Australia's Apate Deploys Over 350,000 AI Agents to Counter Scam Calls

01: Turning Defense into Offense

The Scammer vs. Digital Human War

This Australian AI startup named Apate doesn't do any large-model R&D; instead, it has deployed over 350,000 AI Agents specifically tasked with playing all sorts of potential scam victims.

The more time and energy scammers are tied up, the less time they have in reality to spend on real victims. This anti-fraud logic is rather unconventional, but on closer thought it actually makes sense.

These Agents not only disguise themselves as elderly men and women to proactively answer scam calls, but also cooperate with scammers' instructions to click phishing links or reply to phishing text messages, and can even infiltrate scam group chats—embodying the idea that "the advanced hunter often appears in the form of prey."

So how effective is it really?

Judging from the many countermeasure cases posted on Apate's official website, it's honestly a bit hard to keep a straight face.

There are also quite a few Agents with Indian accents

It's the same abroad—many telecom fraud routines start with something like broadband issues, incorrect phone bills, or administrative penalties, and the follow-up is basically always about needing to handle it remotely.

After a flurry of maneuvers, the scammer waits for the hooked fish to download remote-control software disguised as normal software or links.

But when it comes to this crucial step that could successfully extract money, the elderly start flexibly losing IQ points and dropping the ball left and right.

They can't figure out the browser, the internet is too slow to load, and they suddenly need to go pick up their grandchildren—in short, they just can't sort it out right now. No matter whether you're Ma Dongwhat, what Dongmei, Ma whatmei—everyone just wait wait wait wait wait........

In the end, usually the scammer themselves can't take it anymore and hangs up, while the elderly are still hopping back and forth at that same step.

According to the company's latest business data, this AI system has now completed over 2.5 million conversations, occupied scammers for 17 years of time, and through close conversations with scammers, successfully extracted over 250,000 pieces of useful intelligence (such as scammers' account addresses, crypto wallet addresses, etc.).

Both voice and text can stall

This isn't clickbait—think about it from another angle: this time was originally going to be used to scam real people.

What's even more bizarre is that this track record actually came from an utterly ordinary scam call......

The Anti-Fraud Mastermind Team

A few years ago, Dali Kaafar, a professor at Macquarie University in Australia, received a call from an unknown number during a family picnic—clearly a scam.

But this guy didn't hang up; instead, he chatted with the other party for 44 minutes, during which he had the whole family laughing.

As expected of a researcher, his stamina was impressive—but he didn't chat for nothing. Right after hanging up, he immediately thought of a question: what if this kind of thing could be handed over to machines?

That idea later became the company mentioned above in 2024, whose name Apate is taken from the Greek goddess of deception.

Currently, the partner Australian telecom provider TPG first uses its own system to identify suspicious calls, and instead of blocking them directly, transfers them to Apate's AI to answer. In addition, numbers that ordinary people instantly hang up on, block, and report are also marked as potential transfer targets.

It is reported that during the pilot alone, TPG transferred over 280,000 scam calls, among which more than 20,000 impersonated organizations were identified. The longest call lasted 54 minutes, and the scam traffic identification accuracy is claimed to reach 98%.

Afterward, TPG's cumulative transferred calls approached 590,000, and the company estimates that about $13 million in losses were avoided.

Beyond transfers, Apate also scatters a batch of "bait numbers" online, specifically waiting for scammers to bite.

A massive identification and scanning network

Of course, besides answering calls, these AI Agents also infiltrate various investment groups and scam groups on social platforms. It is said that they handle hundreds of thousands of calls and chats daily on platforms such as WhatsApp and Telegram.

But while the business can be built smoothly, can this thing make money?

Yes, and possibly quite a lot. It looks like public welfare, but in fact it hits many industry pain points.

