The Verification Tax: Why Checking an AI Financial Model Costs More Than Building One

Two numbers from 2026 do not fit together.

The first: an AI model can produce a three-statement forecast in under a minute. Revenue build, cost structure, working capital, a cash flow that looks finished.

The second: finance professionals spend close to 13 hours every week reconstructing, validating and defending AI outputs. In the US, 49% of them report 15 hours or more, and 19% report 30 or more. Those figures come from an IDC survey of 2,275 senior finance decision-makers across North America and EMEA, fielded in February 2026 and published in July as The Emerging Economics of AI in Finance (IDC for Sage, 2026, summarised by ERP Today). The research is vendor-sponsored, which is worth knowing when you read it, and the sample is large enough that the order of magnitude is hard to dismiss.

Put the two together and the productivity story inverts. Generation went to zero. The work moved somewhere else, and nobody put a line item on it.

That line item is the verification tax.

What the verification tax is, and what it is not

It is not "AI outputs need review". Finance has always reviewed. A junior analyst's model gets checked. A statutory account gets a second pair of eyes. A ledger entry gets an approver. Review is a control, not a cost overrun, and any argument that treats human review as a defect misunderstands the function.

The tax is something narrower and more specific: the review is undirected.

When a colleague hands you a model, the review has a shape. You know what they touched, because they tell you. You know their habits. You can ask "why is D42 hardcoded" and get an answer in ten seconds. The artefact comes with a person attached, and that person is a queryable index of the changes.

When a model comes back from a generation step, none of that exists. Twelve numbers changed out of four hundred and nothing marks which twelve. So you check all four hundred. Not because you distrust the tool more than you distrust a junior, but because you have no way to aim.

That is the whole mechanism. The tax is not the existence of the check, it is the absence of a target.

Why an AI artefact is more expensive to check than a junior's

There is a fair objection here: a junior's model also has to be re-derived, and juniors also hardcode. True. Three things make the AI case structurally worse.

The errors pass the eye test. Wall Street Prep ran the four leading AI modelling tools head to head in 2026 on an integrated three-statement model. The pattern in the failures was not obvious nonsense. It was individual lines that were wrong while the totals they fed added up correctly (Wall Street Prep, 2026). An error that reconciles is invisible to the scan you would normally rely on, so the only review that catches it is cell by cell.

Complexity degrades accuracy faster than it degrades confidence. On FinSheet-Bench, models averaged 89.1% on simple lookups and 19.6% on complex aggregation tasks, with the tone of the answer unchanged between the two (Ravnik et al., 2026). The output does not signal when it has entered the regime where it is usually wrong. More figures of this kind are collected on our AI financial model risk statistics page.

There is nobody to ask. Three sessions later, the reasoning behind a line is gone. Not hidden, gone. A junior can reconstruct their thinking; the session cannot. This is the part that converts a five-minute question into a forty-minute reconstruction, and it is the part that compounds every time the model is touched.

This is also why the strongest consensus in the Financial Modeling Institute's 2026 survey of its Global Leaders Council was on review, not on accuracy: 0 out of 63 senior modellers said they would rely on an AI-generated model for a high-stakes decision without independent human review. We covered that survey in Can You Trust an AI-Generated Financial Model?.

The number that reframes the problem: 71%

The most useful figure in the IDC research is not the 13 hours. It is this one:

71% of finance leaders say they would reject an AI tool that is 99% accurate if it cannot explain its answers.

Read that carefully, because it kills a whole category of vendor roadmap. The market is not waiting for the accuracy curve to bend. A tool that gets to 99% and stays opaque gets rejected by seven finance leaders out of ten, and 54% of organisations in the same study say they would pay more for visibility into how an output was produced.

The constraint is not correctness. It is whether the output arrives with its reasoning attached in a form a human can read and a reviewer can challenge.

Which is a relief, in a way. Accuracy is somebody else's research problem and it moves on its own schedule. Explainability is an artefact design problem, and it is available today.

Where the tax is permanent, and where it amortises

This is the part most takes on the subject get wrong, and getting it wrong is how the argument gets dismantled by the first practitioner who has actually industrialised a workflow.

On a repetitive process, the tax amortises. Teams automating revenue recognition or a monthly close describe the same pattern: replay the script over every prior month, compare line by line against the existing entries, then run in parallel for two to three months before switching over. That is a heavy verification cost, paid once. After the cutover the process runs, and often the automation surfaces errors the manual process had been missing for years. Anyone telling you verification is a permanent drag on this kind of work is describing an investment as if it were a subscription.

