---
title: "Can AI Build a Three-Statement Model? 4 Audits"
slug: can-ai-build-a-three-statement-model
date: 2026-11-18
excerpt: "We audited four AI-built three-statement models: 16,573 formulas, 983 of 983 figures matched to SEC filings, and where the build fell short."
category: Product Updates
tags: ["Research Workflows", "AI & Agents"]
author: Jennifer Ma
authorRole: "Co-founder & CEO"
cta: product
ctaPrompt: "Show Alphabet's revenue, operating income and capital expenditure for the last eight quarters, and cite the SEC filing behind each figure."
status: published
---

Can AI build a three-statement model? Yes, for the parts of the job you can check. An AI agent can read a company's filings and return a linked income statement, balance sheet and cash flow statement in minutes. The forecast it writes is still an opinion, and none of it should be trusted on sight. So the useful question is which parts you can verify, and how quickly. To answer with numbers, we audited the four public sample models built by FocusAlpha Model Builder, our own tool: NVIDIA, Alphabet, Anthropic and Oura. Together they hold 29,349 cells and 16,573 formulas. **For the two companies that file XBRL, all 983 statement figures the builder compared matched the facts the companies filed with the SEC.** The same files also record where the build fell short, and we report that too.

**TLDR:**

- The four sample models contain 16,573 formulas with no formula errors. The balance sheet balances in all 89 periods that have one, and ending cash ties to balance-sheet cash in the same 89.
- 4,894 of 5,001 reported figures (97.9%) link to the place they are printed. The other 107 are share prices, market caps and metrics quoted from text, each with its date or its sentence.
- NVIDIA changed its revenue lines in fiscal 2027. The model keeps 49 restated figures on their own tab, each with its source, and leaves the years the company never recast blank.
- The weak part is the forecast. The first pass of the Anthropic model compounded revenue to $1.18 trillion by 2028 before the sense check graded it "not sensible."
- Model Builder is part of the FocusAlpha Fund plan. The four samples open without an account.

## What a three-statement model has to get right

Building a three-statement model is four jobs, and they fail in different ways.

1. **History as reported.** Every historical figure has to be the number the company printed, on the basis it printed it.
2. **Integration.** Net income flows to cash flow and equity, cash flow ends in the balance sheet's cash, and assets equal liabilities plus equity in every period.
3. **Drivers.** Revenue is split into the things that produce it: units and price, customers and revenue per customer, backlog and conversion.
4. **Forecast.** Each driver gets an assumption with a reason behind it.

Spreadsheets were error-prone long before AI. Raymond Panko's review of field audits found errors in 91% of the 54 spreadsheets audited from 1997 onward, and cell error rates of 1.1% to 2.5% in the large spreadsheets audited most thoroughly ([Panko, 2000](https://arxiv.org/abs/0802.3457)). A model with thousands of formulas is likely to contain errors unless something checks it.

General-purpose AI tools have not fixed that. In a test last updated in June 2026, Wall Street Prep asked four of them to build an Apple three-statement model from the 10-K and scored the best one 5.9 out of 10, below its benchmark of 6.4 for a low-bucket analyst ([Wall Street Prep, 2026](https://www.wallstreetprep.com/knowledge/ranking-the-best-ai-tools-for-financial-modeling-2026/)). Its two sharpest findings were about jobs 1 and 2: "Claude and Shortcut both hallucinated significant amounts of historical data," and on the balance sheet, "Most relied on plugs rather than proper integration."

That gives a fair test for any AI model builder, ours included: is the history sourced, do the statements tie, and can you see every assumption?

## What FocusAlpha Model Builder does

You type a ticker. A team of agents reads the company's filings, earnings releases, presentation decks and call transcripts, and returns an Excel workbook. Each model has:

- **Three linked statements**: income statement, balance sheet and cash flow, by quarter and by year, with check rows under each.
- **Source links on history**: every reported figure opens the filing at the place it is printed.
- **Segments and drivers**: each revenue line broken into the drivers the company discloses.
- **Management's words**: guidance and comments from calls and conferences, quoted next to the forecast they support.
- **Restatements**: recast periods kept apart from the figures as first reported.
- **An Assumptions tab**: every forecast input in one list, with its basis.

