The Operations Bridge

Point eleven in the AI Strategy Points series. Last time was about solving real problems for real people. This post returns to a theme from point six, but pushes it further: the prompt itself is only ever as good as the data underneath it.

Prompting gets the attention. Data quality does the work

There’s an entire industry now built around teaching people to write better prompts — clever phrasing, role assignments, structured formats. Some of that is genuinely useful. But the art of the prompt depends on the integrity of the underlying data, and no amount of prompting skill fixes what’s broken underneath.

I learned this directly while developing Power BI reporting with ChatGPT’s help. I could write an extremely precise, well-structured request for a DAX measure. If the underlying data had an inconsistent category, a missing definition, or a field that meant something different than I assumed, the result was a perfectly formatted, confidently wrong answer. The prompt wasn’t the problem. The data was.

Garbage in, eloquent garbage out

The old data adage — garbage in, garbage out — undersells what happens with AI. A traditional broken spreadsheet formula tends to produce an obviously broken result: a blank cell, an error, a number that’s clearly nonsense. AI doesn’t fail that way. It takes the garbage in and produces something fluent, well-organized, and entirely plausible-sounding on the way out. The failure is much harder to catch precisely because it doesn’t look like a failure.

That’s a more dangerous failure mode than the old kind, not a less dangerous one, and it’s the reason I’ve become almost obsessive about data definitions before I let a tool anywhere near a real decision.

What this looks like in practice

Before I trust a prompt-driven answer on anything operational, I check a few things: does the source data have a clear, agreed-upon definition, has it been reconciled against something the team already trusts, and would the people closest to the process recognize the underlying categories as accurate? Those checks take real time. Skipping them is exactly how a fluent, wrong answer ends up driving a real decision.

It’s not a coincidence that this point sits right after the reminder that AI is hyper-literal and inexperienced. A hyper-literal system built on unreliable data will confidently misrepresent your operation, using exactly the words you’d expect a trustworthy answer to use.

A question for leaders

Before your team’s next AI-assisted analysis goes into a decision, has anyone actually verified the underlying data, or did the polish of the output do the persuading?

Last post in this series: why getting the most out of AI requires asking different questions entirely — specifically, learning to live in “now what” and “so what” instead of stopping at “what.”