What AI doesn’t replace in CFO work.

The most overrated AI use case in finance right now is “AI CFO.” It's a product category I'd avoid both as a buyer and as a builder.

Here's the honest read on what's working and what isn't.

What AI is actually good at today

• Transaction categorization. Categorizing 90%+ of routine expense transactions to the right GL account, with the same accuracy as a junior bookkeeper but at 10x speed. Real product, real ROI.
• Anomaly flagging. “This vendor charged 4x more than the trailing 6-month average — worth a look.” Pattern-matching against a known baseline. Excellent at this.
• First-draft document generation. A board deck draft from your actuals. A weekly briefing. A variance commentary. These are first drafts — they need editing — but they save the analyst hours per week.
• Question answering on structured data. “How much did we spend on AWS in Q1?” Across the books, fast, sourced. A version of this lives in the financial-os dashboard and gets used dozens of times per day per client.

These four are real and shipping. If your AI finance tool does these, it's useful.

What AI is not replacing

The hire / no-hire call. “Can we afford this hire?” looks like a math question. It isn't. It depends on conviction in the pipeline, runway risk tolerance, the team's bench, the founder's ambition, and whether the role is replacing a current bottleneck. Mostly judgment, partly data. AI helps surface the data; the call is the operator's.

The pricing conversation. When to raise price, when to discount, how to structure a contract. The model has no view on the customer's pain, your competitive positioning, or what the room felt like in the last sales call.

The fundraising decision. Whether to raise, how much, from whom, at what terms. The model can simulate scenarios; it can't tell you whether your investor is going to support the next round or check out.

The trust relationship with the founder. Most CFO work is a multi-year conversation about the business — what's working, what's bothering them, what they're afraid of. That conversation doesn't happen with an LLM.

The framing I'd use

Treat AI in finance the way you'd treat a strong analyst hire: terrific at the rep work, useful as a thinking partner, not the person you'd hand the strategic call to. Give it the tasks that have an answer. Reserve the calls that have a decision for the operator.

The fractional CFO with AI tooling is faster, sharper, and cheaper than the fractional CFO without. The full-AI CFO is a category that doesn't exist yet, and probably shouldn't for a while.