AI Can Build a DCF in Ninety Seconds. That Was Never the Hard Part.

Two years of building AI tools for valuation work taught me that the model was never the bottleneck. Judgment was. And judgment is exactly what is getting harder to learn.

A discounted cash flow model that used to eat a full working day now takes a few minutes. Feed a tool the 10-K, the last four earnings transcripts, and a sector, and it returns a three-statement model with a debt schedule, a WACC build, and a sensitivity grid already wired up. I have built versions of this. I use them. They work.

And none of it changed the part of the job that actually matters.

What the tools are genuinely good at

I want to be specific, because most writing on this is either hype or denial.

The tools are excellent at the work that used to consume junior analysts and taught them almost nothing. Pulling line items out of a 10-K. Standing up the skeleton of a three-statement model. Writing the first draft of a formula you could write yourself but would rather not. Drafting the boilerplate commentary that wraps around an equity research note. Toggling scenarios faster than you can think of them.

This is the grunt work, and AI does it well enough that fighting it is pointless. An analyst who refuses to use these tools in 2026 is choosing to be slower than the analyst sitting next to them, for no reason.

Where they break, every time

Here is what does not show up in the productivity decks.

An AI will hand you a model with a balance sheet that ties. It ties because a plug somewhere absorbed an error you cannot see, not because the logic is sound. The most dangerous output is not a model that fails to build. It is a model that builds cleanly, looks institutional, and is quietly wrong in a place you would only catch if you had built one by hand a hundred times.

The tool will pick a comp set that looks defensible and is not, because it matched on industry labels rather than on business model, capital intensity, or growth profile. It will accept a terminal growth rate that produces a valuation no buyer would ever pay, and report it with the same confidence it reports everything else. It will give you a beta without a view on whether the comparable companies actually share your subject’s risk. The error hides inside a formula you did not write, which is worse than no model at all.

None of this is the AI being bad at its job. It is the AI doing exactly what it was asked, which is to produce something plausible. Plausible and defensible are different words, and the gap between them is the entire job.

The shift nobody is pricing in

So the work is moving. It is moving away from production and toward judgment. The analyst’s value used to sit in the ability to build the thing. Now it sits in the ability to look at a thing that was built in ninety seconds and know, on the numbers and in the gut, whether it is right.

That sounds like good news for senior people and it is. It is brutal for junior ones.

Here is the problem the headlines miss. You used to learn judgment by doing the grunt work. You built the schedule by hand, you broke the model, you fixed it, and somewhere in the thousandth reconciliation you developed a feel for when a number is lying to you. Remove the grunt work and you remove the apprenticeship. The juniors entering finance now are being asked to exercise judgment earlier, with fewer of the reps that used to produce it. The learning curve did not get easier. It got steeper and shorter at the same time.

Analyst classes are shrinking. The competition for the seats that remain is harder. And the analyst who cannot tell a defensible model from a plausible one is now competing against one who can, with neither of them having built enough by hand to be sure.

What to actually do about it

If you are early in finance, the instinct to outsource the tedious work to AI is correct and also a trap. Use the tools to go faster, then force yourself to audit what they produce as if a managing director is about to tear it apart, because one will. Select a row in an AI-built model and check that every formula in it is identical. Trace whether the schedule labelled “debt schedule” is actually calculating debt. Rebuild a model by hand once a quarter, not for output, but to keep the muscle that lets you catch the errors the tool will not.

If you are hiring, stop testing whether candidates can build a model. The tool builds the model. Test whether they can find the three things wrong with a model you built to be subtly broken. That is the skill that is now scarce, and it is the one your interview process is probably not measuring.

The part that does not get automated

The most valuable person in finance over the next decade is not the one with the best prompts. It is the one who can stare at a clean, confident, institutional-looking output and say, that terminal value is wrong, that comp does not belong, that margin assumption has never held for a company at this scale.

That skill does not come from a tool. It comes from reps, from having been wrong enough times to recognize the shape of a wrong answer before the math confirms it. AI made the model free. It made the judgment to evaluate the model more valuable than it has ever been.

The spreadsheet is not dying. The analyst who trusts it blindly is.

I have spent the last two years building AI tools for financial modelling. If you are doing the same, or arguing with me about any of the above, I would like to hear it.