Your AI Strategy Is a Data Strategy

Every company I talk to has an AI strategy. Far fewer have noticed that it's mostly a data strategy with a new name.

Models are improving every few months and getting cheaper as they do. What isn't improving on its own is the state of the data those models need to do anything useful inside a business. That's where most AI initiatives quietly stall.

The demo-to-production gap

The pattern is familiar. A team builds an impressive prototype in two weeks. Leadership gets excited. Then the project spends six months stuck on questions like:

• Which system is the source of truth for this number?
• Why does the agent's answer differ from the dashboard?
• Who is allowed to see this, and how do we enforce it?
• How do we know it's still right next month?

None of those are model problems. They're data problems we've had for decades, now arriving faster and with a chatbot attached.

Five foundations for AI-ready data

1. Purpose before platforms

Start with the decision you want to improve, not the model you want to use. "Help sales leaders forecast pipeline more accurately" is a strategy. "Deploy an LLM" is not. Purpose determines which data matters, which definitions apply and what "good" looks like.

2. Shared definitions

If two teams define a metric differently, an AI will cheerfully use both, and confidently. A governed semantic layer (metrics, dimensions and their business rules, defined once) is the single highest-leverage investment for AI readiness. It's also exactly what good analytics teams have been building for years.

3. Connectivity through governed channels

Agents need access to live, trusted data, not exports. Standard connectivity layers like the Model Context Protocol let you expose curated tools in front of your warehouse and analytics platforms, so agents reach data the same way your dashboards do.

4. Permissions that follow the user

AI shouldn't become a back door around access controls. The safest designs pass the user's identity through to the data layer, so an agent can never show someone more than they could already see.

5. Evaluation, not vibes

Data teams have always reconciled numbers against a known truth. Apply the same discipline to AI: build a small set of real questions with known correct answers, and test every change against it. If you can't measure accuracy, you can't scale trust.

Where to start

You don't need a multi-year transformation to begin. Pick one high-value workflow, then:

1. Write down the decision it supports and who makes it.
2. Identify the five to ten metrics it depends on and agree on their definitions.
3. Expose those metrics to an agent through a governed, read-only interface.
4. Create twenty test questions with known answers.
5. Ship to a small group of real users, and iterate.

That's it. Each workflow you do this way leaves behind reusable definitions and connectivity, and the next one gets faster.

Companies don't win with AI because they picked the best model. They win because their data was ready when the models got good enough.

The good news for anyone with a data background: you already know how to do most of this. The work hasn't changed nearly as much as the headlines suggest.