How Far Can AI Go on a Customer Value Proposition (CVP)? Part 3 with Anthropic, from Value Proposition Canvas to the Selling Point Customers Pay For

Last time the AI picked three strategies for Anthropic, and all three rested on one assumption: "customers pay for quality". This part tests that assumption from the customer's side, with the tool consultants use to answer "why do customers actually buy?", the Customer Value Proposition (CVP). It is also the part where the AI meets its biggest limitation: it has never talked to a customer.

1. What a Customer Value Proposition is, and why it is the hardest exam for AI

We chained two tools. The first is the Value Proposition Canvas (VPC) by Osterwalder and colleagues (2014), a zoom into the "value propositions" block of the Business Model Canvas from part 1. It has two sides. The customer side is the jobs customers are trying to get done, the pains along the way, and the gains they want. The offer side is products, pain relievers and gain creators. When the two sides match, that is "fit"; a pain with nothing to relieve it is a gap.

The second is the Customer Value Proposition framework by Anderson, Narus and van Rossum in Harvard Business Review (2006), which says there are three kinds of value proposition and only one of them works.

The last line of that table is why this part is the hard exam. Parts 1 and 2 could be done from public documents because they are about the company. The customer side of a VPC has to come from customers, and public documents contain almost none of that. What they contain is the customer-story page the vendor chose to publish. So our question was not only "can the AI do it?" but "does the AI know it is looking at the customer through the vendor's eyes?"

2. The method: build the Value Proposition Canvas first, then examine it with the HBR framework

Two steps, as before. Step one built a VPC for the three customer segments from the part-1 BMC: A, large enterprises in regulated industries; B, software engineering teams; C, everyday office staff. Step two examined that VPC with the HBR framework against the next-best alternative each segment would consider. The 25 attached documents were the 21 from part 2 plus four this job needs: the section of the Menlo Ventures report on why enterprises choose a vendor, Anthropic's customer-story page, the Anthropic Economic Index reports on what people actually do with Claude, and a summary of the part-2 SWOT/TOWS.

You are a business strategy consultant. Build Anthropic's (the developer of Claude)
Customer Value Proposition with a Value Proposition Canvas from the attached
documents, for three customer segments:
A. Large enterprises in regulated industries (finance, health, government)
B. Software engineering teams and developers
C. Everyday office staff in organisations (documents, analysis, workflows)

Rules
1. Plain working language, no preamble.
2. Two sides per segment. Customer side: jobs to be done · pains · gains.
   Offer side: products and services · pain relievers · gain creators.
   3–4 items each, coded A-J1, A-P1, A-G1, A-PR1, A-GC1 ...
3. Every offer-side item must cite [n]. Every customer-side item must state its
   source as one of: (a) evidence from a real customer in document [n],
   (b) inferred from what the vendor or press say [n], (c) prior knowledge,
   may be out of date.
4. State the "fit" per segment: which pains or gains have nothing addressing
   them (gaps), and which offers match no pain or gain at all.
5. Do not guess numbers.
6. End with a table "What must be confirmed with real customers", listing which
   customer-side items come from the vendor's view and what question to ask.
7. No web search.

The second prompt attached the full VPC plus the same documents and asked for three layers per segment: points of parity (what rivals also have), points of difference (with evidence) and resonating focus (the one or two items the customer would weigh most), every item cited and tagged fact or interpretation. Three rules were added. Every resonating focus must state the evidence present, the evidence missing and "what must hold first". It must connect back to the part-2 TOWS and say whether the "pay for quality" assumption behind WT1 and ST1 is supported or challenged. And it must write one value proposition statement per segment in the form "For [customer] who [job or problem], Claude is the [category] that [key difference] because [evidence]", marking which parts of the sentence are interpretation. How to write rules like these is in our guide to writing prompts for Claude.

The two steps took about 15 minutes in total and returned roughly 14,000 and 17,000 characters in Thai, the longest of the three parts, because half of it is tables of questions for customers.

