Tech Race Summit Leaves AI Pilots Stuck at 8%

About 1,500 people at Warsaw’s Tech Race Summit on 10 September 2026 dropped every hand when asked which AI stack would matter in three years. Georgios Chatzimilioudis of Oracle had just watched those same hands rise for more AI, then stay up for more than one model.

The inaugural meeting, organised by SOFTSWISS, put AWS, Cloudflare, Tesla and licensed online casino operators on the same stages. Buying a model is easy. Turning an AI pilot into profit was the part that stalled the room.

Every Hand Went Up, Then Every Hand Came Down

Chatzimilioudis, a principal account cloud engineer at Oracle, used the show of hands as a map of the year. Firms expect to spend more. They expect to juggle more than one model. They do not expect to know which model, chip, platform or framework will still matter in three years.

That uncertainty is not a side issue for high-load trades. Latency, fraud, payouts and live odds already punish slow systems, and a wrong bet on a stack is expensive to unwind. The sold-out debut brought around 1,500 professionals to Warsaw on 10 September 2026, with more than 40 speakers across the Vision, Solution and Experiment stages.

Javier Verdura, Tesla’s global director of product design, talked about teams that own outcomes. Dhiraj Mukherjee, a co-founder of Shazam, talked about shifts that look impossible until the pipes exist. The AI sessions kept returning to a smaller, greyer problem: pilots that never leave the lab.

iGaming’s Digital Score Does Not Cover AI

Later in September, SOFTSWISS published the fifth edition of its iGaming trends report with WorldGaming, drawing on more than 500 industry professionals, 65 experts and more than 480,000 media headlines. The report gave the trade a Digital Maturity Index of 4.67 out of 5, the highest among the eight industries assessed, on AI adoption, mobile-first systems, digital customer acquisition, payments innovation, RegTech and data-driven decisions.

Sergey Khurson, who leads betting and iGaming technology work at AWS, put a colder pair of figures on the same stage. He said 88% of organisations use AI in at least one part of the business, and only 8% have reached advanced maturity, where AI runs in production across the firm rather than in isolated pilots.

THE ADOPTION NUMBERS IN THE ROOM

The index can sit beside that 8% figure because it rewards having AI somewhere. Khurson was counting whether it runs across the business. Fifty-two percent of professionals surveyed for the report named AI the leading macro trend for 2027, which is a spend forecast, not a production one.

Zombie Projects Survive Because Nobody Will Kill Them

Viktor Vedmich of AWS had a name for the experiments that outlive their purpose. He described zombie projects that continue indefinitely, burning staff time and model bills with no decision to expand them or stop them. In a market where nobody will name next year’s stack, a living pilot is a hedge. It is also how budgets disappear.

Vedmich wanted fewer projects, chosen by the value they could create and the effort they would take, and shared platforms so teams stop rebuilding the same scaffolding. Before work starts, he said, a firm should define what success looks like and when management will decide on a wider rollout. Parking the programme under the CEO or COO, with targets tied to revenue and margins, makes that go or stop call harder to dodge. It also forces a blunt question: is the work being automated worth doing at all?

HOW VEDMICH WOULD KILL A ZOMBIE PILOT

• A success line: Write what good looks like before anyone books the model.
• A decision date: Set when leaders will expand the work or shut it down.
• A named owner: Put the programme under the CEO or COO, with revenue and margin targets.
• A shared platform: Run fewer projects on common tools instead of a new stack per team.

One fix discussed in Warsaw was to bound the work around specific AI tools so projects cannot creep. Without that fence, a chatbot trial becomes a platform, then a cost centre, then a line item nobody wants to own.

Why AI Pilots Stall After the Demo?

A model that cannot find or read the company’s own files is a demo, not a product. Chatzimilioudis pointed to Gartner for the scale of that miss, and his remedy was dull on purpose: give data owners, record where it came from, and set rules for who can touch it. For firms sitting on payment trails and player files, he argued for bringing the model to the data rather than copying those stores again and again.

WHAT GARTNER FOUND ON AI-READY DATA

• The survey: In July 2024 Gartner asked 1,203 data-management leaders about practices fit for AI.
• The gap: 63% of organisations either do not have those practices or are unsure they have them.
• The forecast: Gartner says firms will abandon 60% of AI projects unsupported by AI-ready data through 2026.
• The analyst: Roxane Edjlali, a Gartner senior director analyst, treated AI-ready data as a practice that has to keep changing with each new use, not a one-off clean-up.

