operation.

The victim doesn't see that smaller organization. They see a convincing email, a believable identity, a professional-looking service, or a person who appears to know exactly what they're talking about.

AI lets a tiny operation present the surface area of a much larger one.

For crypto, that can be particularly effective because the distance between persuasion and payment is so short. A scammer can spend days or weeks building trust, but once the victim agrees, moving crypto can take seconds. Making the first part cheaper means criminals can push far more people toward the second.

The scam factories aren't disappearing

There's a tempting joke in all of this about AI taking scammers' jobs, but it stops being funny around the people working inside actual scam compounds.

Many of them aren't willing participants.

Amnesty International's 2025 investigation into Cambodian scam compounds documented at least 53 sites and interviewed 58 survivors from eight nationalities. It found evidence of trafficking, forced labor, confinement, and violence.

Those operations haven't been replaced by two men and a server. FinCEN's September analysis of Southeast Asian scam centers describes large transnational criminal organizations operating alongside AI-enabled services and other specialized criminal infrastructure.

The two models can exist at the same time because automation doesn't necessarily cause an industry to shrink. Sometimes it simply increases how much the same workforce can produce.

A scam operation that can automate part of every conversation has two choices. It can attempt the same amount of fraud with fewer people, or keep the people and dramatically increase the number of targets.

For victims, the difference isn't comforting at all.

This is why the idea that AI will put scammers out of work misses the more important economic change. The relevant unit isn't how many people the criminal enterprise employs. It's how cheaply it can attempt another fraud.

If the cost of producing a convincing fake identity falls, more fake identities become economical. If one person can supervise dozens of conversations instead of five, the number of people a gang can approach expands accordingly.

The internet already made distribution almost free. AI is now reducing the cost of persuasion.

When the bots start talking to each other

The defense industry has responded with automation of its own.

O2 created Daisy, an AI grandmother designed to keep phone scammers talking for as long as possible. In the company's account of the experiment, Daisy had answered scammers more than 1,000 times and spent hundreds of hours in conversation, with some calls lasting around 40 minutes.

The idea is wonderful because the scammer expects an elderly victim and instead receives an inexhaustible woman with nowhere else to be and no bank account to empty. It works by attacking something that used to be scarce: the scammer's time.

But that becomes less powerful when the scammer's side of the conversation is automated too.

If one bot spends 40 minutes discussing a fictional bank problem with another bot pretending to be somebody's grandmother, neither criminal nor victim has lost 40 minutes of human life. Somewhere, two organizations can congratulate themselves on engagement while the electricity meter does most of the work.

That absurd endpoint shows where the defensive problem is moving. Making scammers waste time still helps when human attention is expensive. As AI makes that attention cheaper, defenses have to move closer to the point where criminals still need something real.

They need an account capable of receiving money. They need an exchange, payment service, mule network, or bank that lets them move the proceeds. They need victims to authorize transfers and institutions to process them. Those parts are harder to automate away.

Interpol's 2026 global fraud assessment describes more than 1,500 transnational fraud cases involving $1.1 billion in reported losses and highlights international efforts to stop payments after fraud has been detected.

CryptoSlate has previously covered how AI is increasing the scale of crypto scams. The next stage is about what happens when convincing fraud becomes abundant.

Familiar voices can no longer carry as much trust when voices can be generated for fractions of a cent. Video calls also become weaker evidence when faces can be synthesized, and a long, thoughtful conversation means less when maintaining it costs almost nothing.

That leaves ordinary people doing more verification while criminals do less manual work.

The great promise of automation was that software would take tedious tasks away from humans. Fraud has found a particularly irritating implementation of that idea: the machine handles the impersonation, while the person receiving the message has to investigate whether anybody involved is real.

AI may eventually reduce the number of humans required to run a scam. Unfortunately, it also makes running another scam much cheaper.