On modelling work, it does not amortise. A deal is new. A budget is new. A downside scenario built on a Tuesday afternoon because the lender asked is new. There is no second run to spread the cost over. Each artefact is verified once, used once, defended once, and the cost lands whole every time.

That distinction is the honest version of the argument, and it happens to be exactly the work most finance modelling actually is.

What makes verification cheap

If the tax comes from the review being undirected, the fix is to make it targeted. Four properties do that, and none of them require a particular tool:

1. Reasoning attached to the number. Every input carries what it is, what unit it is in, where it came from, and who set it. Not in a comment nobody reads, in the artefact itself. This is what turns "why is churn 2.1%" from a research task into a lookup.

2. Dependency traceability in both directions. From any number you can go up (what feeds this) and down (what breaks if I change this). Excel gives you precedents and dependents cell by cell, which works until the model has thirty tabs and the trace crosses a lookup table. What you need is the same answer for a variable, not a cell.

3. A diff between two versions. The single highest-leverage property. If you can see that this revision changed twelve values and one formula, you review twelve values and one formula. Without it, every review is a full review, forever. This is the difference between reading a pull request and re-reading the codebase.

4. Checks that run on every change. Balance sheet ties to zero. Cash flow closing equals balance sheet cash. No hardcoded constant sitting inside a row that is supposed to be a formula. Sign conventions hold. These are mechanical, they should never be a human's job, and they catch precisely the class of error that passes the eye test.

A model with those four properties still gets reviewed. It gets reviewed in twenty minutes instead of an afternoon, and the twenty minutes are spent on judgment rather than reconstruction.

A verification pass you can run this week

No new tooling required. This is the sequence, in the order that finds the most per minute:

1. Run the mechanical checks first. Balance check at zero on every period, cash flow closing tied to the balance sheet, no negative where a negative is impossible. If any of these fail, stop. Everything downstream is unreliable and detailed review is wasted effort.
2. Hunt hardcodes inside formula rows. In Excel, select the row, Ctrl+G then Special then Constants. Anything a constant that should have been a formula is the single most common AI failure and the hardest to spot by reading.
3. Trace three numbers to their source. Pick the three that carry the conclusion: the exit value, the covenant line, the funding need. Follow each one back to a named assumption. If the trail dies in a cell with no story, you have found the thing you cannot defend in the meeting.
4. Re-test the model, do not re-read it. Change one driver by 10% and read the outputs. Numbers that do not move are the hardcodes you missed in step 2. Numbers that move absurdly are broken links. This finds in ninety seconds what an hour of reading misses.
5. Write down what you checked. Three lines at the bottom of the file. Next month's review starts from there instead of from zero, and that is where the amortisation actually happens on non-repetitive work.

If the model is going to a board, a lender or a buyer, the audit trail your PE board will actually ask for is the next step up from this.

What this means for how you buy

Two things follow for anyone choosing a tool in this space.

Stop scoring vendors on generation speed. Every tool in the category is fast, the benchmarks agree on that, and the fast part was never where the hours went. A demo that impresses you in ninety seconds is telling you about the cheap half of the job.

Score them on what happens on the second visit. Can it tell you what changed since last week. Can it explain a number without you opening the file. Does the reasoning survive a new session with a different person. Those questions map directly onto the 13 hours, and they are the questions the market has apparently already decided matter more than the last percent of accuracy.

Where Layerz sits in this

Layerz keeps a financial model as a structure separate from its data. The logic (what depends on what, which assumption drives which line) is versioned and addressable; the values sit on top of it.

The practical consequence for verification: you get a diff between two versions, so a review can be aimed at the twelve things that changed. Every value traces back to a named assumption in both directions. The mechanical checks run on every change rather than the night before the board pack. And because Claude drives the model from outside through MCP rather than rewriting a grid, the reasoning behind a change is recorded as a change, not lost with the session.

None of that removes the human review, and it is not meant to. The 0 out of 63 modellers were right. What it removes is the reconstruction work that has to happen before the review can even start.

The Excel export stays a clean, standard, auditable workbook with live formulas, because that is what actually gets sent to the other side.

Further reading: AI Financial Model Generator: What These Tools Actually Produce · Can You Trust an AI-Generated Financial Model? What 63 Expert Modellers Say · How to Build an Auditable Financial Model with AI · The Financial Model Audit Trail Your PE Board Will Actually Ask For · AI Financial Model Risk and Governance: Statistics

Layerz keeps a financial model as structure separate from data, so a review can be aimed at what changed instead of starting from scratch. Excel export is clean, standard, and never paywalled. Explore Layerz →