Traditional data work is slow, rigid and poorly documented, Edjlali said, with files sitting in silos and little metadata to judge whether they are fit for a model. If the data is broken, the data is not ready for AI. That is a boring sentence, and it explains more failed pilots than a missing chip.

Cloudflare Built an OS After Setup Became the Blocker

Cloudflare’s field CTO, Michael Tremante, described a path that started with staff, not with a model catalogue. The company set up an email address where people could ask an agent team for help, then built a shared context layer in a GitHub repository so answers and facts could be reused. Engineers next received the open-source tool OpenCode, preloaded with enterprise model accounts. Non-technical staff still tripped on setup and login.

Cloudflare OS was the answer to that second wall. Tremante said a new marketing hire could write a blog post in the company’s voice on day one, because the tool already held the guidelines, the key facts and the approval path. In May 2026 every Cloudflare employee got the first version. Thousands now use it daily, many of them outside engineering, to draft documents, automate repeat jobs and build small internal apps.

In May of this year, we gave every person at Cloudflare access to the first version of Cloudflare OS. Thousands of people across every function, many of them outside of engineering, use it every day to create documents and slides, automate repeatable tasks, and build small apps to visualize data and help them do their work.

Phillip Jones and Dan Carter, Cloudflare Blog

On 5 August 2026 the company open-sourced its internal agent workspace, so other firms can deploy it on their own accounts and plug in their own systems. Agents start with no access. Credentials stay behind gatekeepers. Policy follows whatever the agent has already seen, so a private table cannot leak through a shared dashboard.

Cost sat under that design. Tremante questioned sending every request to the most powerful, and most expensive, model when a smaller one can do the job. Cloudflare routes work between models and runs open-source alternatives on its own kit. The point is not a cheaper chatbot. It is a bill a finance team can read.

Three Months From Idea to a Live Fraud Model

The room was not only a catalogue of stalled work. A few teams had already forced a pilot through to something a customer or an on-call engineer could touch, and those jobs were narrow enough to judge.

THE PRODUCTION WINS THAT DID GET OUT

• BGaming fraud: CTO Alexander Krivetsky walked through an AI fraud system that went from idea to production in 3 months, covering data prep, training, integration and the last mile into live traffic.
• Maincard ops: CTO Igor Borzunov showed an autonomous AI operations agent that cut incident reaction time to 30 seconds, down from 3 a.m. pager cycles.
• SOFTSWISS Orchestrator: Software architect Illia Slepau described an internal layer that coordinates AI agents through the software lifecycle while humans keep validation and approval.

Those cases share a shape. The job is specific, the test is visible, and a person still signs the change. They do not prove that a company can sprinkle a frontier model across finance, support, legal and marketing and call it maturity. They prove that a bounded problem, with a kill date, can leave the lab.

The Human Still Signs the Output

The CTO panel split on whether today’s bills are worth it. That argument will not be settled by another pilot.

WHERE THE CTO PANEL SPLIT

• The sceptics: Modest productivity gains will be hard to keep funding once the invoices stack up.
• The bulls: The spend is small next to what the best engineers can now produce with the same tools.

Accountability is the piece that survives either view. Output from a model is still hard to verify, and the room’s working rule was that the business remains on the hook. Andrey Doronichev, a former Google executive who now runs the drug-discovery firm Optic, described a bank of questions and answers that human pharma experts use to test agents as they are built, so “good” is defined before the system ships. Delegates still wanted a person in the loop who can say the output is good enough, explain how it was produced, and know what it cost.

Denis Romanovskiy, SOFTSWISS’s chief AI officer, has been saying the same thing inside the company that staged the meeting. He said individual staff are 20 to 50% more productive with AI, and that a jump toward 200% means handing whole workflows to agents and teaching people to supervise them. He also said the work is still early.

We are still in early days of AI. At this moment, we’re not there yet.

Denis Romanovskiy, Chief AI Officer, SOFTSWISS

Tech Race Summit will return to Warsaw in 2027, and early-bird tickets are already on sale. The next room will raise its hands for more AI again. The useful document to bring is not a model list. It is the date a failed pilot dies.