Who Trains the Next Controller When AI Does the Junior Work?

In Silicon Valley Bank’s 2026 survey of more than 200 CFOs at venture-backed companies, 35% of finance leaders said AI productivity gains had led to fewer junior hires. Only 5% had made AI-related layoffs. I take venture-backed companies to be early adopters, so I read this as a leading signal. In this data, the cuts are not coming as layoffs. They are coming at the bottom, as junior hires that never happen.

Good controllers can often point to the wrong line on a trial balance before they can say why it is wrong. That is rarely learned in a classroom. It comes the slow way: years of reconciliations that would not tie, and a senior colleague who shows you what you missed. AI is now taking that work. So who trains the next controller?

This is not only a career question. It decides which finance teams can still check their own numbers, and what software they need to do it.

The volume was the tuition

Finance does two jobs. The work layer produces the numbers. The decision layer judges what to do about them. The second was always the point, but nobody started there.

The path in was never designed as a training program, but it worked like one. A new analyst got the reconciliations the seniors did not want, ran hundreds of them, and moved up once the mistakes stopped. Twelve closes a year taught you which accounts drift and which entities cause trouble.

Now the bottom rungs are going. Dost, for one, builds AI that reads, matches and posts supplier invoices, so finance teams handle only the exceptions. Matching and coding invoices is work juniors learned on. That is good news for productivity. It is bad news for the ladder, because those rungs were where judgment got made.

Finance jobs are not disappearing. Junior work is.

In my view, the forecast that AI would empty finance departments has not held up. What it removes is tasks, and the junior ones go first. The rest of the data points the same way:

• Revelio Labs and Ramp found that companies investing most in AI grew employment by more than 10% after adopting it, and increased entry-level hiring. The authors add that this does not rule out displacement in particular roles or tasks.
• The Bureau of Labor Statistics projects employment of bookkeeping, accounting and auditing clerks to fall 6% from 2025 to 2035, about 85,600 jobs. Over the same decade it projects accountants and auditors to grow 5% and financial analysts 7%.
• In a field study by Jung Ho Choi of Stanford and Chloe Xie of MIT Sloan, accountants who adopted AI moved about 8.5% of their time from routine data entry toward work such as business communication and quality assurance.

The Revelio caveat matters most. A company can keep hiring juniors and still stop giving them the work that used to train them.

In March I wrote that finance jobs would end. By July I was calling the job-loss story overstated. Both were half right. The work layer is getting thinner, and more weight falls on the people trusted to judge what it produces.

Clean output hides wrong numbers

As AI produces more of the numbers, the core skill in finance moves from calculating to validating. AI drafts the reconciliation, the forecast and the commentary. A person decides whether to believe it.

That would be fine if bad numbers still looked bad. They don’t. A wrong answer used to arrive in a messy spreadsheet with a broken link, and the mess was the warning. Now it arrives with clean formatting, a confident summary and a chart.

The Choi and Xie study saw both sides of this. When the AI suggested a different category for an uncertain transaction, accountants tended to go along with it. The paper also reports that accountants with more expertise stepped in more often when the AI’s confidence scores signaled doubt.

Knowing when an answer is wrong means knowing how the number is made. Which accruals are judgment calls. Which intercompany balances never quite tie. Which large customer always pays in the last week of the quarter. That knowledge came from building the numbers by hand, the very work now disappearing. You cannot check a number you have never made.

Start juniors on the exceptions

The answer is not to keep juniors on manual work. That wastes the time AI gives back.

Judgment comes from decisions that get corrected. You make a call, someone who knows more tells you why it was wrong, and you remember. Exceptions are where those moments cluster, so that is where juniors should now start. When an agent flags a variance, the first explanation can come from the newest person on the team and the verdict from the most senior. I think a year of arguing real exceptions teaches more than a year of tying out clean accounts.

Hiring shifts too. Earlier this year I asked CFOs what they look for in hires. The answer, every time, was curiosity. Revenue recognition can be taught. The habit of asking why a number moved, before anyone asks you, mostly cannot.

The model has one constraint: senior attention. A senior can only correct what they see.

The exception queue is the classroom

That constraint turns the talent question into a software question.

Finance AI, I think, is too often judged by its output: the close finished, the deck drafted, the commentary written. In a team built on exceptions and senior correction, the surface that matters is the exception queue and the review record. Every override a controller makes, with the reason attached, is a lesson for the junior who sees it and a signal for the agent that made the call. Over time that record becomes the team’s memory: why an accrual is booked the way it is, which reconciling item comes back every quarter. Without it, that memory lives in people’s heads and leaves with them.

That splits finance AI into two kinds of product:

• One hides the process and hands over a number. The number may be right, but the product trains no one.
• The other shows what the agent assumed, where it was unsure and where each figure came from. That is the record the person who signs needs before trusting a number, and it is what a junior learns from.

As teams lose their bottom rungs, I expect buyers to favor products that show their work, because those products do a job the career ladder used to do.

Who gets trained is a design choice

In the work layer, finance probably needs fewer people, though headcount is a poor way to measure what it needs. What it cannot run short of is people who can tell when a clean-looking number is wrong, and the future CFOs who will one day sign the accounts.

Those people used to appear by accident, as a side effect of volume. With the volume gone, I don’t expect the accident to repeat.

Tomorrow’s juniors will tie out far fewer accounts. Whether they still get a senior who shows them what they missed is a design choice: who sees the exceptions first, what the software keeps, where senior time goes. I suspect many finance teams are making it now without noticing.

Wouter Born is the founder and General Partner of Born Capital. He writes AI CFO Office, a newsletter on AI for CFOs and finance leaders with more than 50,000 subscribers.