Financial Modelling & Model Quality

The 10-Minute Financial Model Exists. It Just Isn’t Finished.

Sebastian Szameitat, AFM

Originally published 5 August 2026

AI vendors sell machine minutes. Finance pays for human hours.

I have good news and bad news.

The good news is that AI can build a financial model in ten minutes, and less. Yes, you read it, I wrote it.

The bad news is that the stopwatch appears to have stopped rather earlier than the work did.

DCF models in five minutes. LBO models in eight. Three-statement models in ten. Investor-grade. Error-free. Ready to use. Ninety-five per cent time saving.

These are not straw men. They are published market claims.

And, inconveniently for those of us who have spent rather a long time learning how to build financial models, the first few minutes may actually be real.

AI can produce an impressively complete-looking workbook in minutes. It can create tabs, formulas, schedules, forecasts and formatting before a human analyst has finished deciding whether the coffee is helping.

What those minutes do not tell us is how long it takes to produce a model that somebody understands, trusts and is prepared to sign off.

The 10-minute financial model exists. It just isn’t finished.

AI vendors sell machine minutes. Finance pays for human hours.

An awkward concession

Let us begin by giving the technology its due: the speed is remarkable.

Shortcut’s own marketing page advertises a DCF in 5–10 minutes rather than 4–8 hours, an LBO in 8–12 minutes rather than 6–10 hours, and a three-statement model in 7–10 minutes rather than 5–8 hours. The same page claims formula accuracy above 90%, changes in two or three minutes, and an incremental cost of USD 0–5 per model compared with USD 300–600 of loaded human cost.

V7 describes a narrower historical-spreading and template-population workflow as reducing one or two days to around 20 minutes, or 95%. Openbox promoted a custom model in under five minutes. Ajelix uses the phrase “investor-grade” for an Excel modelling workflow completed in under ten.

These products do not all perform the same task. Their inputs, templates, use cases and definitions of completion differ. We should not pour the numbers into a large bucket, stir gently and call the result science.

But the underlying capability is no illusion.

In Wall Street Prep’s 2026 comparison, Shortcut and Claude produced the initial setup of an Apple three-statement model in roughly 15 minutes. Copilot took around 25 minutes. ChatGPT took close to an hour. All were faster than the stated human baseline for the initial assignment.

In Ankura’s test, five tools each produced structurally complete models for three finance scenarios in 7–10 minutes.

In a live financial-modelling test summarised by CFO Connect, Claude produced a five-year model in approximately 15 minutes, including scenarios, three statements and debt logic.

So yes: minutes to a first workbook are real.

This matters. A great deal.

The mistake begins when we treat that result as the end of the modelling process rather than the beginning of a different one.

The wrong stopwatch

A productivity comparison requires two equivalent measurements.

If the human side includes understanding the assignment, collecting data, choosing assumptions, designing the architecture, building the model, checking it, correcting it, documenting it and preparing it for use, then the AI side must include the equivalent work.

Instead, many headline comparisons place broad human labour on one side and narrow AI generation time on the other.

In that comparison, the human clock begins with the assignment. The visible AI clock begins after some combination of data, requirements, assumptions, templates and instructions has arrived, presumably delivered during the opening credits.

The human clock stops when the model is usable. The AI clock stops when a file appears.

Those are not the same clocks.

Shortcut’s own LBO example makes the distinction unusually clear. The page describes a full LBO model generated in 12 minutes. It then says that an associate spends two hours validating the model and adjusting assumptions.

The twelve minutes are real. They are simply not the total human effort required to reach the described result.

The model was built in twelve minutes. The next two hours appear to have been placed in a different column.

That does not make the workflow unattractive. Two hours after a rapid first build may still be substantially better than two days from scratch. It may be a very good result.

But it is not a 12-minute end-to-end model.

The distinction becomes even more important when a provider claims a 90% or 95% saving. A percentage is reassuringly precise. It wears a tie. It looks as though somebody has measured something.

Yet before accepting it, we need one very simple answer:

If you claim a 95% time saving, show us what was counted in the original 100%.

