There was supposed to be AI, there were people from the Philippines. Investors must watch out for fraudulent start-ups
NATE is the most famous example in recent months. This fintech boasted that he could finalize the purchase in any online store “without any human intervention”. In April 2025, however, the Department of Justice accused the founder, Albert Saniger, Fr. Conscious misleading investors.
The actual level of application automation was – according to the indictment – round zero percent,, And all purchases were finalized by hundreds of contractors from Manila. A company that collected over $ 50 million (The last round of Series is 38 million in 2021), sold out assets and practically reset the value of fund shares.
The mechanism looked similar in Presto Automation, a “voice AI” supplier for the drive -thru. Reports to the American SEC have revealed that over 70 percent Employees from the Philippines served remotely, although marketing materials promised restaurants to reduce costs due to full automation. In 2023, the company became interested in the securities supervision authority, examining the discrepancies between communication and real technology.
In turn, Evenup – a real unicorn, priced in October 2024 above a billion dollars – declared that his algorithms can read medical documentation and create claim letters on their own in matters of health damage. Eight former employees, however, told African Business Insider that in practice The managers told them to completely avoid the tool because they generated too many mistakes: It was able to skip real injuries, “hallucin” non -existent medical visits or confusing the leg with the shoulder.
The company admitted that it uses the hybrid team: people and algorithms. At the same time, it was people for hours that improved the results generated by the AI model, often working at night and in holidays. The role of AI was small.
Read also: A new way for bots. Protection against AI more effective than copyright and courts
Human work cheaper than the algorithm
Why are such companies at all? First, labor costs in the most popular BPO centers are extremely low. In 2025, the Filipino labor law allows a minimum rate of 573-610 PHP per day, i.e. only $ 5.70— 11 dollars. – which brings a working hour even below two dollars (approx. 7.50 PLN).
Second, multiplying funds specialized in AI They compete for exclusive transactions and often finance projects at the stage of promisebefore a working model is created.
Thirdly – from the perspective of the founders – such a “hand prosthesis” can be the simplest path to a rapid increase in the number of users, sometimes with the intention that the algorithms take the tasks later. Although it happens that it never happens.
The fraud scheme towards investors is usually repetitive. The product demo presented at conferences works on narrowly trained scenarios, and the moment of delay resulting from sending tasks to a call center is translated by a network delay. In Pitch Deck, high margins “after full implementation” are emphasized, while The real operating cost increases with the number of orders, because each of them serves man.
Settlements with subcontractors are hidden in the “General costs” column, and the investor who does not carry out the technical due diligence only learns the truth when the company must recapitalize the unnaturally expanded operation department.
Check also: Protection against quantum computers will not be easy. AI attacks are “Pikuś”
Investors must be more careful
So how do you verify start-ups not to be fooled? From the investor's perspective, it is necessary to carry out a multi -layered inspection, which is attended by independent Machine Learning engineers, industry experts and lawyers.
In practice, this means demand full access to system logs and automation statistics, observing live data flow, analysis of the cost structure for outsourced jobs and conversations not only with the C -Level team, but also with first -line employees. AI models should also be tested on a set of random cases that the team did not know before, and the engineer should be able to indicate where the most important decisions are made in the code. Recommendations coincide with the best practices described in industry guides on AI -Powered Due Diligence, which emphasize the analysis of the source code, Machine Learning processes and data version control.
The user, although he has less audit, is also not defenseless. The answer is simple tests: starting the application in offline mode, checking whether information about “human review”, comparing the system response time and the quality analysis of the results at unusual queries appears in the regulations. If the service claims that it does not send data out of the device, and at the same time each query lasts several seconds, there is a probability that traffic returns to outsourcing centers. It is also worth noting whether the company publishes bright records of the model effectiveness and updates them in time. Lack of such data can be a warning signal.
Not all young AI companies that benefit from the work of people deserve a patch of fraudsters. Many companies use an approach Human -in -The -Loop Consciously to train the model and gradually reduce human participation. The problem begins where this participation is hidden, and investors and customers have no chance to know the cost and ethical truth.
When the number of funds interested in AI, NATE, Presto or Evenup is growing on the Vistula, show that healthy skepticism and deep verification of technology become not so much an advantage but necessity. Otherwise, even the most innovative slogans and groundbreaking technologies can mean – literally —Call -perceter on the other side of the globe.
Author: Grzegorz Kubera, Business Insider Polska journalist
