Tech Race Summit participants warn of obstacles to turning AI pilots into profits [Sponsored]

At the first Tech Race Summit, the room was full of people who know the AI bill comes due long after the demo looks good. Georgios Chatzimilioudis of Oracle opened with a simple show of hands: everyone expected to use more AI in three years, and many expected to use more than one model. But when he asked who knew which model, chip, platform or framework would matter most, the confidence vanished. That was the point. The hard part is no longer access to AI. It is turning it into something a business can actually run, govern and defend.

Sergey Khurson of AWS put a number on the gap. He said 88 per cent of organisations use AI in at least one part of their business, but only 8 per cent have reached advanced maturity, where AI is running in production across the company rather than sitting in scattered pilots. The Tech Race Summit audience, drawn from fintech, e-commerce, iGaming and other high-load sectors, heard a familiar story: AI is moving into engineering, fraud prevention, customer support, compliance and infrastructure, but the organisations around it are still catching up. Buying a model is the easy bit. Feeding it the right data, limiting what it can do, and proving it was worth the spend is where the work starts.

That same problem shows up in the graveyard of unfinished experiments. Viktor Vedmich of AWS described “zombie projects” that keep eating resources after their purpose has gone fuzzy. His answer was fewer projects, shared platforms and hard definitions of success before anything starts. Management needs to decide in advance when a pilot gets scaled and when it gets killed. If a programme is sitting under a CEO or COO, with targets tied to revenue and margins, there is less room for it to drift forever.

Data quality came up as the other quiet blocker. Chatzimilioudis cited a Gartner report saying 63 per cent of organisations lack the data management practices AI needs. His fix was basic but not easy: assign ownership, track where data came from and control access. For sensitive financial or customer data, he argued for bringing AI to the data instead of copying data everywhere. Cloudflare said it followed that logic by building a shared context layer, then Cloudflare OS, so employees could use it from day one. A new marketing hire, for example, could draft a blog post in the company’s voice because the tool already carried the guidelines, facts and approval process.

Cost and accountability hovered over everything else. Cloudflare’s Michael Tremante said not every request should go to the most powerful model, and the company routes work between models and runs open-source alternatives on its own infrastructure. A CTO panel at the summit split over whether the productivity gains justify current spending. Andrey Doronichev of Bioptic added a quality check that felt refreshingly old-school: human pharma experts keep a bank of questions and answers to test agents and define what good output looks like as the system is built. The crowd’s verdict was plain enough. AI can help, but for now the business still owns the result.