The biggest blocker to scaling AI isn’t the technology. It’s the data no one can fully explain.

Most AI strategies are being built on data that organisations cannot clearly trace, govern or defend. That is the real constraint that is already showing up in wasted investment, unreliable outputs and decisions that look confident but rest on weak foundations.

Across all sectors, organisations are moving quickly to embed AI into core operations – in pursuit of efficiency and a competitive advantage. But the underlying data environment has not caught up. As AI scales, those weaknesses do not get resolved, but instead compound.

Salesforce found that a staggering 84% of data and analytics leaders believe their data strategies need a complete overhaul before their AI ambitions can succeed. At the same time, 89% of organisations with AI already in production report inaccurate or misleading outputs, while 55% say they have wasted resources training models on bad data (Salesforce, 2025).

This is not a tooling problem. It is a visibility and control problem that is pervasive.

Organisations are not short of data. They are short of understanding what that data actually represents and whether it can be trusted when important decisions depend on it.

Most organisations are still solving for data quality. Cleaner datasets, better structure, fewer errors. That work does obviously matter, but it does not solve the true problem they are facing.

Provenance.

It is not enough for data to look usable. It needs to be explainable. That means being able to trace where it came from, how it moved across systems, what changed along the way, and whether there is enough evidence behind it to support the decisions built on top of it.

That distinction becomes critical as AI moves closer to decision-making. Outputs are easier to generate and easier to trust at face value, hiding the uncertainty underneath.

A dataset can look complete and well-structured while still being fundamentally misleading if its data origin is unclear or its context has been stripped away.

The risk is not sitting in the model, but in the data feeding it.

There is a widening gap between how organisations describe their data and how that data behaves in practice.

MindBridge describes this as a data paradox. Teams report confidence in their data, yet still experience operational delays, hidden errors in AI systems, and direct financial impact linked to data issues (MindBridge, 2026).

Salesforce’s findings point in the same direction. Many organisations describe themselves as data-driven, yet still struggle to turn data into meaningful, AI-powered outcomes. Nearly half admit they sometimes draw incorrect conclusions because context is missing (Salesforce, 2025).

Organisations have built systems that produce outcomes, but those data-driven outcomes are crucially, neither visible or defensible.

As AI scales, that gap becomes operational risk. Outputs may look more sophisticated, but the underlying weaknesses remain. They are simply amplified and harder to detect.

At the same time, the composition of data is changing.

Synthetic data is becoming a larger part of the AI ecosystem. It offers clear advantages around privacy, cost and scalability, and its use is accelerating. But it introduces a more subtle risk.

As generated data is reused and fed back into future systems, it becomes harder to distinguish what is original, what is synthetic, and what can actually be trusted.

Research into model collapse suggests that repeated training on synthetic outputs can reduce diversity and accuracy over time. Systems can drift away from real-world signals while still appearing coherent.

This is not a problem with synthetic data itself. It is a visibility problem.

If organisations cannot clearly track the origin and composition of their datasets, they lose the ability to judge whether outputs are grounded in reality or gradually diverging from it. That makes provenance not just useful, but essential.

The regulatory environment is beginning to reflect the same shift.

The UK Data Use and Access Act 2025 does not replace UK GDPR, but it does change how organisations are expected to handle data in practice (Genc, 2025). The emphasis is shifting towards demonstrability. Not just having policies in place, but being able to show how data is accessed, processed and used across systems.

AI raises the stakes further. It increases the volume, speed and complexity of data use, making it much harder to rely on static governance models or retrospective checks. If data cannot be traced and explained, it becomes difficult to justify how decisions were made or defend them when challenged.

Governance can no longer sit alongside the business as a compliance layer. It is now business-critical to embed it into how data flows, how systems operate and how decisions are produced.

The conversation needs to move beyond adoption.

A more useful, and prescient question is whether the organisation can trust the data behind its AI systems. That means understanding where data originates, how it is validated, how it moves across systems, what permissions apply to it, what has changed over time, and whether there is enough evidence to explain outcomes when those outcomes are challenged.

Those are no longer theoretical governance questions. They are operational questions, financial questions and increasingly leadership questions.

Data has always mattered, but the consequences of getting it wrong are becoming much harder to hide.

AI has made the strengths and weaknesses of data environments more visible. Regulation is moving towards evidence rather than intention. Taken together, all of that points in the same direction.

Poor data control is no longer a technical nuisance sitting somewhere in the background. It is a business liability with massive financial, operational and strategic consequences.

The organisations that treat governance, provenance and control as active capabilities will be in a stronger position to scale AI with confidence. The ones that do not may still move quickly, but they will find it much harder to trust the outputs, defend the outcomes or sustain the progress.

That is the real divide now. Not between organisations using AI and organisations that are not, but between organisations building on data they can trust and organisations hoping the weaknesses underneath will stay hidden. For those organisations, data provenance for AI is quickly becoming one of the clearest tests of whether that trust is real.

If you can’t clearly explain where your data comes from, how it moves across its lifecycle, your AI strategy is already at risk. That’s where the work needs to start.

If you are unsure whether your data can support AI at scale, the first step is to test how visible, governed and explainable it really is. Get in touch.