AI Doesn't Fix Broken Data. It Amplifies It

The phrase ‘data strategy’ has come to mean almost anything. A warehouse migration, a governance committee, a Power BI rollout, a cloud platform decision or a roadmap document that lives in a shared drive and gets dusted off once a year. When all those things can sit under the same heading of ‘strategy’, that framing has stopped being useful.

 The pilot worked, the rollout won’t…this is exactly what industry leaders need to address before the next AI investment decision.  

The pilot worked. The rollout won't.

There’s a moment that catches a lot of organisations off guard, the one where the AI pilot runs smoothly, the demo lands well in the boardroom and budget is released for enterprise deployment – the perfect scenario… But, then somewhere between proof of concept and production, the cracks start to show, and the outputs drift, confidence drops and the same model that looked transformative six months ago is now generating answers nobody fully trusts.

What’s changed? It’s rarely the model but the data underneath it. Pilots run on curated, controlled, well-understood data, enterprise deployments run on everything else; the spreadsheets, the legacy systems, the half-migrated CRM, the customer records that have been re-keyed across three platforms over the last decade. AI doesn’t surface those problems and resolve them, it uses whatever it is given and produces outputs at speed and scale. If the foundation is wrong, the outputs reflect that at a similar speed and scale.

This is the single most underestimated dynamic in mid-market AI investment right now, and it is the reason so many organisations are quietly entering what we’ve started calling the pilot graveyard; the growing collection of AI projects that proved the technology works, then ran into the data reality and stopped.

 The Ferrari engine problem  

I use an analogy in a lot of these conversations because it makes the dynamic tangible. If you put a Ferrari engine into a car with flat tyres, on a road it was never built for, nothing happens the way it should, not because there’s anything wrong with the engine, but because it isn’t in the right place.

That’s the situation most AI implementations are in. The models are remarkable, the platforms are mature, but the foundation underneath them; the data, the integration and the architecture wasn’t designed for this. It was designed, in most cases, to support yesterday’s operational reporting, not tomorrow’s autonomous decision-making.

McKinsey's State of AI 2026 reports 78% of organisations now use AI in at least one function and 65% only 6% are seeing material EBIT impact from it. That isn’t a technology problem, the technology is doing what it’s supposed to do. The gap between adoption and impact is the gap between buying capability and being able to use it and that gap almost always runs through the data layer.

Gartner’s insight suggests the same: that through 2026, organisations will abandon 60% of AI projects unsupported by AI-ready data! Not paused, not deprioritised but abandoned, because at some point in the lifecycle the data reality catches up with the AI ambition, and there isn’t a productive way forward.

What 'broken data' looks like

When organisations hear ‘broken data,’ they tend to think of obvious errors such as duplicate records, missing fields, the customer who’s been entered into the CRM twice with slightly different spellings, those things matter but they’re not what makes data unfit for AI.

Data that breaks AI deployments tends to break in five specific ways, and most organisations have gaps in more than one:

• Quality: accuracy, completeness, consistency. Whether the same fact is recorded the same way across the systems that hold it, whether the values in the field mean what the field name suggests.
• Integration: whether the data exists in a form the AI can reach. Most organisations have customer data spread across a CRM, a billing platform, a marketing automation tool, and three spreadsheets that someone in operations maintains by hand, the reality is that AI doesn’t stitch that together, it works with what is connected.
• Governance: who owns the data, who can change it, who has access to which fields, and whether those rules are enforced in practice or just documented in a policy nobody opens. AI applied to ungoverned data scales not just the insight but the risk.
• Literacy: whether the people interpreting the AI’s outputs understand what the underlying data is and isn’t telling them. A model is only as useful as the human decisions it informs.
• Business alignment: whether the data being used to train and run the AI is the data that actually reflects what the business cares about. A model tuned on the wrong inputs can be 99% accurate and still useless.

The reason these matter together, not individually, is that AI compounds them. A 2% improvement in model accuracy delivers nothing if the underlying data is incomplete, stale, or contested. MuleSoft's 2025 Connectivity Benchmark found 95% of IT leaders attributing integration hurdles impeding AI implementation, it’s a higher-ranked barrier than skills, than budget or than executive sponsorship.

The C-suite is asking different questions about the same investment

In most organisations, the CEO is asking how AI is going to transform the business, the CFO is asking where the return is and over what timeframe, the CIO is asking what the platform needs to look like and how it integrates with everything else, and the CTO is asking what the underlying architecture has to support.

All four questions important, the problem comes when those questions are being answered separately by different teams, in different conversations, against different success criteria. The result is AI investment that’s technically capable and strategically adrift. Pilots get funded, tools get bought but there isn’t a coherent picture of what the organisation is trying to do with any of it.

This is why the data foundation conversation matters at C-suite level, not just at the data team level. The technical work; quality, integration, governance is achievable. The harder work is aligning the four executives funding it around a shared answer to a simple question: what does our data need to be able to do, for the business to be able to do what we want?

What 'broken data' looks like

We see this play out regularly, and the pattern is consistent regardless of industry. A financial services firm preparing for a regulatory reporting deadline, looking at how to bring calculation logic spread across spreadsheets into a more reliable, auditable structure. A retailer ready to personalise at scale, working through what it takes to align customer records across the CRM, the loyalty platform, and the e-commerce stack. A professional services business with growing ambition around resource forecasting, looking at how to build trust and consistency into time-recording data. A housing provider exploring how AI could help prioritise repairs, thinking about how tenant and asset data, originally captured for compliance reporting, can be reshaped to support operational decisions too. Different sectors, different starting points but the same underlying opportunity. The data wasn't designed to support what the business now wants to do with it, and that's a foundation question worth getting right before the next stage of investment.

The making the most meaningful progress aren't necessarily the ones with the cleanest starting position, they're the ones who recognised the data foundation question early, addressed it in parallel with delivery, and built the architecture that the next phase of their ambition could run on.

Read the full whitepaper >

What to do before your next AI investment decision

None of this is an argument for putting AI on hold while a multi-year data transformation runs in the background. That sequence doesn’t work for most mid-market organisations, the appetite and the competitive pressure is real, and a two-year wait isn’t a credible answer to either.

What it is an argument for is a different conversation before the next significant AI investment. Specifically, three things:

1. Align the C-suite around one question, not four.

Get the CEO, CFO, CIO and CTO into the same conversation about what the organisation needs AI to be able to do, and what the data has to support for that to be possible. The four-question problem doesn’t get solved by better project management, it gets solved by a single, shared picture.

2. Establish where the foundation is, honestly.

Map the data estate against the five dimensions; quality, integration, governance, literacy, business alignment. Be specific about where the gaps are and what they mean for the use cases being considered. Most organisations skip this step or do it superficially. The ones who do it properly tend to make smaller, better-targeted AI investments that scale.

3. Scale from what’s ready, not from what’s ambitious.

Sequence the first AI deployment against the part of the data estate that is already in good enough shape, while remediation happens in parallel on the rest. This is how you build momentum without building on sand.

The bottom line

AI doesn’t fix broken data. It amplifies it. That’s the part most organisations underestimate at the point of investment decision, and it’s the part that determines whether a year from now they’re scaling AI or quietly winding down another set of pilots.

AI amplifies broken data, it also amplifies good data at the same speed and at the same scale. The organisations getting genuine return from AI right now aren’t necessarily the ones with the best models, they’re the ones who did the foundational work, made the C-suite conversation a shared one, and sequenced the investment against what the data could actually support.

In our experience this is the conversation that makes the difference.

What to do next

 If this is the conversation you’d benefit from, book a 30-minute conversation with our Data & AI Practice below, no preparation needed, no obligation.