We Don't Have an Agent Problem. We Have a Judgment Problem.

I sat in a demo a few months ago where a vendor had an agent pull actuals, compare them to plan, write the variance commentary, flag the three cost centers that needed attention, and draft an email to the business partner—all in about ninety seconds. The room was quiet in the way rooms get quiet when people are doing career-anxiety math in their heads.

Here’s the thing nobody said out loud: the agent did the steps well. It did the judgment badly. It flagged the three biggest variances by dollar amount, not the three that actually mattered. One of them was a timing difference everyone already knew about. It missed a smaller variance that was the actual story that quarter—a vendor contract renegotiation that was going to change the cost structure for six months.

That’s the part of this job people keep forgetting to talk about.

Agents are good at the parts we already automated in spirit

Pulling actuals, running the comparison, drafting a first-pass narrative—we’ve been “automating” this for a decade with templates, macros, and increasingly capable BI tools. Agentic AI is a real step change in how much of that chain it can chain together without a human clicking “next.” That’s genuinely useful. I’m not writing this to dunk on the tooling.

But the reason FP&A analysts still exist isn’t that nobody built the report yet. It’s that someone has to decide which three things matter this month, and that decision depends on context an agent doesn’t have: which VP just got a new mandate, which cost center is about to get reorged, which “one-time” charge is actually the fourth “one-time” charge this year.

The uncomfortable question

If most of what we ask junior analysts to do is “gather, format, draft”—yes, an agent will eat that, and probably should. The uncomfortable question for finance leaders isn’t “how do we deploy agents,” it’s “how do we teach the judgment layer to people who are about to get a lot less practice doing the mechanical layer that used to be how they learned it.”

I don’t have a clean answer. I’ve started deliberately assigning the “why does this matter” question to the newest person on a project, even when the agent already produced a clean first draft—specifically so the muscle gets built somewhere. It’s slower. It’s also the only thing I’ve found that actually addresses the real risk, which isn’t job loss, it’s a generation of analysts who are excellent at prompting and thin on the pattern-recognition that used to come from doing the boring version of the job for two years.

What I’d actually pilot first

If you’re a finance leader looking at agentic tools right now, my honest recommendation: don’t start with the highest-visibility, highest-stakes workflow (board deck commentary, external guidance). Start with something recurring, lower-stakes, and reviewable—like flagging unusual vendor invoices for AP review, or drafting first-pass budget-vs-actual notes for internal cost-center owners who will catch errors fast because it’s their own number. Let the agent build a track record where the cost of being wrong is a five-minute correction, not a restated external number.

The agent will get faster. The judgment layer won’t get faster just because the mechanical layer did—if anything, it needs more deliberate investment now, not less.

---

This is one of a handful of pieces I’m writing from direct experience rather than summarizing a vendor report—if you want the survey-data version of “where CFOs are placing bets,” that’s covered in the Finance AI strategy hub.