Was the input data already clean? Was a proprietary template supplied? Who wrote the instructions? Who answered clarifying questions? Who checked the historical data? Who reviewed the formulas? Who repaired the debt schedule? Who documented the changes? Who tested the next update? Who signed off?

And how many active human hours did all of that require?

Without those numbers, we have a machine-runtime claim. We do not yet have a productivity case.

What exactly is a first draft?

One reason this debate becomes slippery is that “build a model” can mean almost anything.

It may mean creating a revenue forecast from a clean 100-row dataset. It may mean populating historical financials into an existing template. It may mean building an integrated three-statement forecast. Or it may mean producing a transaction or project-finance model with debt sizing, circularity, taxes, reserve accounts, covenants, waterfalls and returns.

All are legitimate tasks. They are not interchangeable.

The same problem applies to the word “finished”.

A workbook may have all the expected tabs and still contain incorrect historical data. It may balance only because a plug is hiding somewhere. It may contain a debt schedule that is, in a rather philosophical sense, neither a schedule nor debt logic. It may calculate correctly today but fail as soon as an assumption changes. It may be numerically plausible and structurally unreviewable.

There is a meaningful progression:

a workbook has been generated;
its formulas and basic mechanics have been reviewed;
its data, assumptions and economic logic have been validated;
its controls, documentation and maintainability have been tested;
a qualified person is prepared to accept responsibility for its intended use.

The first of these can happen in minutes.

The last is what Finance is eventually paying for.

Wall Street Prep calls moving a model from zero to roughly 60% complete a good use case for current tools. That is a useful and refreshingly honest description.

It is also worth asking what the 60% measures.

Sixty per cent of visible output is not necessarily sixty per cent of professional effort. It is not sixty per cent of risk removed. And it is certainly not sixty per cent of responsibility discharged.

The remaining forty per cent may be smaller on the screen and considerably larger in the risk register.

That last sentence is my interpretation, not a benchmark result. Nobody has yet published a universal allocation of the remaining work. That absence is not a minor inconvenience. It is the central finding.

Fast, impressive and quietly wrong

Wall Street Prep tested Shortcut, Claude, Microsoft Copilot and ChatGPT on the same Apple three-statement assignment. The results are useful precisely because they are neither triumphalist nor dismissive.

Shortcut and Claude were very fast and looked most like analysts. But the highest overall tool score was 5.9 out of 10. Wall Street Prep’s lower-performing human-analyst benchmark scored 6.4.

All four tools scored zero on circularity.

More troublingly, Shortcut and Claude introduced errors into historical data. The errors were not always spectacular. They were subtle, plausible and capable of surviving a casual glance. Wall Street Prep concluded that checking the affected cells carefully could take longer than entering the numbers manually.

This is an important kind of failure. An obvious error is annoying. A plausible error gets invited to the meeting.

It sits neatly in the workbook, wears the correct formatting and waits for a decision to be made from it.

Ankura’s study tells a similar story from another angle. Five tools were tested against a long-range plan, a 13-week cash-flow forecast and a customer-profitability analysis. The researchers prepared fictional company data and detailed prompts of more than 500 words.

Every tool produced a structurally complete model in 7–10 minutes.

None produced a production-ready model.

Ankura found fundamental formula-mechanics problems or incomplete and inconsistent structures in every output. Meaningful human rework remained.

There is an important limitation: Ankura deliberately used a single pass, with no iterative prompting. The firm explicitly says that iteration would likely improve the results.

That is fair. It is also revealing.

Once we add iteration, we have reintroduced human time. Someone must inspect the output, identify what is wrong, explain the required correction, assess the new version and decide whether another round is necessary.

That may still be far faster than manual construction. But once again, the relevant measurement is no longer seven minutes. It is the complete human-assisted cycle.

The CFO Connect recap is even more direct. Claude’s approximately 15-minute model included a scenario structure, financial statements and debt logic. It also contained hardcoded historical figures, a weak debt structure and formulas that did not copy consistently across periods.

Their conclusion was admirably concise: a strong first draft, not a deliverable.

That is not a failure of AI-assisted modelling. It is a useful description of its present maturity.

The professional benchmarks are not feeling particularly festive

The gap becomes clearer as the tasks become more complex.

