A CFO at one of the CFO Alliance roundtables in Texas in early November 2026 described a financial model built with AI that missed a 6% payment-in-kind (PIK) interest strip on lender payments. PIK interest is added to the loan balance rather than paid in cash, so a missed strip understates the debt and overstates the cash in every period after it.
That is the AI in finance problem the rooms kept returning to. Speed is no longer the issue. Confidence is.
Fast, fluent and unchecked
Garbage in, garbage out is an old rule. AI amplifies it in a particular way: it presents a weak answer in the same polished form as a strong one. A model with the wrong interest treatment looks exactly as finished as a correct one.
Several CFOs in the room put it in terms of the Dunning–Kruger effect, the finding from Kruger and Dunning's 1999 studies that people with the least skill in an area tend to overestimate their performance the most. Later researchers have argued that part of the original pattern is a statistical artefact, so it is worth using carefully. The practical point stands without it. AI lets people produce work beyond their experience, and the people furthest beyond their experience are the least able to tell when it has gone wrong.
What the research says about the cost
The time saved by AI is being partly spent checking it. Three studies point the same way.
| Source | Sample | Finding |
|---|---|---|
| Datarails, 2026 CFO Sentiments Survey (October 2026) | 270 US CFOs and finance leaders at companies with $100M+ revenue | Finance teams spend 26% of their working week verifying or correcting output from AI tools |
| Workday, Beyond Productivity (January 2026) | 3,200 employees using AI at $100M+ revenue organisations | Nearly 40% of AI time savings are lost to fixing low-quality output |
| BetterUp Labs and Stanford Social Media Lab (September 2025) | 1,150 US desk workers | 40% had received AI-generated work that looked good but lacked substance in the past month, averaging about two hours to resolve each instance |
Treat the figures as indicative. Each is self-reported, and two of the studies were run by software companies. The direction is still consistent: some of the hours AI saves are going straight back into checking what it produced.
Why the repetitions mattered
Finance people used to learn by doing the same work many times. Tying out a balance sheet, rebuilding a model cell by cell, reconciling an account that would not reconcile. It was slow, and it built a sense of what a wrong number looks like.
AI now does a large share of that first pass. The answers arrive faster, and the repetitions that built judgment are missing. If you lack the domain confidence to push back, you will not catch the error. One participant described the risk of the great pretender: a first-time finance leader using an AI assistant in place of experience, producing work that reads well and has not been tested by anyone who knows better.
Agent checking agent still needs a person
Some teams are now using one AI agent to review another's work. It catches some mistakes, particularly formula and consistency errors. It does not replace a finance expert in the loop, because both agents can share the same blind spot. A missed PIK strip is not usually a formula error. It is a term in the loan agreement that never made it into the model, and it takes someone who knows to look for it to find it.
Controls that make AI in finance reliable
Name a reviewer. Anything AI produces that leaves the finance team, from a board pack to a lender model, has a named person who reviewed it and is accountable for it.
Tie back to source. Reconcile AI-built schedules to the ledger, the loan agreement or the contract, not to themselves.
Check the terms that matter most. For debt models: interest type, PIK, fees, covenants and amortisation. For revenue: recognition timing and contract terms.
Keep the hand-built skills alive. Have junior staff build some schedules manually, so they know what the AI is supposed to produce.
Measure the checking time. If verification is eating the savings, the process needs redesign, not more AI.
What better tools will not fix
They will not fix a team without the experience to review the output. A more capable model makes errors less often and makes the remaining ones more convincing.
They will not remove accountability. The lender, the board and the auditor will hold the finance function responsible for the number, whoever or whatever produced it.
Frequently asked questions
What is the main risk of using AI in finance?
Confident, wrong output that nobody catches. AI presents a weak answer in the same polished form as a strong one, and people without enough experience may not notice the difference.
How much time do finance teams spend checking AI output?
In Datarails' 2026 survey of 270 US CFOs and finance leaders, finance teams reported spending 26% of their working week verifying or correcting AI output. Workday's 2026 research found nearly 40% of AI time savings were lost to fixing low-quality output. Both are self-reported.
Can one AI agent check another's financial work?
It can catch some errors, especially formula and consistency mistakes, but it does not replace a finance expert. Both agents can miss the same thing, such as a term in a loan agreement that never made it into the model.
How should finance teams review AI-generated models?
Assign a named reviewer, tie every schedule back to source documents and the ledger, check the terms that drive the numbers most, and track how much time verification takes.
What is the Dunning–Kruger effect and why does it matter for AI?
It is the finding, from Kruger and Dunning's 1999 studies, that people with the least skill tend to overestimate their performance most. Researchers still debate how much of it is a statistical artefact. For AI, the practical point is that the people most able to produce work beyond their experience are the least able to check it.
Where CFOLogic fits
Our teams use AI in the close, reconciliations and reporting, with a doer, a reviewer and a QA check on every engagement, and a finance professional accountable for what leaves the team. See Finance Team as a Service.