The Skill I'd Actually Tell a New FP&A Analyst to Build Right Now (It's Not Prompting)

A younger colleague asked me recently what they should be learning to “stay relevant” given everything happening with AI in finance. Their assumption going in was that the answer would be prompt engineering, or maybe a specific tool. I get why—it’s the visible, learnable-in-a-weekend skill, and every LinkedIn post about AI and careers points at it.

I told them something different, and I want to actually explain the reasoning instead of just asserting it, because I think the popular advice here is optimizing for the wrong time horizon.

Prompting is a shrinking-half-life skill

The specific syntax and tricks of prompting a model well are already changing every few months as the tools themselves get better at inferring intent from vague instructions. What worked as a clever prompt technique a year ago is often just unnecessary now—the newer models don’t need the workaround anymore. If you build your professional identity around being good at a specific interaction pattern with a specific generation of tool, you’re building on ground that moves under you faster than almost any other skill in this field.

That doesn’t mean don’t learn to use the tools. Use them constantly, get fast and comfortable with them. Just don’t mistake tool fluency for the durable skill underneath it.

What I’d actually prioritize

Knowing what a right answer looks like before you ask the question. This is the single most durable skill in an AI-assisted workflow, and it’s not new—it’s the same instinct experienced analysts have always needed, just now more urgent because the tool will confidently hand you a wrong answer with the same fluency as a right one. If you don’t already have a rough sense of what the number should be in the ballpark of, you can’t catch when the tool is off. This is built the old-fashioned way: doing the work, comparing your estimate to the actual, doing it again.

Being able to ask the second and third question, not just the first. Anyone can ask an agent “what drove the variance.” The valuable skill is knowing to ask “is that driver seasonal or structural,” “does that match what the business partner told us last week,” “what would make this explanation wrong.” That’s a judgment skill, not a tool skill, and it transfers across every generation of AI tooling that comes next.

Writing clearly, in your own voice, without the tool. This sounds almost quaint next to “learn to use AI,” but I mean it seriously: if you can’t draft a clear, well-reasoned two-paragraph explanation of a variance yourself, you won’t be able to tell whether the tool’s version is actually good or just fluent. The skill of writing clearly and the skill of judging clear writing are the same skill, and you lose it if you outsource all of your drafting before you’ve built it.

Basic data literacy: where numbers live, how they’re defined, why two systems might disagree. This is unglamorous and it’s the thing that actually determines whether you catch a wrong AI-generated answer or pass it along trusting it.

The honest pitch to a younger version of myself

If I were starting over right now, I’d spend far less time trying to become an expert prompter and far more time building the pattern recognition that lets me sense when a number is wrong before I can articulate why. That skill was valuable before any of this AI tooling existed, it’s more valuable now because the tools remove the friction that used to force you to build it slowly, and it’ll still be valuable whatever the next generation of tools looks like. Prompting skill has a half-life measured in months. This one doesn’t.