Dissonance in the data layer

Most AI projects that fail do so because organisations are quietly feeding their AI a data foundation they have long since stopped questioning. And while the large language models (LLM) available today are genuinely impressive, even the best model can only work with what it is given. In most organisations, LLMs are given incomplete records, inconsistent fields, no clear ownership, no meaningful metadata, and data spread across systems that have never properly talked to each other. The predictable result is an AI that sounds confident but is frequently wrong. What is much worse is that the people relying on it are often the last to know.

Three things AI will quietly punish you for

Across our work, we keep seeing the same root causes behind disappointing AI. And technology is rarely it. Instead, we see three areas where getting the foundation right makes the difference between AI that creates value and AI that quietly causes damage.

1. Your agent is only as good as the data it draws from

Data quality is not a new problem, but AI is making it an urgent one. An analyst working with imperfect data senses when something is off. They pause, ask questions, dig into the source, and apply judgement. But an AI model has no such human instincts; it takes whatever it is given and produces a confident output, regardless of what lies underneath.

This matters most when those outputs feed straight into decision-making:

A finance team we worked with deployed an AI assistant to support forecasting, and early results were strong, the kind of early win that builds belief in a programme. But the model was drawing from actuals that were inconsistently maintained across business units, with unreconciled adjustments and miscategorised entries that had never been cleaned up. The model had no way of flagging this. It kept producing outputs that looked precise, and those outputs fed quietly into decisions, until a number surfaced in a board presentation that no one in the room could account for, resulting in the kind of silence you do not soon forget.

Poor data quality used to slow down reporting; now it actively misleads the AI sitting on top of it, at speed and at scale.

2. AI inherits every permission gap you have ever ignored

Most organisations have permission gaps. Folders that were opened for a project years ago and never closed, access rights that quietly followed someone into their next role. You have lived with these for years without much consequence.

AI changes what those gaps cost you. When a person opens a document they were never meant to see, there is usually a trail, a hesitation, a moment of human judgement involved. When an AI tool pulls the same data, the exposure can be systemic, silent, and already woven into outputs before anyone notices.

A senior leader at an organisation used an AI tool to surface insights from internal financial data and received a summary that included margin information from a business unit outside their remit. No system was breached. The data was simply accessible, and no one had defined what the tool was permitted to use or for which users it was intended. By the time legal and compliance had assessed the exposure, three months of rollout momentum had been lost, and with it, the board's confidence in the programme. This is the kind of story that ends with uncomfortable questions being asked at the top, about liability, about governance, and about who was supposed to be responsible.

Clear permissions and enforced access boundaries are not constraints on what AI can do. They are what makes it safe to let AI loose on your data. Without them, scale is not an achievement but an exposure.

3. Feed AI everything and it broadcasts the mess

One of the most common mistakes we see in AI projects is treating data access as a proxy for data readiness. “Connect AI to everything,” the thinking goes, “and let it figure things out.” What you actually get out of that approach is every inconsistency in your business, every disagreement about what a given number means, every workaround your analysts have been carrying for years, neatly surfaced at speed and packaged as insight.

A group finance team connected AI to every available source to produce a consolidated view of revenue performance. Within days, it was returning three different figures for the same period, the result of different revenue recognition methods, different currency conversion points, and different definitions of what counted as closed revenue. None of this was new. Analysts had been quietly reconciling these differences manually every month. What the AI did was strip that reconciliation away and put the unfiltered inconsistency in a board report. What had been a quiet operational issue suddenly became the CFO's problem.

A well-structured data foundation, with a single trusted view of what the numbers actually mean, is what allows AI to produce outputs people will act on. Getting there is a data problem and a governance problem in equal measure. Someone has to decide which definition of revenue is the right one, and that decision is not made by a system. It is made by people, in rooms, with all the politics and diplomacy that come with it. And ultimately, it is a decision that belongs to the CFO, not something to delegate to a data team while hoping for the best.

The hidden blocker

The reason why data foundations do not get fixed is rarely ignorance. Most finance and data leaders know the problems exist. The barrier is that fixing them does not have a natural owner, does not appear on a P&L, and competes for budget against initiatives with clearer ROI narratives.

There is also a political dimension that is easy to underestimate. Agreeing on a single definition of revenue, or deciding who owns a dataset, requires a decision on what is ‘correct’ – and that inevitably surfaces competing views and leads to demanding discussions that are often harder to resolve than simply buying a new tool.

The organisations that do make progress tend to start small and stay concrete. Which three or four datasets does your AI most depend on? Who owns them? What standard do they need to meet? Starting there, with clear accountability and a fixed timeline, is more likely to create lasting change than a broad data strategy that tries to solve everything at once.

This is why this is primarily a leadership investment, rather than a technology investment. Such investments rarely get celebrated in the moment but are what separates the organisations whose AI creates value from those still wondering why theirs does not.

The question worth asking

The organisations getting their AI projects right are treating data quality, governance, and structure as prerequisites, not afterthoughts. They invest in the unglamorous layer and apply the same rigour to their data that they already apply to their financial controls.

The question is not whether your AI is capable enough. It almost certainly is. The question is whether your data is ready, and whether your organisation is willing to do the very human work that getting it right requires. Those that answer yes will not necessarily have better AI than their competitors. They will simply have AI they can trust.