Your AI strategy is a data strategy wearing a costume

Ask most organisations for their AI strategy and you will be handed a list of capabilities: a copilot here, a forecasting model there, document extraction somewhere in operations. It reads as a technology plan. Underneath, almost every item on it is a data question that has been deferred for years and is now being asked under deadline.

This is the uncomfortable arithmetic of applied AI. The model is increasingly a commodity — rentable, swappable, improving without your involvement. What is not commoditised is knowing what your data means, where it came from, who is permitted to see it, and whether it is still true today.

The questions that actually gate delivery

In practice, AI programmes stall on a short and predictable list. Nobody owns the customer record. Three systems disagree about revenue. The field everyone relies on was quietly repurposed in 2019 and never documented. The dataset that would make the model useful sits behind a governance process nobody has run in two years.

None of these are solved by a better model. All of them are made worse by one, because a model industrialises the ambiguity. A human reading a suspect field notices something is off and asks a colleague. A pipeline does not. It propagates the error at scale, with confidence, into a decision.

Where the leverage actually sits

• Ownership before architecture. Every dataset feeding a decision needs a named owner accountable for its definition, not merely its uptime.
• Lineage you can query rather than lineage you can draw. If tracing a number to source requires a meeting, it will not survive an audit or an incident.
• Access as a designed path rather than an exception process. Where obtaining data requires heroics, teams copy it, and your governance posture becomes fiction.
• Definitions expressed in code. A metric that lives in a slide deck will drift; a metric that lives in a tested transformation will not.
• Freshness treated as a contract with consequences. A model trained on the promise of daily data and fed weekly data does not fail loudly. It fails quietly, for months.

The reframe worth making

This is not an argument for a multi-year data programme before any AI work begins. That sequencing fails too, because a platform built without a consumer optimises for the wrong things and arrives to an audience that has stopped waiting.

The better approach is narrower and more disciplined. Pick one decision that genuinely matters. Follow the data behind it all the way to source. Fix what you find. You will finish with a working feature and a small, real improvement to the foundation. Repeat that eight times and you have both an AI capability and a data strategy — earned in the order that actually holds.