MBABench evaluates agents on end-to-end spreadsheet work in finance using professional criteria for accuracy, formulas and format. Its authors report that even the strongest agents frequently fall short of professional finance standards, with performance deteriorating sharply as complexity rises beyond a few chained calculations.

FrontierFinance goes further into long-horizon work. It contains 25 complex financial-modelling tasks across five core model types, requiring more than 18 hours of skilled human labour on average. Human experts achieved higher average scores and were more likely to produce client-ready outputs than the tested systems.

And then there is BankerToolBench, an open benchmark of 100 end-to-end investment-banking tasks developed with 502 bankers. The work includes Excel models, presentations and written deliverables, with human completion time averaging about five hours and reaching up to 21.

The best tested model failed nearly half of the rubric criteria. Bankers rated 0% of its outputs client-ready.

“Zero per cent client-ready” is not quite the same headline as “ninety-five per cent time saving”, although it does have a certain clarity.

We must be careful here. BankerToolBench is not a financial-model-only benchmark. Its result covers multi-artefact investment-banking workflows. Nor does FrontierFinance tell us how much human time an expert working with AI might save before sign-off.

These studies do not prove that AI creates no value. They do not even measure the most promising expert-plus-AI workflow directly.

What they do show is that rapid artefact generation has not yet removed the quality gap in complex professional work.

And they expose the most important missing number in the market:

How many active human hours does an AI-assisted workflow require to reach the same professional acceptance threshold as the human baseline?

The public evidence does not currently give us a reliable market-wide answer.

My own inconvenient experiment

This would be a simpler article if my own experience suggested that AI was useless. Unfortunately for a simpler argument, it suggests the opposite.

To test these claims against my own work, I have been building an AI-assisted project-finance model. The case is a Spanish solar-plus-battery project: 120 MWp of photovoltaic capacity, a 50 MW / 100 MWh battery, one construction year, 30 operating years and quarterly periods.

I did not want a polished little toy in which an AI changes the font colour, produces a chart and receives a round of applause from people who have never met a debt-service reserve account.

I wanted something uncomfortably close to real modelling work.

That meant a full specification, realistic operating logic, a catalogue of deliberate review issues, review tooling, version control, evidence chains and a model that experienced practitioners could inspect without having to lower their standards for the occasion.

At the measurement date, the project comprised:

18 controlled workbook versions;
approximately 20 worksheets;
around 24,473 formula cells;
approximately 150–200 generated or updated controlled file instances;

operating logic for construction, energy, battery operations, revenue, operating costs, tax, working capital and cash flow.

Twenty-four thousand formula cells is not proof of quality. It is merely proof that I have made certain lifestyle choices.

Here is the commercial baseline that matters.

Based on my own assignments, I typically allow 60–80 active human hours for a substantial new-build financial model developed in line with the FAST Standard – from blueprinting through build, testing, documentation and closure of agreed review comments.

It is not an industry benchmark. It is the baseline I can actually define and defend. By “substantial”, I mean an integrated, decision-ready model – not a template, a single-sheet forecast or a cosmetic update.

The AI-assisted case had already required approximately 35–55 active human hours, with 45 hours as the most plausible midpoint.

At that point, the Operating Model was approximately 90–95% complete. The complete Financial Model was only around 55–65% complete. Funding, debt sizing, drawdown, interest, fees, reserves, covenants, equity funding, distributions and returns still had to be built.

At the midpoint, I had therefore used 45 hours against a normal complete-job budget of about 70 — roughly 64% of the usual time. Yet the complete Financial Model was only around 55–65% finished.

The position at the measurement date was this:

Normal commercial baseline for a complete model: 60–80 human hours.
AI-assisted case so far: 35–55 human hours.
Complete Financial Model at the measurement date: 55–65% finished.
Demonstrated end-to-end saving: not yet measurable.

The assignments are not perfectly like-for-like, progress is not linear and this experiment included unusually deep governance and evidence work. The figures therefore do not prove that AI made the job slower. They do not prove that it made the complete job faster either.

Could the finished AI-assisted model still come in below my normal 60–80-hour range? Possibly. Could the remaining financing, returns and review work consume the difference? Also possibly.

