The Operations Bridge

Point six in the AI Strategy Points series. Last time was about AI as a multiplier on existing judgment. This post is about a sequencing mistake I see constantly: starting with training instead of starting with data.

The natural starting point isn’t the right one

Most AI initiatives start with platform training or prompt training. It’s the natural instinct — get people comfortable with the tool, teach them good prompting technique, and adoption will follow. I understand the appeal. It’s visible, it’s schedulable, and it produces a training completion metric leadership can point to.

But the real starting point is data. If I said this before I actually lived it, I’m not sure I would have believed how much it mattered. Building operational reporting with AI taught me that no amount of prompting skill compensates for messy, undefined, or unreliable underlying data. You can write the most elegant prompt in the world and still get a confidently wrong answer if the data feeding it doesn’t mean what everyone assumes it means.

What training-first initiatives tend to miss

An organization that trains everyone on prompting first often discovers, months later, that the tools produce inconsistent or untrustworthy results — and the instinct is to blame the training, or the tool, or the people. Rarely does anyone go back and ask whether the underlying data had clear definitions, consistent categories, and a trustworthy source in the first place.

This mirrors something I wrote about in an earlier series on this blog: before you can ask what a metric means, you have to establish that it’s a trustworthy fact in the first place. AI doesn’t change that requirement. If anything, it raises the stakes on it, because AI will happily generate a fluent, confident-sounding answer from bad data without flagging that anything is wrong.

A better sequence

Before investing heavily in platform or prompt training, I now push for a data audit first: are the definitions clear, are the categories consistent, is the source reliable, and do the people closest to the process trust what’s being recorded? Only after that foundation is reasonably solid does prompt and platform training actually produce results people can rely on.

This doesn’t mean waiting for perfect data before doing anything — that’s its own trap. It means being honest about which data is solid enough to build on now and which needs work first, instead of assuming the AI layer will somehow smooth over problems underneath it.

A question for leaders

Before your next AI training rollout, ask: do we actually trust the data this tool will be working with? If the honest answer is “not really,” that’s where the real work needs to start.

Next: why I think of AI as a super intelligent, hyper-literal, eager, but inexperienced intern — and what that metaphor should change about how you work with it.