A chat sits beside the model. You ask it to change an assumption, and it changes the cells and says what moved. You can download the workbook and keep your own rows and formulas. When the company reports, the Excel add-in brings the new quarter into your workbook and keeps your changes.

These are the four samples we audited. Each link opens the model in a browser, no account needed.

| Sample model | Built from | Reported history | Tabs |
| --- | --- | --- | --- |
| [NVIDIA](https://app.focusalpha.ai/share/_0XRIzI1b8jFEWcHwKZ2a6vW) | 10-Ks, 10-Qs, 8-Ks | 5 years, 10 quarters | 16 |
| [Alphabet](https://app.focusalpha.ai/share/7r2ZnpMxQUcY--plAxdbFc97) | 10-Ks, 10-Qs, 8-Ks | 5 years, 10 quarters | 14 |
| [Anthropic](https://app.focusalpha.ai/share/ARKj2wsrUzgw_wBtj4ixln3O) | S-1 | 3 years, 6 quarters | 11 |
| [Oura](https://app.focusalpha.ai/share/_PdEscI-NwmavdRefFBB4xwG) | S-1/A | 2 years, 7 quarters | 10 |

*Source: the four public sample models, FocusAlpha Model Builder, versions of October 28 to November 2, 2026.*

One caution before the numbers. These are our samples, chosen by us as the best build of each company. They show what the tool can do, not its average. The checks below are the ones a reader can repeat on the public files.

## How much of an AI-built model is formulas?

More than half. Across the four models, 56% of all cells are formulas, 17% are reported figures, and the rest are labels, notes and assumption inputs.

| Sample model | Cells | Formulas | Assumptions |
| --- | --- | --- | --- |
| NVIDIA | 12,060 | 7,219 | 289 |
| Alphabet | 8,886 | 5,151 | 302 |
| Anthropic | 3,943 | 1,844 | 164 |
| Oura | 4,460 | 2,359 | 224 |
| **Total** | **29,349** | **16,573** | **979** |

*Source: FocusAlpha analysis of the four public sample models.*

![Grouped bar chart of four sample models. NVIDIA has 7,219 formulas and 1,776 source-linked reported figures; Alphabet 5,151 and 1,630; Oura 2,359 and 750; Anthropic 1,844 and 738](/images/blog/can-ai-build-a-three-statement-model/fig-1.svg)

*Figure 1: Formulas and source-linked reported figures in each sample model. Source: FocusAlpha analysis of the four public sample models.*

None of the 16,573 formulas returns an error. That is a low bar, but Panko's cell error rates explain why it matters at this size: at 1%, a model like NVIDIA's would carry about 70 wrong formulas. The 979 assumption cells are the forecast inputs. Each one is listed on the Assumptions tab with its default and the reason for any change.

## Can you trace every historical number?

Almost every one. Of the 5,001 reported figures in the four models, 4,894 carry a link to the place they are printed, which is 97.9%.

The 107 without a link fall into two groups. Seventy-two are share prices and market caps on the Peers tab, each stamped with its date. Thirty-five are operating metrics that management stated in words, on a call or in the filing, such as a customer count. Each of those carries the quoted sentence.

Most links go to sec.gov. The NVIDIA model links to 75 separate SEC filings and the Alphabet model to 78. The Anthropic and Oura models each link to one, because a company before its IPO has only its registration statement. That is why their histories are shorter: the S-1 prints what it prints.

This is the check that answers Wall Street Prep's first finding. A hallucinated figure has no page to open. A linked one takes a click to verify.

## Do the figures match the filings?

For NVIDIA and Alphabet, the builder compared 983 statement figures with the XBRL facts each company filed with the SEC: 540 for NVIDIA and 443 for Alphabet. All 983 matched. Where our data layer and the filing disagree, the model is built to show the filing's figure.

There is no such comparison for Anthropic and Oura, and the model says so. A first registration statement has no history of XBRL facts to compare against, so those two model pages state that the reported figures come from the registration statement and show no count.

