How Far Can AI Go on a Business Model Canvas? A Real Test on Anthropic, the Company Behind Claude

This article is not a theory lesson. It answers one question executives ask more and more often: "If I let AI write my Business Model Canvas, how far can I trust it?" The most direct way to answer is to run a real test on a company with plenty of public data that changes fast. So we picked Anthropic, the company behind Claude, had Claude analyse its own parent company, and then checked the result figure by figure.

1. What a Business Model Canvas is, and why it suits AI

A Business Model Canvas (BMC) is a single page that explains how a business makes money, in nine blocks. It was created by Alexander Osterwalder and Yves Pigneur in the 2010 book Business Model Generation. Its strength is that it forces you to write short, see the whole picture on one page, and keep all nine blocks consistent. If the revenue block says you sell to enterprises but the channels block only lists a consumer app, something is wrong.

The BMC suits AI because it is a "sort the facts into a frame" task, which AI does very well. That is also its weakness. If the input is incomplete or old, AI will sort old facts into the frame neatly and beautifully, and you will not know until you check.

2. The method: one prompt, two passes, the only difference is "data attached or not"

We used the same prompt in both passes. The first pass attached nothing and let the AI answer from what it already knew. The second pass attached 15 documents we collected ourselves: 8 Anthropic announcements plus 7 news reports and pricing pages, each with a date. Web search was off in both passes, so that any difference in the result comes from the attached data, not from the model's ability.

You are a business strategy consultant. Analyse Anthropic (the developer of Claude)
with a Business Model Canvas, all nine blocks.

Rules
1. Answer in plain working language, no preamble.
2. Write 3–6 short bullets per block.
3. For every number (revenue, valuation, customer counts, prices) state
   "as of when" and "from where". If unsure, say you are unsure. Do not guess.
4. End with a table titled "What a human must verify", listing which blocks
   may be out of date or incomplete.
5. Answer from your own knowledge only, no web search, and say when your
   knowledge ends.

These five rules are the heart of the test. Rule 3 forces a date and a source on every number. Rule 4 makes the AI write its own "what to verify" list, which is the rule that saves the human checker the most time. The second pass changed only rule 3 to "use the attached documents as the primary source, put a reference number [n] on every figure, and if you must fall back on prior knowledge, mark it as possibly out of date". Everything else stayed identical. We ran the prompt in Thai; the version above is the English translation. How to write instructions like this is covered in our guide to writing prompts for Claude.

Each pass took about 5 to 7 minutes, because this is the top-tier model and it thinks for a long time. Each answer came back at roughly 7,000 to 8,000 characters in Thai, with the verification table at the end.

3. Pass 1, no data attached: all nine blocks right, but the numbers almost a year behind

The first line the AI wrote, before any analysis, was a statement of its own limits: "My knowledge runs to about mid-2026, but the part I am really confident about ends in late 2025. Today is 26 September 2026, so there is a gap of at least three months. And I am an Anthropic model myself, so my view may be biased." That sentence is the most valuable one in the whole answer, because it tells you how to read everything that follows.

All nine blocks came back, and the structure was correct. Main customers are enterprises and developers calling the API (Application Programming Interface). The value proposition is coding and agent capability plus a safety brand. There are three compute partners. The largest cost is compute. All of that is still true today. But when you go through the numbers, every one stops at the end of 2025.

GW = gigawatt, the power capacity of a data centre · TPU = Tensor Processing Unit, Google's AI chip · IPO = Initial Public Offering · numbers in [n] point to the references at the end

To its credit, the AI did not make things up. Every number carried a date, and the ones it was unsure of were marked unsure. The "what to verify" table at the end had 10 rows and pointed correctly at every figure that really was out of date. But a reader who skips the first line gets a BMC that looks complete while its headline numbers are off by 5 to 9 times.

4. Pass 2, 15 documents attached: a BMC where every line can be traced

The second pass changed the whole result. Every line with a number carried an [n] pointing back to an attached document. Where the AI gave its own opinion, it wrote "(interpretation)" next to it, without being told to. And it did three things we did not expect.

• It flagged that the $65B and $100B run-rate figures were "media citing sources, not official announcements", and recommended waiting for the S-1 (the registration document a company files before listing on a US exchange) before treating them as confirmed.
• It warned that every compute figure is an "up to" figure from an agreement, not installed capacity in use.
• It separated the "about 80% of revenue from API" share as a third-party estimate, not an Anthropic figure.

The "what a human must verify" table grew from 10 rows to 15. That sounds worse but is actually better, because each row is more specific. "The figures may be old" became "check the revenue-share terms with the cloud providers" or "separate the investment announced as 'up to' from what has actually been paid in". Those are the questions analysts ask.

Below is Anthropic's BMC from pass 2, after we checked it against the sources. We removed the items the AI had marked as "prior knowledge" and kept only what could be confirmed against a document. The image shows the canvas in the standard nine-block layout; the table that follows has the same content with reference numbers.

