When data means different things to different teams, AI outputs can't be trusted

Only 12% of organisations report sufficient data quality and accessibility for AI. If you've tried to build anything on top of enterprise data, that number probably doesn't surprise you. But data quality is the surface issue. The harder problem is that the same data means different things depending on who's using it and why; semantic data, considered critical yet mostly ignored. And until that's sorted, quality metrics are insufficient for trustworthy AI systems.

This is the second in a four-week series on AI governance that gets AI deployed. Last week: the overview and three pillars — data, algorithm transparency, people. This week: data governance, the first pillar and the foundation where trust starts being built. If AI governance closes the information gap between business strategy, technology capability, and operations, data is where that gap first becomes visible.

This is only a high-level overview as AI Governance in practice is incredibly context-specific.

Data doesn't carry its own meaning

Ask a fraud team what "high-risk customer" means and you'll get an answer rooted in transaction patterns, velocity, and geographic anomalies. Ask the credit underwriting team and you'll get a completely different definition built around repayment probability and exposure thresholds. Same label, different concepts, different data behind each one. Neither is wrong. But feed both into the same model without reconciling what the label actually represents, and the output becomes uninterpretable.

This happens everywhere. "Monthly active users" means one thing to the product team (unique logins), another to marketing (engaged sessions above a threshold), and something else again to finance (billable accounts). "Revenue" gets measured differently by sales (bookings), finance (recognised), and customer success (expansion). The further organisations specialise, the more these semantic conventions diverge. Every domain develops its own shorthand, its own assumptions about what data represents. That's normal and necessary for domain expertise. The problem is that nobody coordinates across those boundaries.

64% of organisations identify data quality as their top AI challenge, up from 50% just two years earlier. 43% cite it as their single biggest obstacle. But quality metrics like completeness, accuracy, and timeliness all assume the data's meaning is agreed upon. If two teams are using the same field to represent completely different concepts, measuring accuracy doesn't help. You're measuring accuracy against the wrong definition.

The coordination burden nobody budgets for

Data governance in practice is a coordination and alignment or consensus problem. Getting teams to align on shared definitions, agree on what each data element represents in each context where it's used, and make those agreements explicit and transparent takes more effort than most organisations plan for. It sits somewhere between "everybody's responsibility" and "nobody's job."

91.2% of executives say cultural challenges are the principal barrier to becoming data and AI-driven, ahead of technology. Data meaning is one of the places where cultural misalignment becomes concrete and measurable. When two teams can't agree on what a field represents, that's a coordination failure with specific consequences for every model that touches that data.

76% of data leaders say AI governance doesn't keep pace with employee AI use. Part of that gap is the semantic coordination that nobody assigned to anyone. And the effects compound. 65% of employees trust AI data, but 75% need data literacy upskilling to actually work with it. The gap between trusting data and understanding what it means is exactly where these problems live. People act on outputs they trust but can't evaluate. Blind trust is worse than scepticism; at least scepticism makes you pause.

In many programmes, teams discover months into a project that they had been operating with incompatible definitions of the same entity. Not because anyone was careless, but because nobody had created the space to surface those differences before they mattered. And they always matter eventually. With AI, they matter sooner, because the model amplifies the inconsistency across every prediction it makes.

From quality checks to shared understanding

Better data pipelines help, but the real shift is making data meaning explicit and agreed across every team that produces or consumes it. This is where business intuition first gets translated into something technology can work with and operations can run. What the business knows intuitively about customers, risk, or value has to become concrete enough to flow through data pipelines and into workflows that produce consistent outcomes. Data governance is that translation layer, and it's the first place where business, technology, and operations need to build consensus before alignment on algorithms or accountability is even possible.

Data governance failures are the primary reason most AI projects don't deliver the expected results in production. Sometimes the data is clean; it's just being used to mean something it doesn't. A model trained on one team's definition of "customer lifetime value" will produce outputs that are technically accurate and practically misleading to any team using a different definition.

AI is the first technology that makes this coordination problem impossible to ignore. Previous platforms needed it too: ERP implementations, data warehouses, analytics systems all suffered from the same semantic fragmentation. But the consequences were slower and easier to work around. AI compresses the feedback loop. A model trained on misaligned data produces wrong outputs immediately, and at scale.

Governance that works is alignment embedded in how teams work together throughout the AI lifecycle, fluid and context-specific, surfacing where meanings diverge and creating shared understanding before those divergences become blockers.

What comes next

AI Readi's governance framework surfaces data alignment where and when it matters for each specific context. No blanket checklist before deployment, no gate that slows everything equally. Data governance that shows up where it's needed (within the specific use case, for the specific teams involved) so businesses move faster with higher trust in what they're building on. The outcome: AI solutions that perform the same in production and at scale as they did in the pilot.