Until the model is complete and the hours are measured, the honest finding is simpler:

Not measured. Not finished. Not proven.

By the ten-minute standard, I was running roughly 44 hours and 50 minutes late.

By any defensible end-to-end standard, the stopwatch was still running.

The distribution of time was more interesting than the headline total.

At the midpoint, the actual model-build phase accounted for about 19% of my active human effort. Review and validation accounted for approximately 27%. Iteration and evidence closure accounted for another 20%. Review, iteration and governance together represented roughly 56%.

AI clearly accelerated individual construction and iteration steps. But construction was not the majority of the AI-assisted human process, and this case does not yet demonstrate that the complete assignment was faster.

There is another revealing number. Of the 18 controlled workbook versions, approximately 89% primarily addressed correction, review, compatibility, governance or evidence rather than classic first-build activity.

Again, this does not prove that every AI project will look the same. My case was unusually controlled, and some of that control was part of the experiment itself.

But it does illustrate why “minutes to first draft” is an incomplete economic measure.

The first draft was no longer the bottleneck.

Understanding, correcting and proving the model remained the work.

The AI did not replace the modeller. It gave the modeller considerably more model to worry about in considerably less time.

Where the value actually appears

The most useful lesson from my case is not that the marketing claims are too optimistic.

It is that they may be looking for value in the wrong place.

AI’s contribution was not confined to writing formulas faster. It helped compress an entire loop:

specification → testable workbook → findings → corrections → retesting → evidence

It helped turn detailed requirements into working structures. It produced repetitive formula logic. It supported documentation, comparison, challenge and patching. It made it possible to test more versions and alternatives within the same period.

That is a better value proposition than “type one sentence and collect your investor-grade model before the kettle boils”.

The economic benefit may lie in several places:

faster movement from specification to something testable;
rapid implementation of well-defined changes;
cheaper exploration of alternative structures;
more complete documentation and model maps;
assistance in designing controls and review procedures;
shorter cycles between finding a problem and testing the repair;
the ability to perform more informed iterations before delivery.

Not all of these benefits have been independently quantified. Some are conclusions from my own documented case and from the patterns across the available tests.

But they point towards a more credible proposition:

AI does not eliminate the modelling process. It increases the number of informed iterations an expert can complete within it.

That can improve speed. It can also improve quality — if the saved time is used for better questions, deeper testing and clearer communication rather than merely for producing more unchecked files at industrial speed.

AI did not remove the difficult part.

It allowed us to reach it sooner.

The modeller was never merely typing

The arrival of AI exposes an uncomfortable truth about how financial modelling has sometimes been valued. If the value of a modeller were simply the ability to enter formulas into Excel, then yes, the outlook would be rather bleak.

Fortunately, formula entry was never the whole job.

The modeller defines scope. Chooses architecture. Identifies missing assumptions. Understands economic relationships. Decides whether an inconsistency is immaterial, dangerous or deliberate. Designs controls. Tests changes. Explains outputs. Maintains the model. And eventually accepts responsibility for the result.

AI can assist with every one of those activities.

It cannot make responsibility disappear.

This is why the democratisation argument needs care. AI absolutely lowers the barrier to producing spreadsheet output. That is useful and, in many situations, liberating.

But access to generation is not the same as competence in validation.

AI democratises spreadsheet generation faster than it democratises judgement.

The person receiving an AI-generated model still needs to know whether the historical data are correct, whether the cash flow links, whether the debt really amortises, whether the tax logic follows the stated rules, whether the circularity is controlled, whether a check is independent and whether a balanced balance sheet is balanced for the right reason.

The barrier to generation is collapsing.

The barrier to assurance is not.

That does not mean every model must be built or reviewed by somebody with a particular title or qualification. It means the required capability does not materialise simply because the workbook did.

A generated workbook is an output. A trusted financial model is a professional conclusion.

This is not a defence of billable hours.

It is a demand for honest measurement.

A modest proposal for anyone claiming 95%

If a vendor can genuinely reduce professional end-to-end modelling time by 90% or 95%, that would be an extraordinary result.

I would like to see it.