## Does the balance sheet balance?

Yes, in every period that has one. The check row under the balance sheet reads OK in 89 of 89 periods across the four models, reported and forecast. Ending cash on the cash flow statement ties to balance-sheet cash in the same 89.

| Sample model | Balance check OK | No balance sheet printed | All checks OK |
| --- | --- | --- | --- |
| NVIDIA | 28 of 28 | 0 | 503 |
| Alphabet | 28 of 28 | 0 | 371 |
| Anthropic | 17 of 17 | 5 | 144 |
| Oura | 16 of 16 | 5 | 164 |
| **Total** | **89 of 89** | **10** | **1,182** |

*Source: FocusAlpha analysis of the check rows in the four public sample models. A check with nothing to compare stays blank; 503 check cells are blank and none reads as failed.*

The ten blank periods are informative. The Anthropic and Oura registration statements do not print a balance sheet for every quarter. The model leaves those quarters empty and flags them as missing. It does not fill them with an estimate, and the check stays blank instead of reading OK.

The checks go beyond the balance. Each model also tests that segment revenue sums to consolidated revenue, that net income on the cash flow statement equals net income on the income statement, that each year equals its four quarters, and that the forecast sits inside the company's guided range where one exists.

## What happens when a company restates?

NVIDIA is the test case. In the first quarter of fiscal 2027 it changed how it presents revenue by market platform and recast the comparable periods. In the second quarter it moved one company between platforms and recast again ([NVIDIA, Q2 FY27 revenue by market platform](https://s201.q4cdn.com/141608511/files/doc_financials/2027/Q227/Q227-Revenue-by-Market-Platform-Slides.pdf)).

A model that pastes the new lines over the old ones mixes two bases, and a growth rate that spans the change is then wrong without any error showing. This is how the sample model handles it:

- The new lines (Hyperscale; AI Clouds, Industrial, & Enterprise; Edge Computing) run from Q1 FY25, the first quarter NVIDIA recast.
- The old lines stay as first reported, in their own block.
- Fiscal years the company never recast are blank in the new structure. They are not back-filled.
- A Restatements tab lists all 49 restated figures, each with its source. The first one, $11,633 million of AI Clouds, Industrial, & Enterprise revenue for Q1 FY25, matches NVIDIA's own slide.

The model also records a change to NVIDIA's geographic lines in the 10-Q filed in November 2025. That accounts for 21 of the 49.

## Where did the AI model builder fall short?

The models keep their own build logs and flags. Four entries are worth reading before you trust any AI-built forecast.

**The first forecast can be absurd.** The first pass of the Anthropic model took growth rates from the first half of 2026 and carried them forward. Revenue reached $1.18 trillion in 2028. The sense check that runs after a build graded it "not sensible," with five serious findings, and wrote that growth "is far beyond anything the company has printed, on a tiny base." The growth assumptions were then cut, and the model records a reason for each change. The sample as published forecasts $149bn for 2028. Our [Anthropic S-1 analysis](/blog/anthropic-ipo-s1-financial-model-2028-revenue) used an earlier version of this model that reached $195bn on different customer assumptions. Two versions of one model differ by $46bn, and the gap comes entirely from assumptions.

**A driver the company does not disclose cannot be found.** The NVIDIA build tried to split data-center revenue into gigawatts deployed and revenue per gigawatt. Its log says no document gives either factor for Hyperscale in any reported quarter. One of the three agents assigned to a revenue line finished without a split the model could use. Those lines are forecast on growth, with management's words quoted beside them.

**An earnings release is not always readable.** Alphabet's Q3 release was filed on October 28, 2026. The builder could not read its balance sheet, so the quarter stays a forecast in the sample until the 10-Q, and a flag on the model says so.

**The review leaves items open.** The NVIDIA review ran two rounds, made one fix and left one item open: a memo line on the balance sheet that is blank in the forecast columns.

![Dot timeline of the NVIDIA sample build on October 25, 2026. History is logged at 14:39 UTC, research, driver and forecast steps run to 14:49, balance and checks are logged at 14:50, and the review finishes at 14:56](/images/blog/can-ai-build-a-three-statement-model/fig-2.svg)

*Figure 2: Logged steps of one NVIDIA build, by stage. Source: build log of the public NVIDIA sample model.*

The log of that build runs 10 minutes 55 seconds from its first entry to the checks, and the review ends six minutes later. Speed is the least interesting result. A fast model that needs a day of checking has saved nothing, so what matters is whether the errors are in places you can see.