3. Step 1 result: of 36 customer-side items, the AI admitted only 12 came from real customers

Here is the VPC in the standard Strategyzer layout, one canvas per segment. The square on the left is the offer side, the circle on the right is the customer side, and the arrow in the middle asks whether the two sides match. The green "real customer", yellow "vendor's view" and red "prior knowledge" tags were applied by the AI itself under rule 3, and the yellow badge on each image counts them for that segment. Click any image for full size.

Three things worth noticing.

• Two thirds of the customer side is the vendor's view. Of 36 customer-side items, the AI tagged 12 as coming from real customers, 22 as inferred from what the vendor, partners or analysts say, and 2 as prior knowledge. Notably segment A, the one with the largest contracts, has the least real-customer evidence, 2 of 12, because the customer stories the vendor publishes talk about outcomes, not about the pain of procurement.
• The AI found the gaps the vendor does not talk about. For segment A the offers in the documents cover only US frameworks (FedRAMP, HIPAA) and are silent on the EU AI Act even though it is already in force. For segment B there is no offer competing at the "cents" price level. For segment C the pain "no confidence in accuracy, must re-check" has no offer addressing it at all, and the time-saved evidence comes from specialist professions such as law and pharma, not general office staff.
• The AI also pointed out what customers may not want. Being on all three clouds, a major strength in part 2, may not matter at all to a dev team buying Claude Code by the seat, the AI said. The $1 deal applies only to US federal agencies. And the Public Benefit Corporation governance the vendor is proud of has no evidence in the documents of being used as a purchasing criterion.

The full "what must be confirmed with real customers" table has 24 questions and is in the raw output we kept on file · SCIM = automated user provisioning standard · LTBT = Long-Term Benefit Trust · RSP = Responsible Scaling Policy

4. Step 2 result: only one selling point passes "different, measurable, evidenced" in full

Step two was the exam. The AI worked through all three segments with the HBR framework and concluded on its own that only one Anthropic value proposition meets all three resonating-focus tests, different from the next-best alternative, measurable and evidenced: code quality for segment B. And even that one has a problem with the age of its evidence.

The value proposition statement the AI wrote for segment B, the one that passed, reads: "For engineering teams that must ship code faster without losing quality and take agents from pilot to production, Claude is the model and agent-tooling platform for coding that delivers measurably higher code quality than the next-best alternative, because it has sat at the top of the coding leaderboard for almost 18 months, holds 54% of enterprise coding versus 21%, and customers such as Deepgram get code that is 4 to 10 times more durable." Then the AI added at once that "measurably" rests on vendor figures and "higher than the next-best alternative" rests on November 2025 data.

The "what a human must decide" table in this step has seven issues. We picked the four that part 2 did not already cover.

And the most valuable output of this step is not the statements but the list of questions to ask real customers. Step 1 produced 24, step 2 added 15 that do not overlap, 39 in total. The three sharpest: "In the same console, what do you compare Claude with the cloud's own model on: price, quality or contract terms?", "After rivals cut prices by 80%, did the team re-test, and what work did you move?", and "If ChatGPT Work or Gemini cost the same or less, what would make you choose Claude?"

5. What we learned: AI forms hypotheses about customers very well, but it is not the customer

Across the three parts, the AI got through four stages: sort facts, separate fact from opinion, pair them into options, and now form hypotheses about customers with a way to test them. What it cannot do, and knows it cannot do, is confirm those hypotheses, because the answers sit with customers, not in any document.

• Where AI did better than expected: tagging every item as customer or vendor sourced, finding the gaps the vendor does not mention, refusing to compare with rivals it had no data on, and comparing the date of the evidence with the date of the threat.
• What stays human: taking the 39 questions to at least five real customers, adding internal data such as lost-deal reasons and churn, deciding which segment to lead with, and owning the statement that goes into real selling.
• What to watch for: the prettiest VPC is the one the vendor writes for itself. Count the "vendor's view" tags before reading the content, every time.

6. Running it on your own company: the customer side comes from real data and from asking, not from your own brochure

Most companies write a VPC by having sales and product people guess the customer's pains in a meeting room. The result is the same kind of VPC the AI wrote for Anthropic in step 1: two thirds vendor's view. The fix has two parts: the data that already records customer behaviour in your systems, and asking.

