AI is eating the work that trains your next CFO
More than 80% of the code Anthropic ships is now written by its own AI. Software engineers are the ones sweating about that number. Finance should be reading it more carefully, because the same shift is already here, and it arrives disguised as good news. AI will run your reconciliations, post your journal entries, and write your variance analysis.
Every CFO I talk to wants exactly that (I would too), and Gartner expects 90% of finance functions to run at least one AI tool by 2026. Almost none of them ask what comes next. If the machine does the work a junior analyst used to cut their teeth on, where does the next controller come from?
The boring work was always the training
Nobody loves reconciliations. The tie-outs, the journal entries, the variance walk that does not foot until 7 p.m. Everybody wanted that work gone. We forgot it was doing a second job. It was teaching a 23-year-old where the numbers come from, where they hide, and what a wrong one feels like in the hand.
You do not learn professional skepticism from a textbook. You learn it by doing five hundred reconciliations and getting burned by the five hundred and first.
I ran a consulting firm for twenty-five years, and nobody on my team became a good modeler by watching. (The first real model I built had the rows and columns transposed. I have never made that mistake since, which is precisely the point.) They got good by building a bad model, having a partner shred it, and building a better one.
Do not take just my word for it. The Journal of Accountancy ran a March 2026 cover story under the headline “How will accountants learn new skills when AI does the work?” When the AICPA’s own magazine is asking the question, it is not a fringe worry. PCAOB chair Erica Williams said the apprenticeship that builds judgment was already eroding, and that AI compounds it.
Somebody still has to know when the AI is wrong
The standard reassurance is that people shift from doing the work to reviewing it. Reviewing it with what, exactly? You cannot catch an error you never learned how to make.
IBM Research tested frontier models as code reviewers and found they caught about 45% of the errors in another model’s work. That is the machines grading the machines, and they missed more than half. We are now betting that humans do better than that while doing less and less of the underlying work themselves.
In finance the downside is not a failed unit test but a restatement, an audit finding, a board deck built on a number nobody can trace. A controller who never did the reconciliation is a smoke detector with the battery pulled out: installed, certified, and silent at the exact moment it matters. Elizabeth Mason, a firm CEO quoted in that same article, said it plainly. Shipping polished AI output fast “raises the stakes for judgment,” and the juniors may not be able to review it.
The pipeline was already running dry
All of this would be survivable if finance had people to spare, and it does not. Accounting degrees fell 6.6% in the most recent year on record, the lowest count of college accounting graduates in two decades. New CPA exam candidates dropped from about 42,600 in 2023 to roughly 28,100 in 2024. Over the same horizon, the Bureau of Labor Statistics projects around 120,000 accounting and auditing openings a year.
So the arithmetic is ugly. Fewer people are coming in, and the ones who do arrive will get less of the hands-on work that turns a hire into a senior. (We are narrowing the funnel and pulling up the on-ramp at the same time. Even my Cowboys do not draft that badly, and that is saying something.)
In May 2026, PwC made it concrete. It stood up a business unit it literally named the “Office of the CFO,” built on Anthropic’s Claude, and set out to train thirty thousand of its people on it. The early tasks are journal entries, variance analysis, and annual planning, which is the exact work that used to fill somebody’s first year.
To be fair to the optimists, the labor data has not caught up to the fear. The Yale Budget Lab finds no measurable AI dent in white-collar employment yet. But “yet” is carrying a lot of weight in that sentence, and the close runs every month whether the statistics have noticed or not.
Design the apprenticeship on purpose, because it will not happen by accident
For forty years, training was a free byproduct of the work. AI just unbundled the two. The judgment still has to be built, and it will not build itself anymore. So build it on purpose. Three moves I would make tomorrow:
- Make juniors review the AI’s output and defend that review out loud to someone senior. Reviewing is the skill now, so teach it like one instead of assuming it.
- Rotate people through the work the AI could have done anyway, early and deliberately, the way pilots still hand-fly with the autopilot off. The reps build the instinct, and the instinct is what catches the confident, wrong number.
- Promote on judgment, not throughput. If the only thing you measure is tasks closed, the AI wins every contest and your bench quietly rots.
None of this is free. The cost of the apprenticeship used to hide inside the grunt work, where nobody had to budget for it. Now it is a visible line item, and it is yours to fund.
AI is the best thing to happen to the finance function in my lifetime, and I mean that without hedging. It is also quietly dismantling the ladder that produced every person who runs that function today, and both of those are true at the same time. The CFOs who win the next decade will be the ones who notice that the monthly close was also a classroom, and who keep teaching after the textbook learns to teach itself.
So put this on your own board deck. If the AI does all the work your juniors used to learn from, who signs the statements in 2040, and what will they actually know?