## How to check an AI-built financial model in ten minutes

These five checks work on any model, whoever or whatever built it.

1. **Open three historical figures at random.** Each should take you to the page that prints it. If a figure has no source, treat it as unverified.
2. **Read the balance check across every column**, forecast included. Then look for a plug: a line whose only job is to absorb the difference.
3. **Tie the cash.** Ending cash on the cash flow statement should equal cash on the balance sheet in every period.
4. **Find a restatement.** Pick a company that changed its segments and see whether old and new lines are mixed in one row.
5. **Read the assumptions before the valuation.** A valuation is arithmetic on the assumptions. If the revenue path makes no sense, neither does the price target.

## Who can use Model Builder?

Model Builder is part of the FocusAlpha Fund plan, which is sold to investment teams. [Contact us](/contact) to see it on a company you cover, or compare plans on the [pricing page](/pricing).

The four sample models are open to anyone: the links above open in a browser, and downloading a workbook takes a free account. If you build models with your own agents, reported financial statements are available through the [financials API](/docs/api/financials) and MCP.

## Methodology

We downloaded the four public sample models on November 18, 2026, in the versions frozen between October 28 and November 2, 2026. Each is a grid of cells in the same shape as the Excel export. We counted cells, formulas, and cells the builder marks as reported figures, and how many of those carry a link. Check results are the values of the check cells; a check with nothing to compare is blank. "Figures compared with filings" is the builder's own count of statement figures compared with SEC XBRL company facts during the build. Restated figures are the model's list of them, and we checked the NVIDIA example against the company's slide. Build times are differences between timestamps in the build log. The samples were selected by FocusAlpha, so these results describe four models and are not an average across companies.

## FAQ

### Can AI build a three-statement financial model?

Yes. An AI agent can assemble linked income statement, balance sheet and cash flow statements from a company's filings in minutes. In our audit of four models built by FocusAlpha Model Builder, the balance sheet balanced in all 89 periods and none of 16,573 formulas returned an error. The forecast still depends on assumptions that a person should read and change.

### How accurate are AI-built financial models?

It depends on whether the history is sourced. In a 2026 test by Wall Street Prep, general-purpose AI tools scored at most 5.9 out of 10, and some invented historical figures. In our four sample models, 4,894 of 5,001 reported figures link to where they are printed, and all 983 figures compared with SEC XBRL facts matched.

### How do you check an AI-generated financial model?

Open a few historical figures and confirm each one against its source. Read the balance check in every column, forecast included. Confirm that ending cash ties to the balance sheet. Look for mixed bases after a restatement. Read the assumptions before you read the valuation.

### What is FocusAlpha Model Builder?

Model Builder is a feature of FocusAlpha that builds a three-statement model of a public company from its filings, earnings releases, decks and call transcripts. Every reported figure links to its source, revenue is forecast from the drivers the company discloses, and the model downloads as an Excel workbook that an add-in updates each quarter.

### Which plan includes Model Builder?

Model Builder is part of the FocusAlpha Fund plan, which is sold to investment teams. The four sample models (NVIDIA, Alphabet, Anthropic and Oura) open in a browser without an account, and downloading a workbook takes a free account.

### Can an AI model handle restatements and segment changes?

It can if it keeps the bases apart. In the NVIDIA sample model, the revenue lines the company introduced in fiscal 2027 sit in their own structure, the old lines stay as first reported, and a Restatements tab lists 49 restated figures with their sources. Years the company never recast are left blank.

### What is FocusAlpha?

FocusAlpha is a [SEC filings API](/docs/api/filings) and agent-ready financial data
layer: it turns SEC filings (10-K, 10-Q, 8-K, 13F), earnings-call transcripts, and
other trusted company communications into structured, normalized data where every
value keeps its citation back to the source document. AI agents connect via
[API or MCP](/docs) to research public companies from complete, trusted information.

