Most companies point AI agents at shipping software faster. Australian startup Apate.ai points them at scammers. TechCrunch reported on October 14, 2026 that the company runs thousands of agents posing as ordinary victims, with different ages, genders, accents and languages, to keep fraudsters engaged long enough to reveal bank accounts, crypto wallets, phone numbers, malicious sites and the brands they impersonate.
The idea is old scam-baiting, automated and sold as a service. Startup Daily's report on the seed round likewise names Commonwealth Bank as a client and describes 2.5 million autonomous conversations with threat actors, though it does not independently confirm those figures. That makes it a useful case study for what agents are good at (endless patience, scale, persona variety) and for which claims about them need caveats. Apate is a finalist in the Startup Battlefield 200 at TechCrunch Disrupt in San Francisco, which runs through October 15.
TL;DR
| Question | Answer | Source |
|---|---|---|
| What is it? | Fleet of AI agents that pose as scam victims on calls, messages, email and the web | TechCrunch |
| Who built it? | Founded in 2024 from Macquarie University research; CEO Dali Kaafar | TechCrunch |
| What do customers get? | Scam intelligence in an insights portal; they can assign missions such as getting a bank account number | TechCrunch |
| Scale claimed | 2.5M conversations, 17 years of scammer time wasted | Company, via TechCrunch |
| Customers named | Commonwealth Bank of Australia, TPG Telecom, others | Company, via TechCrunch |
| Funding | $8.15M seed in August; just under $10M total | TechCrunch |
| Independent proof | None found | Our check |
How the agents work
Co-founder Brad Joffe told TechCrunch the company uses both frontier models and its own fine-tuned models, trained on conversations its agents had with scammers plus Kaafar's Macquarie research on using machine learning and language processing to understand scam calls. Voice agents answer calls, text agents work messages and email, and web agents search forums, investment groups and social platforms for scam activity, then enter conversations under a fabricated identity.
Joffe's pitch is about timing: "Most of the industry works after the fact. Telcos block numbers, but scammers just get new ones. Banks watch transactions, but by then the scammer has already reached the customer." Apate instead "goes to the scammer directly." Its output is intelligence, not blocking by itself: customers feed it into their own anti-fraud systems.
Origin story
Kaafar got a scam call at a family picnic and played along for 44 minutes. The scammer was not calling a real victim during that time, which raised the question of whether the stalling could be automated and made useful. That became research at Macquarie University and then the company.
The numbers, and who gave them
Everything below is attributed to Apate via TechCrunch.
- 17 years of scammer time wasted across 2.5 million conversations.
- Tens of thousands of scammers contacted every day, per Joffe.
- TPG Telecom diverted nearly 590,000 calls to Apate's agents.
- Commonwealth Bank saw a 76% drop in publicly reported scam losses.
- An unnamed Australian bank used outbound agents to disrupt 22 scam campaigns.
Treat these as vendor claims. The 76% figure is the one to question: a drop in publicly reported losses over some period cannot by itself be attributed to one vendor, since banks run many fraud controls at once. Neither the period nor the method is given in the article.
Why this matters: AI is on both sides
TechCrunch cites FBI figures of more than 22,000 AI-related crime complaints in 2025, about $893 million in losses, with over $632 million from investment scams where fraudsters use AI to run conversations and deepfake celebrities and executives. For the wider picture of how large the AI scam economy is claimed to be, see our check of the $442 billion AI scam economy claim. On the defensive side, Google has put scam detection on devices in Google Messages enhanced protection, and DeepMind has worked on on-device scam detection with Gemma in Singapore.
Apate's approach is different in kind: offense against the scam operation's resources rather than filtering at the victim's end. Related deepfake-driven fraud is covered in the $25 million video call scam.
From hobby to product: how scam-baiting changed
TechCrunch notes that people have long made a hobby of stringing scammers along, sometimes comically, such as getting them to read entire books aloud or accept absurd demands. The hobby had two limits: it needed a human with time and patience, and what the baiter learned rarely reached the institutions that could act on it.
