Most failed AI projects don't fail because the model was wrong. They fail because of a decision made or not made - long before anyone chose a model: the state of the data underneath it.

It's an uncomfortable truth, because the model is the exciting part. It's where the demos happen, where the budget conversations start, and where most of the attention goes. But the evidence is now hard to ignore. Gartner has projected that, through 2026, organisations will abandon around 60% of AI projects because the underlying data cannot support them. The model rarely breaks first. The data does.

This article looks at why data - not the algorithm - is the binding constraint on enterprise AI, why the problem is so easy to miss, and what "data-ready" actually looks like in practice.

The pattern everyone recognises but few name

The story tends to run the same way. A small team runs a proof of concept on a curated slice of historical data. The results are genuinely promising. Confidence builds, budget is released, and the project moves toward production - at which point it hits a wall. The pilot that dazzled on clean, narrow data behaves very differently against the messy, live, full-scale reality of the organisation's systems.

This is not a fringe outcome. In regulated and infrastructure-heavy sectors it is especially acute: McKinsey's 2025 analysis found that around seven in ten energy AI initiatives remain stuck in pilot - among the highest stall rates of any sector. Across industries, commentators note that a large share of organisations discover their data infrastructure is fundamentally inadequate only after launching ambitious AI initiatives, not before.

The reason the wall appears so suddenly is that small datasets, narrow use cases and manual oversight mask deeper problems. At pilot scale you can tolerate imperfect data. At production scale, those same issues compound: the model meets more edge cases, touches more of the business, and the cost of every error becomes real and visible.

Why it's the data, not the model

There's a structural reason this keeps happening. Capable models have become broadly available - the differentiator between organisations is increasingly not which model they use, but whether the data feeding it is governed, traceable and reliable. The organisations getting the most from AI in 2026 are generally not the ones deploying the newest models. They're the ones that built strong data foundations so those models can operate on trustworthy information.

AI also changes how data is consumed, which is the part traditional data programmes weren't built for. Conventional governance grew up to support reporting and analytics: curated, modelled, visualised data feeding dashboards. Modern AI does something different - it ingests vast volumes of structured and unstructured content, from documents and emails to operational logs, across a far broader and more dynamic lifecycle. Governance now has to follow data across all of that, not just the tidy reporting layer.

And when the foundation is weak, the failure mode is quietly dangerous. Models trained on inconsistent or outdated data drift in production and break silently - often without obvious detection - feeding wrong answers into customer interactions, automated decisions or operational systems before anyone notices. Poor data controls are also, as Accenture's research has highlighted, a leading cause of AI compliance and regulatory risk. In other words, bad data doesn't just make AI underperform; it makes it a liability.

The trust paradox

There's a subtler problem layered on top, and Informatica's 2026 survey of data leaders put numbers on it. A clear majority of employees believe the data behind their AI is sound - yet most don't have the literacy to question it. That combination, blind trust in data nobody is scrutinising, is precisely the kind of thing that turns a quiet data-quality issue into a confident, wrong decision at scale.

The same research found that even as organisations race toward production, data reliability remains one of the top barriers to scaling, and for agentic AI specifically - where systems don't just generate but act - around half of data leaders cite data quality and retrieval as their single biggest challenge. The more autonomous the system, the less forgiving it is of a shaky foundation.

Governance has fallen behind - and that's now a regulatory problem

Here is the gap that ties it together. Three out of four organisations admit their governance has not kept pace with their AI adoption, and only a small minority describe their data governance as genuinely mature. Meanwhile the regulatory expectation is moving the other way. Frameworks such as ISO/IEC 42001 and the NIST AI Risk Management Framework are becoming reference points, the EU AI Act is in force, and a growing share of organisations now name compliance and regulatory readiness as their biggest single adoption challenge.

For organisations in regulated sectors, this is not abstract. The ability to show where data came from, how it was transformed, who can access it, and whether it can be trusted is moving from good practice to a requirement you may have to evidence. Lineage, quality and access control are becoming the things a regulator - or your own board - expects you to demonstrate, not just assert.

Reframing governance: enablement, not enforcement

The instinctive response to all this is more control, more sign-offs, more friction - which is exactly why so many governance programmes stall and get a reputation for slowing the business down. The more useful shift, and one a number of leading organisations are now making, is to treat governance as enablement rather than enforcement. Some have gone as far as renaming the function "data enablement" to make the point. The intent matters more than the label: governance done well doesn't block AI, it's the thing that lets AI move safely and at speed. Clean, well-described, trustworthy data is what makes a model's output accurate, consistent and explainable - to an executive, a frontline user or a regulator.

What "data-ready" actually looks like

If the model isn't the question, what is? In practice, an organisation that is genuinely ready to scale AI can usually answer "yes" to a short list of unglamorous questions:

• Quality and consistency - is the data accurate, current and consistent enough that the same question returns the same trustworthy answer every time?

• Lineage and traceability - can you show where data originated, how it was transformed and where it flows, automatically rather than by hand?

• Access and permissions - does the system surface only the data a given user is allowed to see, with no new access pathways quietly opened up?

• Semantic consistency - do "customer", "asset" or "incident" mean the same thing across systems, so AI isn't reasoning over contradictory definitions?

• Governance that travels with the data - do quality, privacy and policy controls follow the data across its full lifecycle, including the unstructured content AI now consumes?

None of this is exciting. All of it is what separates an AI programme that scales from one that stalls.

Where we'd start

In our work with data-rich, regulated organisations, the most valuable first step is rarely a model or a platform decision. It's an honest, structured picture of where the data foundations actually stand - what's strong, what's fragile, and the few changes that would most reduce risk and unblock the use cases that matter. A focused data-readiness assessment does exactly that: it surfaces the governance gaps that would otherwise derail an initiative, before significant money is committed.

That's the approach we favour - technology-agnostic, grounded in the organisation's existing estate rather than a product to sell, and delivered as a short, sharp engagement that produces a defensible plan quickly. The work it points to is foundational rather than headline-grabbing: quality, lineage, access, consistent meaning, governance that enables. But it is the work that decides whether AI delivers value or quietly becomes a liability.

The lesson of the failed projects is not that AI doesn't work. It's that the success was decided earlier than anyone was looking. Get the foundations right, and the model has something solid to stand on.

Wondering whether your data is genuinely ready for AI? A short data-readiness assessment can give you a clear, honest view of where you stand and the practical steps that matter most before you scale.