Everyone wants to talk about the 95% that fail. Almost nobody wants to talk about what the 5% did first because it isn't a model, a vendor, or a strategy offsite. It's data nobody wanted to touch.
🎧 Hit play on Alex Luketa — Why 95% of AI Projects Fail, and What the 5% Do Differently (Ep 121, 49:44). Worth the walk to work.
Alex Luketa spent fourteen years inside Goldman Sachs, Morgan Stanley and Credit Suisse before co-founding Xerini, where he's now CTO. That background matters, because he's watched the same failure twice now, once in finance, once in infrastructure. The AI project doesn't collapse at the model. It collapses at the point someone asks a simple question and discovers the answer lives in four systems, three spreadsheets and one bloke's head, and that no two of them agree.
His argument is that data silos and data quality aren't the boring prelude to AI. They are the AI project. And the organisations getting value out of it start embarrassingly small weith one dataset, one decision, one measurable outcome, rather than commissioning a transformation programme that promises everything and lands nothing. He's pointed on HS2 here: the lesson isn't that the technology was wrong, it's that scale amplifies whatever data discipline you already had. If it was poor, you've now bought poor at volume.
The neat inversion is what you do about it. Most people treat bad data as the thing blocking AI. Alex uses AI as the thing that finds the bad data, pointing models at the mess to grade quality, spot the contradictions and tell you where the rot is, before you've committed a penny to the shiny use case. Cheap, fast, and it produces the one thing a business case actually needs: a number you can defend.
He's not evangelical about it. The back half of the conversation is governance, privacy, the environmental bill for the data centres, and a genuinely uncomfortable stretch on inequality — who gets the productivity gain and who gets automated. He thinks local models on local hardware change that maths more than most people expect, and that legislation will arrive late and blunt, as it always does. His 2031 prediction is worth the last ten minutes on its own.
Why it matters: the 95% failure rate isn't a technology problem, it's a data-governance problem wearing a technology costume. If you can't answer what your data says today, a model won't rescue you — it'll just be wrong faster, and with more confidence.
👉 One thing to do this week: take the AI use case you're most excited about, and before you scope it, ask what single dataset it depends on and who owns its quality. If nobody can name that person in ten seconds, you've found your real project.
Links and Stuff
Alex Luketa (CTO & Co-Founder, Xerini) — LinkedIn
Xerini — xerini.co.uk · LinkedIn
Listen on
Subscribe to the Project Flux newsletter — the week's hottest AI and project-delivery news in your inbox every Monday.
Project Flux is independently owned, funded, and operated by its founder. Any external references or features are for community benefit only and do not imply endorsement, control, or ownership.