Scale AI too early and you automate the wrong thing faster

Almost every leadership team I speak to has the same tension. Demand for AI is high. The board wants a plan, colleagues are already using tools on their own, and there’s pressure to show something at scale this year. Meanwhile, the people closest to the data are quietly worried that the numbers underneath aren’t ready.

They’re usually right to be worried. In most organisations the constraint on AI isn’t ambition or budget. It’s trust and repeatability. If the business already argues about which revenue figure is correct, an AI that summarises, predicts or acts on that figure doesn’t settle the argument. It spreads it, faster and with more confidence.

AI doesn’t fix untrusted data. It industrialises it.

So I work along one path, in sequence. I first used it to shape a three-year Data and AI strategy for a PE-backed technology distributor, and I’ve since generalised it so it can frame any engagement. It has three phases, each with a badge, a plain-English goal and a clear test for moving on.

1. Trust at scale: make it trustworthy

Before anything scales, the data that matters most needs named owners, agreed definitions and checks that run every day. Run costs need to be visible, so nobody gets a surprise invoice when usage takes off. And there needs to be one front door for requests, so priorities are set deliberately rather than by whoever shouts loudest.

This phase is unglamorous and it is where most of the value is protected. It doesn’t need to take years. It needs to cover the data behind your priority journeys, not the whole estate.

Move on when priority data is owned and measured, and controls are in place.

2. Velocity through reuse: make it repeatable

The second phase replaces bespoke one-off builds with things you can use again: a handful of proven architecture patterns, shared data products and certified metrics. The point is that your fifth use case should be noticeably faster and cheaper than your first, because it reuses what the first four built.

If every use case is a fresh project with its own integrations and its own definitions, you don’t have an AI capability. You have a series of experiments.

Move on when new use cases reuse patterns and land faster than the last.

3. Value at scale: make it pay

Now you scale: assistants, automation and decision support, applied to measurable business problems in your customer and colleague journeys. Every use case has one primary outcome, a baseline and an owner, and investment expands only where the scorecard shows value.

Move on when benefits are reported with Finance and investment follows evidence.

This is not a waterfall

The path describes how an organisation’s capability matures, not a rule that nothing can happen until phase one is finished. A use case whose data is already reliable can go ahead early. One that depends on data nobody trusts waits, however exciting the demo was.

That is the real discipline: each use case moves when its own dependencies and controls are ready, and the organisation invests to make more of them ready.

Signs you’re about to scale too early

• Executive meetings still argue about whose numbers are right.
• Most of the data team’s time goes on tickets and fixes rather than improvements.
• Pilots can’t name an owner who will run them in production.
• Nobody can tell you what the platform costs per month, or why.
• Requests arrive through side doors: a direct message here, a favour there.
• Success is described as “people like it”, with no baseline to compare against.

If three or more of those sound familiar, the fastest route to value is to slow down for a few months and fix the foundations behind your priority journeys. Then scale with confidence, instead of scaling the argument.