Key Takeaways
- →A demo proves the model can work. It does not prove that users will change how they work, that the costs hold at volume, or that the data and governance will pass review.
- →The AI Use Case Blueprint scores one use case on twelve blocks in four pillars: Foundation, Value Engine, Execution and Sustainability.
- →Where a red falls decides what happens next: stop for Foundation, pause for Value Engine, proceed with care for Execution, plan before scale for Sustainability.
- →Strength in one pillar does not make up for a red in another.
Formerly published as the AI Use Case Canvas.
The AI Use Case Blueprint tests whether one AI use case deserves funding before the money is spent. It asks twelve questions in four groups: is the problem real and the data there, will anyone pay or change their work for it, can it be built and adopted, and will it still pay and pass review at scale. Each answer is scored red, yellow or green, and where the reds fall tells you whether to stop, pause, proceed with care or plan before you scale.
Updated 4 October 2026. Renamed and rebuilt as a short reference. Client stories and advisory figures without a public source were removed, and every external figure was checked at its source. The changes are listed at the end.
Why Is a Working Demo Not a Business Case?
A demo shows that a model can do the task on data someone chose. It says nothing about whether the people who will use it will change how they work, what it costs at production volume, or whether the data pipeline will pass a privacy or regulatory review. Those questions decide most outcomes, and they are the ones a good demo makes easy to skip.
The pattern is common. In 451 Research's 2025 enterprise survey, 42% of companies surveyed were abandoning the majority of their AI initiatives before production, up from 17% a year earlier, and respondents said 46% of projects on average were scrapped between proof of concept and broad adoption. In McKinsey's State of AI survey for 2026, 44% of respondents said AI was scaling across their enterprise, and 37% attributed any EBIT impact to it. Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. Those three reasons are Blocks 10, 4 and 11 below.
Before you build, you should be able to show four things: the problem is priced, the data exists, someone will pay for it or adopt it, and the costs and the governance hold at scale.
Why Not Use the Business Model Canvas?
The Business Model Canvas has nine blocks: key partners, key activities, key resources, value propositions, customer relationships, channels, customer segments, cost structure and revenue streams. It works for most ventures. It has no block for three things that decide AI projects:
- Data. In an AI product, data is part of the product. A use case needs a plan for where the data comes from, how to start with too little of it, and whether use will make it better.
- Costs that grow with use. Inference is paid per call. At $0.02 a call, 1,000 calls a day costs about $7,300 a year; 100,000 calls a day costs about $730,000 (100,000 × $0.02 × 365). A price that worked in the pilot can fail at volume.
- Governance. Some uses are regulated as high-risk. Under the EU AI Act, systems that score creditworthiness or screen job applicants are listed in Annex III, and the high-risk rules for those uses apply from 2 December 2027 after the Digital Omnibus moved the date.
The Blueprint keeps the Business Model Canvas's best idea, one page that shows the whole case, and rebuilds the blocks around the ways AI projects fail.
What Are the Twelve Blocks?
Each block asks one question. Green means you can show the answer today.
The question each block asks
Foundation and Value Engine come first: a red there stops or pauses the work.
| Block | The question | Green means |
|---|---|---|
1. Problem definition | What business pain does this solve, and what does it cost today? | The pain is priced in money, time or risk, in the business owner's words. |
2. Data strategy | What data does it need, where does it come from, and how does it start with too little? | Someone has opened the data and checked it. |
3. Technology core | Build, tune an existing model, or call one through an API? | The simplest approach that works has been chosen, with reasons. |
4. Value proposition | What outcome does the user get that they cannot get today? | A finance director would accept it in one sentence. |
5. Market validation | Who wants this, and what proves it beyond enthusiasm? | Paid pilots or committed users. |
6. Competitive landscape | What else could the user do, and why can't a rival copy this? | The edge comes from data, workflow integration or domain depth. |
7. User experience | Who uses it every day, and what happens when it is wrong? | The design was tested with the people who will use it every day. |
8. Technical architecture | How does it fit existing systems, and how do you roll it back? | Integration, latency and rollback are planned for the production system. |
9. Go-to-market | How will people adopt it, and who leads the change? | There is an owner for adoption and a pricing or rollout plan. |
10. Unit economics | What does one request cost, and how does that change at volume? | The margin holds at ten and a hundred times pilot volume. |
11. Governance and compliance | Which rules apply, and what happens when it makes a serious error? | The obligations are mapped and an owner for each is named. |
12. Continuous improvement | How does it get better, and how will you know it is working? | Feedback loops, success measures and a review rhythm exist. |
How Do You Score a Use Case?