Has already launched cooperation with multiple Australian financial institutions

According to investor Investible's description, Apate provides scam intelligence to financial institutions, telecom companies, and government departments through its software platform and real-time alerts, and has established commercial partnerships with institutions including Commonwealth Bank of Australia and TPG Telecom.

Among them, the Commonwealth Bank of Australia (CBA) is said to have started a pilot with both parties in 2024 and announced an expansion of cooperation in June 2025. According to CBA's own public data, customer scam losses have dropped 76% compared with the peak.

Another Australian financial institution also stated that after enabling Apate's outbound-call robots, it successfully dismantled 22 scam gang operations.

The Scam Prevention Framework

Another major backdrop is that the policy costs brought by scam activities are continuously rising.

In February 2025, Australia passed the Scam Prevention Framework (SPF), stipulating that banks, telecom providers, and digital platforms can be fined up to AUD 50 million if they fail to take "reasonable measures" to report, block, and respond to scams; victims also have a clear claims channel when institutions violate the rules.

So this account is really easy to calculate.

Large banks already invest astronomical sums each year in anti-fraud and cybersecurity, and now they also face the risk of fines and potential claims, so naturally they have to find a way.

Moreover, these AI Agents that don't fleece old men but fleece scammers have brought a qualitative change in the efficiency and scale of anti-fraud intelligence gathering.

Apate co-founder Joffe said in an interview that generally speaking, when a telecom provider blocks a number, scammers just switch to another one; even if a bank is monitoring the transaction in between, by the time abnormalities are detected, the scammer has long since contacted the customer first.

A full set of anti-fraud defenses

But in the scam chain, there is one link that is not so easy to swap out, and that is "collecting the money."

Numbers, scripts, and identities can be changed at any time, but bank accounts require opening an account, verifying identity, and someone responsible for receiving and transferring funds; once flagged by risk controls and frozen, new ones must be found.

This has spawned an underground business: money mule accounts. As the name suggests, someone specifically provides bank accounts to collect and transfer money for scam syndicates, then takes a commission.

Also called mule accounts

In July this year, Apate's bots discovered online a trading marketplace specifically recruiting real-name Indian bank accounts, with intermediaries offering commissions calculated as a percentage of scam fund flows, settled in USDT.

At this point, attack is often the best defense, so Apate's approach goes straight at this link.

Banks and other clients can first assign "tasks" to AI Agents, such as obtaining these bank account numbers or crypto wallet addresses. After the scammers who take the bait are strung along by the AI, they will think the other side is a customer about to transfer money, and at that point they will certainly reveal their payment collection method.

In the past, banks often had to wait for victims to discover they had been scammed and report it themselves before they could trace the flow of funds, by which time the money had long been transferred away and basically could not be recovered. Now, if the account numbers and wallet addresses are known in advance, there is a chance to intercept midway.

Capital has also seen this path. In August this year, Apate completed an $8.15 million seed round, with follow-up plans including opening an office in London and expanding its US business.

02: One Body, Two Lives

First, on the scam side: in the past, numbers could be bought in bulk, scripts could be copied and pasted, and even many voices could be cloned.

But to actually get the money, the human operations of building trust, emotional manipulation, and step-by-step enticement between people are still far beyond what machines can replace.

Now with AI, through massive learning and distillation, many Agents specifically used for scams can likewise possess multiple personalities, just like the scams we often talk about: it's not that no one was fooled, but that the right one for you hasn't come along.

Shutting them down is even more impossible; even if a scammer is discovered, it does not stop them from immediately dialing the next number.

So Apate's approach can be called innovative: if you can't block them all, trap them all, and even completely break the scammers' composure, because in the end it is impossible to tell which are real people and which are Agents. Whether it is a human scammer or an AI scammer, there are costs behind it.

Time and cost are massively wasted, and it becomes increasingly impossible to tell where the target is. It is simply devastating.

Now rolled out to version 2.0

Actually, as early as 2017, New Zealand's independent online safety organization Netsafe released an AI email bot.