<script type="application/ld+json">
{"@context": "https://schema.org", "@type": "FAQPage", "mainEntity": [{"@type": "Question", "name": "Can AI build a three-statement financial model?", "acceptedAnswer": {"@type": "Answer", "text": "Yes. An AI agent can assemble linked income statement, balance sheet and cash flow statements from a company's filings in minutes. In our audit of four models built by FocusAlpha Model Builder, the balance sheet balanced in all 89 periods and none of 16,573 formulas returned an error. The forecast still depends on assumptions that a person should read and change."}}, {"@type": "Question", "name": "How accurate are AI-built financial models?", "acceptedAnswer": {"@type": "Answer", "text": "It depends on whether the history is sourced. In a 2026 test by Wall Street Prep, general-purpose AI tools scored at most 5.9 out of 10, and some invented historical figures. In our four sample models, 4,894 of 5,001 reported figures link to where they are printed, and all 983 figures compared with SEC XBRL facts matched."}}, {"@type": "Question", "name": "How do you check an AI-generated financial model?", "acceptedAnswer": {"@type": "Answer", "text": "Open a few historical figures and confirm each one against its source. Read the balance check in every column, forecast included. Confirm that ending cash ties to the balance sheet. Look for mixed bases after a restatement. Read the assumptions before you read the valuation."}}, {"@type": "Question", "name": "What is FocusAlpha Model Builder?", "acceptedAnswer": {"@type": "Answer", "text": "Model Builder is a feature of FocusAlpha that builds a three-statement model of a public company from its filings, earnings releases, decks and call transcripts. Every reported figure links to its source, revenue is forecast from the drivers the company discloses, and the model downloads as an Excel workbook that an add-in updates each quarter."}}, {"@type": "Question", "name": "Which plan includes Model Builder?", "acceptedAnswer": {"@type": "Answer", "text": "Model Builder is part of the FocusAlpha Fund plan, which is sold to investment teams. The four sample models (NVIDIA, Alphabet, Anthropic and Oura) open in a browser without an account, and downloading a workbook takes a free account."}}, {"@type": "Question", "name": "Can an AI model handle restatements and segment changes?", "acceptedAnswer": {"@type": "Answer", "text": "It can if it keeps the bases apart. In the NVIDIA sample model, the revenue lines the company introduced in fiscal 2027 sit in their own structure, the old lines stay as first reported, and a Restatements tab lists 49 restated figures with their sources. Years the company never recast are left blank."}}, {"@type": "Question", "name": "What is FocusAlpha?", "acceptedAnswer": {"@type": "Answer", "text": "FocusAlpha is a SEC filings API and agent-ready financial data layer: it turns SEC filings (10-K, 10-Q, 8-K, 13F), earnings-call transcripts, and other trusted company communications into structured, normalized data where every value keeps its citation back to the source document. AI agents connect via API or MCP to research public companies from complete, trusted information."}}]}
</script>

## Disclaimer

This article is analysis only and does not constitute investment advice or a recommendation to buy or sell any security. The sample models are FocusAlpha's own, and their forecasts are model estimates that may be wrong.

## Sources

- [FocusAlpha Model Builder — NVIDIA sample model, version d0b930a, 2026-11-02](https://app.focusalpha.ai/share/_0XRIzI1b8jFEWcHwKZ2a6vW)
- [FocusAlpha Model Builder — Alphabet sample model, version c2fd20d, 2026-10-29](https://app.focusalpha.ai/share/7r2ZnpMxQUcY--plAxdbFc97)
- [FocusAlpha Model Builder — Anthropic sample model, version 92cd36a, 2026-10-30](https://app.focusalpha.ai/share/ARKj2wsrUzgw_wBtj4ixln3O)
- [FocusAlpha Model Builder — Oura sample model, version 6e03afc, 2026-10-28](https://app.focusalpha.ai/share/_PdEscI-NwmavdRefFBB4xwG)
- [NVIDIA — Quarterly revenue by market platform, Q2 FY27 slides](https://s201.q4cdn.com/141608511/files/doc_financials/2027/Q227/Q227-Revenue-by-Market-Platform-Slides.pdf)
- [Wall Street Prep — Ranking the Best AI Tools for Financial Modeling (2026), Matan Feldman, updated 2026-06-29](https://www.wallstreetprep.com/knowledge/ranking-the-best-ai-tools-for-financial-modeling-2026/)
- [Panko, R. — Spreadsheet Errors: What We Know. What We Think We Can Do. EuSpRIG 2000, arXiv:0802.3457](https://arxiv.org/abs/0802.3457)
