Field note · AI in practice · Judgment & governance
Directing AI, not just prompting it.
Everyone can get an answer out of an AI now. The value was never the answer. It is knowing which answer to trust, which to throw out, and which questions the tool should never be handed in the first place.
There is a lot of noise right now about people who are good with AI. Most of it measures the wrong thing. Speed and volume are easy to show off, but in business intelligence the output is not the job. The job is a decision someone is going to make with real money behind it, and a fluent answer that happens to be wrong is more dangerous than no answer at all.
So the way I work with AI starts from a rule that sounds backwards: the most important skill is knowing when not to use it. I treat AI as a drafting, querying, and research partner, never as a source of truth. The source of truth is the data, and a human stays accountable for every number that reaches a decision-maker. That is not a limitation I begrudge, it is the entire point. It is what lets me move fast on the mechanical work without ever putting my name behind something I cannot defend.
Where a human stays in the loop
Experience has taught me where the human judgment has to stay front and center, and these are the moments I am most deliberate about. A number is not decision-ready until it has been traced back to the source and checked, so I do that before it reaches anyone. Confidential, regulated, or non-public data does not belong in an external tool, and I will always find another way. What a business metric means is a governance call, so I treat it as one and reconcile the definition with the stakeholders who own it rather than letting a model settle it quietly. And when the question is why something happened, I slow down on purpose. A model can surface a pattern in seconds, but a pattern is not a cause, and the leap from correlation to root cause is exactly the kind of judgment that is worth talking through with the people closest to the work.
The through line is simple. Anything that carries accountability, a client deliverable, an audit-facing report, a recommendation that drives a budget, gets reviewed, verified, and owned by a named human. If I cannot independently verify the output against the source, it does not ship. None of this is rigid for its own sake. These are starting principles, and the right move on any given project gets refined through deliberation with the team and the stakeholders who live with the results. Rest assured, that discipline does not slow the work down. It is what makes the fast parts safe to be fast.
Where it earns its keep
Inside that fence, I lean on AI hard, and it changes what a single person can deliver. It drafts the recurring report so I start from a structured first pass instead of a blank page. It writes and tests a first-round query so I can spend my attention on whether the logic is right rather than on syntax. It scans a market or a competitive set and gives me a starting map I then verify against real sources. It is a genuine thinking partner for surfacing patterns I might have walked past, as long as I am the one who decides what those patterns mean.
Getting good output is a craft of its own, and most of it comes down to how you frame the ask. I give the model a role, a goal, and a format, then I decompose a vague question into the smaller testable pieces it can actually answer. "Why are we down this quarter" is not a prompt, it is a project. Broken into the right sub-questions, with the right context and constraints attached, it becomes something a tool can help me work through quickly. That decomposition, turning a fuzzy business question into answerable parts, is a skill I have carried from analysis long before AI, and it is the same skill that makes AI useful now.
Why this matters for who I work with
The market is moving toward a world where natural-language-to-query is a commodity and anyone can generate a chart. When that happens, the premium does not sit with the fastest generator. It sits with the person who makes self-service trustworthy, who can look at a confident answer and know whether it holds, who owns the definitions and the verification and the framing. That is the direction I am building toward on purpose: not report producer, but the person a team trusts to decide which answers are real.
This whole site, and the tools in my case studies, were built in exactly this way. I directed the work, made the calls, and verified the output, while AI compressed the mechanical distance between idea and result. That is the modern advantage I bring, and I am careful about it precisely because I take it seriously.
Related: Thirteen iterations, directing AI on a real build →
Trying to use AI without betting the business on it?
That balance, moving fast on the mechanical work while keeping a human accountable for the calls that matter, is exactly the kind of thing I like to help teams get right.