AI Bubble: It Settles Where Precision Isn't Needed

Pavel calls his path a crooked one: a small town near Moscow, a Moscow school from the eighth grade, and a first degree as a historian and Arabist at the Institute of Asian and African Studies at Moscow State University. History, he says, is the same work with large volumes of data: students were taught to work out from a corpus in an unfamiliar eastern language whether it was a chronicle, correspondence or diaries, using basic statistics and NLP methods, all of it in VBA. History would not pay, and after his master's he changed direction sharply: the University of Geneva, where education is free even for foreigners at 500 francs a semester, financial analytics taught in French, and calculus rescued by a seminar tutor from Rostov who explained half of it in Russian. The Swiss visa obliged graduates to leave, so Pavel returned, studied supply chain analytics at the Higher School of Economics and landed at Renault.

At Renault he worked on supply chain analytics and optimisation: the Avtoframos plant, two AvtoVAZ sites, Nissan near St Petersburg and, he thinks, more than a hundred local suppliers; with a colleague he optimised returnable packaging and saved the company serious money through accurate forecasts. The colleague left for Ernst & Young, Pavel followed, and consulting began: projects from Yakutsk and Mirny to the Murmansk region, twelve-hour shifts in protective gear trailing an operator, collecting data by hand, then models, slides and a dressing-down from the partner in Moscow at night. That was when he learned Python, took a dozen courses and got drawn into the Open Data Science community — in his words the best place for networking he knows, where someone from one company would happily discuss approaches with a competitor's data scientist. From Ernst & Young he moved to Accenture, and there wrote the Habr article on decision support systems that S7 read before inviting him to build one.

At S7 Pavel was a hands-on product owner: the decision support system stalled — data sat inside separate departments and every dataset had to be begged for — but the team did test smart contracts for aircraft refuelling and he optimised the layout of in-flight catering, taking the idea from an Air Canada paper. He presented that case at a Dubai meetup, and afterwards the CTO of the Finnish firm Reaktor's local office came up to him: three of his job offers, he reckons, came from being publicly visible, and the catering case was later repeated for Emirates. Reaktor turned out to be radically egalitarian — Pavel coded on a project alongside the global CEO and left him code review comments, and whether to take a client on was decided together — while juniors were not hired at all: they needed people who could both talk to the client and write the code.

He describes the Middle Eastern market as slow: a year can easily pass between a verbal agreement and the first money, "from Ramadan to Ramadan", and how a solution looks often matters more than how it works. One client's CEO saw the finished mobile application at a demo, said he had expected a robot, refused to pay, and never got the solution; a public transport authority tender went to McKinsey, several times more expensive and offering a remote team. Then demand for AI collapsed: a chief AI officer at a rich holding might not tell a regression task from computer vision, and after the owner of a large local marketplace vanished with the investments, a wave of protecting investments, layoffs and falling salaries followed, while Reaktor had put everything on Dubai's tourism ministry a month before the borders closed, and shut the office.

Next came Europe and Beat, a Greek ride-hailing service with its main markets in Latin America, where Pavel led analytics for fraud prevention and safety. The keenest passengers of the branded Teslas in Mexico City were drug dealers: a conspicuous car, money to spend, several stops visible on camera; a three-stop limit per ride thinned out the dealers, and demand fell with them. Drivers were robbed every month, in one case a driver shot a robber inside the car, and the team tried to train computer vision to spot an emergency when the SOS button is out of reach; a bot farm outside Lima was found as a blob of ride requests from one house. He breaks fraud into a pyramid — desire, opportunity and the sense of impunity — and a platform mostly affects opportunity. The company never made the turn from growth to efficiency: the Teslas sat on the balance sheet, several CEOs came and went, one announcing that he ran the company like a rugby team, and it ended in bankruptcy. EPAM saved the three months he had to find a visa, and a year later he chose between TomTom, whose product he never found an audience for, and Microsoft.

The Industrial Solutions Engineering vertical takes no money from clients: its metric is growth in Azure consumption, and it builds what the standard product line does not cover. At a Scandinavian brewery with an OEE far below 50% two shifts simply did not talk to each other: the unions allow not a minute of extra work, and instead of a handover log he was asked for speech-to-text with an automatic summary. He expects the GenAI bubble to behave like the Dubai one, where people asked for computer vision in places a regular expression would do: the technology will settle where full accuracy is not required. Alexander adds that GenAI cannot simply be spread over a developer platform — the effect has to be gathered job by job and measured with tool telemetry. A summer reorg moved Pavel from management to the IC track, and a September one made him a Data Science Champion under the region's vice-president and a resilience buffer for multidisciplinary team leads. The term forward deployed engineer, popularised by Palantir, he translates simply: a jack of all trades who builds it, counts the ROI and presents it. The conversation ran nearly three hours and ended with a promise of a sequel.