Automation removes the first limit. An agent does not tire, can run thousands of conversations in parallel, and can switch language or accent per call. It also changes the second limit, because the output is structured: the same fields (account numbers, wallets, numbers, impersonated brands) appear in every conversation and can be fed into a portal. That is the product Apate sells.
How it compares with other anti-scam approaches
| Approach | Where it acts | Weakness |
|---|---|---|
| Number blocking by telcos | After a number is known to be bad | Scammers rotate numbers, as Joffe notes |
| Transaction monitoring by banks | When money is already moving | The victim has already been reached |
| On-device scam detection (Google, DeepMind) | On the victim's phone, in the moment | Depends on the victim's device and settings |
| Public awareness campaigns | Before contact | Hard to measure and uneven reach |
| Apate-style baiting | On the scammer's side, before real victims | Needs scammers to engage; vendor claims unverified |
These are complements, not substitutes. The honest framing of Apate is as a source of upstream intelligence that makes the other layers better informed, not as a replacement for them.
What people are asking
Is this legal? The article does not address it, and we cannot give legal advice. Recording and impersonation rules differ by country, and Apate says it works with banks, telcos and government agencies, which suggests those customers have reviewed the arrangement. That is an inference, not a reported fact.
Do the scammers lose anything real? Time is the cost imposed. Apate's stat of 17 years of wasted scammer time is a way of expressing that, though scam operations that run many calls at once may feel it less than the number suggests. The more concrete harm to them is exposure of accounts and wallets, if customers act on the intelligence.
Could criminals do the same to defenders? Scammers already use AI to run conversations and clone voices, according to the FBI figures TechCrunch cites. An arms race in which both sides use agents is plausible, and detection of bot counterparts becomes a feature on both sides. We have no evidence yet of scammers specifically hunting for Apate agents.
Why is this a startup and not a feature at a big bank? The article suggests a reason in its customer list: a shared fleet that talks to tens of thousands of scammers a day pools intelligence across a bank, a telco and government agencies, which a single institution would struggle to match alone. Again, that is our reading.
What builders can take from it
Apate is an agent product with a clear, narrow job, a measurable output and a buyer who already has a budget for the problem. Three things stand out for anyone building agent products.
First, the output is data, not actions in the world. The agents converse, but customers decide what to do with the findings. That limits the damage a mistaken agent can do, a design choice worth copying.
Second, persona variety is the core technical asset. Different ages, genders, accents and languages are what keep a conversation going, and Joffe says the company trains on its own conversation data on top of frontier models. That is a pattern we see in other vertical agent companies: a general model for language plus proprietary data from the deployment.
Third, the proof problem is real. The company's headline stats are impressive and also unverifiable from outside. Vertical agent startups selling to risk-averse buyers will be asked for independent results, and the ones that publish methodology will stand out.
Open questions and limits
- Evidence. No audit or customer quote backs the headline numbers.
- Ethics and law. Agents impersonating people raise questions about recording consent, entrapment-like conduct and cross-border rules. TechCrunch does not discuss them; we flag them as questions, not findings.
- Scammers adapt. If scam operations start detecting bot victims, the time-wasting value drops. Apate's reported answer is varied personas and its own trained models, but no data on detection rates is public.
- Data handling. Agents collect bank accounts, wallets and phone numbers of suspected criminals; how that is verified before customers act on it is not described.
What you can do
For individuals: do not engage scammers yourself, never share codes, and verify callers through the number on your card or official app. For teams building or deploying agents, the same lessons apply to your own agents: an agent that talks to untrusted parties needs limits on what it can reveal and do. AgentBeam is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.
Related reading
- The $442 billion AI scam economy claim, checked
- Google Messages scam detection with Gemini
- DeepMind and Singapore on-device scam detection
- Deepfake video call fraud and verification
- Bill Gates on AI cyberattacks
- AI agents not going rogue: containment basics
Details are as reported by TechCrunch on October 14, 2026; company figures are unverified.