SCIM = System for Cross-domain Identity Management (automated user provisioning) · FedRAMP = the US federal cloud security authorisation programme · HIPAA = the US health data privacy law · run-rate = the latest month's revenue multiplied by 12, not actual full-year revenue

Read as strategy, the table shows one picture. Anthropic sells to enterprises through every channel enterprises already buy from, the three big clouds, the tools already in use, and the government channel, and then commits that revenue to multi-year compute contracts in advance. The revenue block and the cost block grow together. That is why the analysis in [11] calls gross margin "the most important number" for the company, and it is the number the official S-1 will finally reveal once it is public.

5. Four things AI still cannot do for you, even with all the data attached

The last row matters most for Thai readers. Nothing in Anthropic's nine blocks tells a Thai organisation how to buy Claude with a tax invoice, handle VAT under form Por.Por.36, or run the procurement paperwork, because that is a channel question in each country, not part of the parent company's business model. If you need that, we cover it in buying direct from Anthropic versus through a supplier in Thailand and in our Team and Enterprise plan FAQ.

6. Running it on your own company: five steps, and where ERP comes in

1. Collect real data before you open the chat. At least five of the nine blocks already live in your systems. Customer segments and channels are in revenue reports by customer or department. Revenue streams are in the general ledger. Cost structure is in cost-centre reports. Key partners are the largest suppliers in your procurement system. Key resources are your headcount and budget. Export them to Excel or PDF and date every file.
2. Use the prompt from section 2. Change the company name and attach your files in place of our 15 documents. How to attach files and the limits of each file type are covered in using Claude to read Excel, CSV and PDF files.
3. Give the whole team the same document set through Claude Projects, so that sales, finance and management get a BMC from one data set instead of each attaching their own.
4. Executives read the "what to verify" table before the BMC. That table is the homework list the AI prepared. An executive who signs off on the BMC without reading it is accepting numbers nobody has checked. We wrote about the executive's role with system data in ERP succeeds or fails at the top.
5. Repeat every quarter. Keep the old version to compare. A block that changes is a signal the business is moving.

Next in the series: we take this BMC forward into a SWOT and a TOWS Matrix with the same method, in How far can AI go on SWOT and a TOWS Matrix?

Conclusion

• AI can write a full nine-block Business Model Canvas in minutes, and the structure was right even in the pass with no data attached.
• Without attached data, the headline numbers stopped at late 2025 and were off by 5 to 9 times, but the AI stated its knowledge limits and its uncertainties clearly.
• With 15 documents attached, every figure pointed back to a source, facts were separated from interpretation, and the AI produced its own list of 15 items for a person to verify.
• The rules that made the difference: require a date and a source on every number, forbid guessing, and make the AI write the verification table.
• What remains human work: finding the right sources, cutting to what matters, making the strategic call, and adding local context such as how to buy in Thailand.
• For your own company, five of the nine blocks already sit in your ERP. Start by exporting dated reports, then use the same prompt.

AI can fill the canvas in ten minutes. The numbers in it belong to whoever attached the data. If nobody did, they belong to the past.

- What we learned from the 26 September 2026 test

References

• [1] Anthropic: Anthropic raises $30B Series G at $380B post-money valuation (12 Feb 2026)
• [2] Anthropic: Anthropic raises $65B Series H at $965B valuation (28 May 2026)
• [3] TechCrunch (citing Bloomberg): Anthropic's annualized revenue surges to $65B (17 Aug 2026)
• [4] Axios (citing The New York Times): Anthropic tops $100 billion revenue pace, report says (18 Sep 2026)
• [5] Anthropic: Expanding our use of Google Cloud TPUs and services (23 Oct 2025)
• [6] Anthropic: Anthropic expands Google and Broadcom compute deal (6 Apr 2026)
• [7] Amazon: Amazon announces $5B Anthropic investment, up to $20B more (20 Apr 2026)
• [8] Microsoft: Microsoft, NVIDIA and Anthropic announce strategic partnerships (18 Nov 2025)
• [9] Anthropic: Seoul becomes third Anthropic office in Asia-Pacific (and Sydney fourth office, 10 Mar 2026) (23 Oct 2025)
• [10] Claude: Pricing page (checked 26 Sep 2026)
• [11] Yahoo Finance: Anthropic's gross margin is the most important number in tech (10 Jun 2026)
• [12] GSA: GSA strikes another OneGov deal with Anthropic (extended to 31 Oct 2026 per Nextgov, Sep 2026) (12 Aug 2025)
• [13] Anthropic: The Long-Term Benefit Trust (Sep 2023)
• [14] Anthropic: Claude Cowork and product pages at claude.com (checked 26 Sep 2026)
• [15] Sacra: Anthropic revenue, valuation & funding (third-party estimate) (2026)
• Strategyzer: The Business Model Canvas (Osterwalder & Pigneur)