1. Pull the signals from your systems first. Customer pain always leaves traces in the ERP (Enterprise Resource Planning) system: receivables that are habitually late, returns or credit requests, deliveries past their date, complaints. What customers want shows in what they buy again and in accounts that raise their orders. Export to Excel or PDF and date every file. That is your first set of "real customer" evidence.
2. Have the AI draft the questions, then ask five customers. Run the step-1 prompt with the data from step 1 plus your own product documents. The output will carry many "vendor's view" tags. That is your interview list for customers who bought, customers who stopped, and customers who chose a rival, at least one or two of each.
3. Build the VPC a second time from the interview notes. Attach the notes and the tags will flip from "vendor's view" to "real customer" one by one. Any item that will not flip is not yet ready to be used in selling. How to attach files and share one document set across the team is covered in using Claude to read Excel, CSV and PDF files and Claude Projects.
4. In step 2, tell the AI who the customer's real next-best alternative is. Not the rival you would like to be compared with, but what customers said in step 2 they compare you with. Often the next-best alternative is "keep doing it the old way" or "do nothing".
5. Tie the resonating focus to a number you will watch every quarter. If the selling point is "we deliver on time like nobody else", the number is on-time delivery in your system. If it is "we save the customer time", collect the figure from two or three real customers rather than estimating. On the executive's role in deciding on real data, see ERP succeeds or fails at the top.

Conclusion

• AI built a six-block Value Proposition Canvas for three customer segments and self-tagged the 36 customer-side items: 12 from real customers, the rest from the vendor's view or prior knowledge.
• AI found the gaps the vendor does not mention on its own: silence on the EU AI Act, no "cents" price tier, and no offer addressing accuracy of output.
• Tested with the HBR framework, only one selling point passed "different, measurable, evidenced" in full, code quality, and its evidence is older than the price threat.
• The part-2 assumption "customers pay for quality" is supported for developers, challenged for volume work, and unevidenced for office staff.
• The most valuable thing the AI produced is 39 questions for real customers, not the value proposition statements.
• For your own company, the customer side of a VPC must come from traces in your systems and from interviews; let AI form the hypotheses and draft the questions, not answer for the customer.

AI can write a fluent value proposition in ten minutes, but the most fluent sentences usually come from the vendor's brochure, not from the customer's mouth. What AI can really give you is the list of questions. The answers still have to come from your customers.

- What we learned from the part-3 test, 26 September 2026

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)
• [16] Yahoo Finance / Quartz (citing Bloomberg): OpenAI annualized revenue tops $40 billion ahead of IPO (14 Aug 2026)
• [17] Menlo Ventures: 2025 — The State of Generative AI in the Enterprise (9 Dec 2025)
• [18] State Street Investment Management: Beyond DeepSeek — China's 2026 model wave and the repricing of the AI stack (21 Sep 2026)
• [19] EU Artificial Intelligence Act: Implementation timeline (updated 31 Aug 2026)
• [20] PYMNTS: Google Cloud rides enterprise AI demand to 82% growth (Alphabet Q2 2026 results) (22 Jul 2026)
• [21] Our part 1: How far can AI go on a Business Model Canvas? — the verified BMC (26 Sep 2026)
• [22] Menlo Ventures (same report as [17]), the section on why enterprises choose a vendor and price sensitivity (9 Dec 2025)
• [23] Claude: Customer stories (Spellbook, EvenUp, Deepgram, Pictet, League, Carvana and others, checked 26 Sep 2026)
• [24] Anthropic Economic Index: Economic primitives (15 Jan 2026) and Cadences (26 Jun 2026)
• [25] Our part 2: How far can AI go on SWOT and a TOWS Matrix? — the WT1/ST1 assumptions (26 Sep 2026)
• Strategyzer: The Value Proposition Canvas (Osterwalder, Pigneur, Bernarda, Smith, 2014)
• Anderson, Narus & van Rossum (2006): Customer Value Propositions in Business Markets, Harvard Business Review