Most companies are running AI projects. Almost none are ready for them.
Nearly 18 months into the generative AI boom, enterprises are hitting a wall that no model upgrade can fix. According to the Dun & Bradstreet AI Momentum Survey — 10,000 businesses across 32 countries, published May 2026 — 97% of organisations have active AI initiatives. Only 5% say their data infrastructure is genuinely ready to support them company-wide. The gap between polished demos and real-world deployment has never been more visible.
The money picture
The financials are mixed but heading in one direction. Sixty percent of businesses report measurable returns from AI, with 24% describing those returns as strong. That's enough to keep spending commitments high: 56% plan to increase AI investment over the next 12 months, and roughly a third have already moved from pilots into live operational use.
Dun & Bradstreet's chief strategy officer Cayetano Gea-Carrasco puts it plainly: a single proof-of-concept can run on messy, fragmented data. Scaling it across an organisation requires something different entirely — a unified, governed, interoperable data ecosystem. Without that, AI stays an expensive experiment locked inside one department.
The infrastructure wall
When companies try to move beyond sandboxed testing, four barriers keep coming up. Data access complexity tops the list at 50%, followed by privacy and compliance concerns (44%), poor data quality (40%), and integration gaps between existing systems (38%).
Only one in ten leaders says their organisation is confident in its ability to manage AI-related risk. That number matters more now that agentic AI — autonomous systems that act on your behalf without human sign-off at every step — is moving from concept to product. These systems need continuous, clean access to live data. Most corporate environments were built around human workflows with manual approvals baked in.
In the UK, the pressure is compounding. The ICO's statutory code on automated decision-making arrived in spring 2026, and its approach diverges from the EU AI Act — meaning UK firms operating across borders face two separate compliance tracks simultaneously. The ICO Agentic AI Report (Feb 2026) explicitly flags supplier accountability for agentic AI controls and liability exposure when AI systems produce incorrect outputs.
What changes now
The practical shift already under way: AI is being repositioned away from "replacing workers" and toward accelerating specific tasks — onboarding, compliance analysis, data preparation for decisions. The productivity gains show up consistently where data is already structured and accessible. Where it isn't, results remain patchy.
The defining challenge of 2026 isn't finding a smarter model. It's the unglamorous work of data architecture — cleaning, connecting, and governing the information those models actually run on. Companies that skip that step will keep running pilots indefinitely while others start collecting the operational gains.