As long as users forwarded suspicious emails to it, the AI would take over and chat with the scammer, similar in idea to the later Apate. It could play multiple identities and imitate human humor and grammatical errors to fool scammers. It would also keep asking questions and telling stories, dragging the conversation out until the scammer stopped replying.

Within days of launch, it sent more than 26,000 emails, consuming more than three months of scammers' time.

By 2023, with the emergence of GPT, individual developer Roger Anderson built an anti-fraud subscription service earning $250,000 a year, with 10,000 subscribers.

Using ChatGPT plus voice cloning, it kept phone scammers on the line for long periods. Among the personas were "confused old man Whitey Whitebeard" and "Salty Sally, the mom always interrupted by her kids." It could also recognize recordings from auto-dialers and press the "transfer key" to prevent the other party from hanging up.

What is chilling is that when the founder first told GPT to "waste time," it actually replied: "As a helpful AI, wasting time is not in my nature." It took carefully designed prompts to get around that.

By 2024, UK carrier Virgin Media O2 also launched an AI grandma, who consumed scammers by telling stories about fictional cats and grandchildren. It chatted with scammers more than a thousand times in total, with the longest session about 40 minutes.

Including Apate, when all AI Agents first started, they had only 120 persona sets, far from enough to "handle" that many scammers in practice.

As technology iterated and developed, its AI Agent count expanded to 197,000 by August this year, and has now reached 350,000+, still growing rapidly.

Each persona has a different age, gender, accent, language, and personality. Even the pauses and hesitation when speaking, and the subconscious "hmm," "ah," and "oh" when thinking, were specially designed.

Technically, Apate co-founder Joffe told TechCrunch that they use both general large models and their own fine-tuned models, with training data coming from real conversations between bots and scammers, as well as research Kaafar did at Macquarie University.

A 2023 Macquarie University report mentioned that at the time calls averaged 5 minutes, with a goal of 40 minutes. Later the company advertised a maximum of 50 minutes. According to the founder's latest disclosure to WIRED, calls with scammers now often exceed two hours.

For scammers, what is the most beloved and most hated kind of customer? One who looks about to pay but always falls just one step short.

So some scammers, as the chat drags on, break down first and start cursing.

The most abstract and satisfying part comes next: Apate simply wrote scammers cursing into its KPI.

Founder Kaafar joked that they may be the only company in the world that uses "the number of times scammers swear at the bot" as a KPI.

Joffe said that every time a scammer curses, their team celebrates. It is a quantitative indicator of how much the scammer has been tormented, and especially when they swear, the sense of achievement simply explodes.

Looking at the middle ground, one can actually also spot a huge industry change that is quietly making a fortune.

From scams to customer service

After all, what the internet lacks most now is content and data. After collecting large amounts of verbatim transcripts of ordinary creators speaking, letting large models "learn" a batch of personas can give each "person" their own personality, catchphrases, and even values.

Next, companies can let massive numbers of AI personalities converse with their customers, from pre-sales to after-sales, from young mothers to elderly men, all finding the right track and providing precise service. In this process, human customer service has not completely disappeared; instead, when human intervention is needed, a person chooses a reply from the A, B, and C options given by the AI.

Plenty of hardcore tricks, but in the end it is still customer service

From a certain angle, it seems humans have not completely left, but humans have instead become something like an appendage.

Looking at broader figures, in 2026 China's intelligent customer service market reached 28.5 billion yuan, and the penetration rate of large models in the customer service field rose from 38% in 2024 to 72%. It can be said that AI customer service is one of the fastest to land and most widespread application scenarios.

This thing is no different from the real-life emotional livestreamers of the past, shouting every day for aunties to help them, and then using livestream selling to scam money.

It cannot help but make people worry: how are we going to get by in the future?

"Is there still a human here?"

This article comes from the WeChat public account "Coolplay Lab," author: Coolplay Lab, published with authorization from 36Kr.