Not a highly edited demonstration. Not a stopwatch that begins after the data have been cleaned and stops before the first review comment. Not a comparison between AI marginal compute cost and the fully loaded cost of a human professional.

A reproducible test.

Publish the input pack. Describe the state of the data. Show the prompt, clarifications, templates, skills and connectors. Record the machine runtime and the active human setup time separately. Publish the first output. Record the review findings by severity. Include correction time, reruns and failed attempts. Define the acceptance criteria before the test. Then report the total human hours to sign-off.

Use the same task and the same quality threshold for the human comparison.

And please include the awkward runs as well as the one that behaved beautifully while the camera was on.

This would not slow progress. It would allow the market to distinguish genuine productivity from impressive generation.

It would also give vendors something much more valuable than another spectacular claim: evidence that finance teams can use for investment decisions.

Until then, the burden does not sit with sceptics to prove that the saving is impossible.

The party publishing the percentage must show what was measured.

Do not show us the build time. Show us the total human hours to sign-off.

Human judgement, AI assistance

AI may yet transform financial modelling more profoundly than today’s marketing suggests.

Not because it makes expertise unnecessary, but because it changes where expertise can be spent.

Less time may go into repetitive construction. More can go into architecture, alternatives, controls, explanation and review. A strong modeller may be able to create, test and document materially more within the same working week.

My own case shows rapid construction and iteration. It does not yet show an end-to-end saving against my real commercial baseline. Provider-independent tests and research benchmarks likewise suggest that professional quality and client readiness still require significant human involvement.

The honest conclusion is neither “AI has replaced financial modellers” nor “AI is useless for real models”.

It is more interesting than either.

The first draft is becoming dramatically cheaper and faster.

Judgement, assurance and accountability are not.

That is a meaningful productivity story. It may eventually become an extraordinary one.

But it is not a ten-minute story.

If your system can produce a professional financial model in ten minutes, publish the evidence. Show us the preparation, the review, the corrections, the acceptance criteria and the total human hours to sign-off.

Until then, ten minutes is a build time.

It is not a productivity case.

Human judgement. AI assistance.

Method, limitations and sources

This article grew from a multiple-rounds research process using ChatGPT, Claude, Gemini and Perplexity as research assistants. Their outputs were not treated as evidence. Every factual claim retained in the article was checked against the underlying vendor page, comparative test, research paper or dated case record.

The evidence was treated in four separate categories:

vendor claims, which establish what was publicly promised but not whether it was independently demonstrated;
provider-independent comparative tests, which provide stronger evidence but remain limited to their selected tasks, prompts, tools and versions;
academic or open benchmarks, which test professional complexity but do not necessarily reproduce each commercial product workflow;
my own documented case, which provides detailed practical evidence but is not a controlled experiment.

AI products and benchmark results change rapidly. All conclusions are therefore dated to the research cut-off of 3 August 2026. The article does not conclude that AI produces no savings. It concludes that the publicly available evidence does not yet support a general, precise end-to-end time or cost saving for professional financial models at the same acceptance threshold.

“Provider-independent” does not mean “interest-free”. Wall Street Prep provides finance training and AI advisory services; Ankura provides consulting services. Their tests are useful comparative evidence, not neutral certification. Research benchmarks have their own task, rubric and model-version limitations.

Disclosure: I earn my living through financial-modelling, review and training services. I therefore have an obvious interest in the continued value of professional modelling expertise. I also use AI extensively and have an equally obvious interest in making it work. My own case is included because it produced an inconvenient result for both extremes: rapid construction, substantial expert human work and, so far, no demonstrated end-to-end saving against my normal commercial baseline.

Principal vendor-claim sources

Comparative and contextual evidence

Research benchmarks

Author case

AI-Assisted Project-Finance Model — Live Experience Record, Aug 2026. The author’s observed commercial baseline is 60–80 active human hours for a substantial, complete FAST-standard modelling assignment. The AI-assisted case had required an estimated 35–55 active human hours while the complete Financial Model was assessed as only 55–65% finished. The figures are reconstructed estimates, not a parallel controlled test; no final end-to-end saving percentage is claimed. The detailed reconstruction is available from the author.