Score each block red (a gap that blocks the case), yellow (a gap with a plan) or green (shown). Then read the reds by pillar, because not every gap is equally fatal. Strength in one pillar does not make up for a red in another: a use case with a green Foundation and a red Execution still fails at launch.
What a red means in each pillar
| Pillar | A red here means | Action |
|---|---|---|
Foundation (Blocks 1 to 3) | The problem is not proven, the data is not there, or the technical approach is wrong. | Stop. Fix the foundation before any further spend. |
Value Engine (Blocks 4 to 6) | Users do not want it, the market is not there, or a rival can copy it. | Pause. Prove demand before building further. |
Execution (Blocks 7 to 9) | Gaps in the design, the integration or the rollout. | Proceed with care. Give each red an owner and a date. |
Sustainability (Blocks 10 to 12) | Costs, governance or improvement are not yet worked out. | Plan before scale. A pilot can run; scaling waits. |
What Does a Scored Example Look Like?
A hypothetical legal-technology company wants to sell AI contract review to the legal departments of large companies. The company and its numbers are invented for illustration, and the arithmetic is shown.
Contract review AI, scored (hypothetical)
| Block | Assessment | Score |
|---|---|---|
1. Problem definition | A target department spends about $2.4 million a year on routine contract review, at about four hours a contract. | Green |
2. Data strategy | 10,000 contracts already reviewed and labelled by the buyer's own process. Start on public legal datasets; every attorney correction feeds back in. | Green |
3. Technology core | An existing language model tuned on contract language, with retrieval over the buyer's clause library. Nothing trained from scratch. | Green |
4. Value proposition | Cut review time by 70%: about $1.7 million a year on a $2.4 million cost (0.7 × $2.4 million). | Green |
5. Market validation | Five legal departments interested; two signed paid pilots at $50,000 to $100,000 a year. | Green |
6. Competitive landscape | Three established rivals. The edge is industry-specific contract types and deeper workflow integration. | Yellow |
7. User experience | Clause-level highlights with confidence scores. An attorney approves or overrides every suggestion. | Green |
8. Technical architecture | Integrations with the main document systems still to build. Rollback planned. Target under 30 seconds a contract. | Yellow |
9. Go-to-market | Sales through legal-technology events and law-firm partners, on a per-seat annual licence. The sales cycle is unproven. | Yellow |
10. Unit economics | About $0.15 of compute a contract. At 500 contracts a month, about $900 a year per client (500 × 12 × $0.15), against $50,000 to $100,000 of revenue. | Green |
11. Governance and compliance | Buyers will ask for a SOC 2 report, data residency for clients abroad, GDPR terms for EU contracts, and explanations an attorney can check. None is in place. | Red |
12. Continuous improvement | Attorney overrides feed a monthly retraining; accuracy is audited every quarter. | Green |
Foundation and Value Engine have no reds, so the case passes the stop and pause tests. The three yellows (Blocks 6, 8 and 9) are ordinary work. The red is in Governance, which sits in Sustainability: the pilots can start, but governance runs as its own workstream from the first day, and no client goes beyond a pilot until it is green. A buyer's security review will ask for that evidence anyway.
Which Seven Anti-Patterns Does It Catch?
Seven ways AI use cases fail, and the block that catches each
| Anti-pattern | What it looks like | Caught by |
|---|---|---|
1. The technology-first trap | A model is built before anyone checks the problem or asks the users. | Blocks 1 and 5 |
2. Data denial | "We have lots of data" stands in for a data plan. | Block 2 |
3. The demo delusion | A demo on chosen data is taken as proof the product works. | Blocks 4 to 9 |
4. The moat mirage | Defensibility is claimed from the model or the prompts, which rivals can copy. | Block 6 |
5. The scale assumption | What worked in the pilot is assumed to work at volume. | Blocks 7 to 9 |
6. Economics ignorance | Inference costs are found only once they eat the margin. | Block 10 |
7. The governance gap | Compliance waits until a regulator, a customer or an incident forces it. | Block 11 |
The demo delusion is the one the headline names. A demo answers Block 3. Blocks 4 to 6 ask whether anyone needs the result, and Blocks 7 to 9 ask whether it survives real users and real systems. Score them separately and the gap shows.
Does This Work With Agile?
Yes. Treat it as a lens you come back to. Fill it in roughly in the first week to find the biggest unknowns, then update it as you learn. The reds become sprint priorities. A block that turns from green to red during development is the method working: you found the problem while it was still cheap. Fast iteration on the model does not fix a red in Block 1 or Block 5. It only gets you to the wrong answer sooner.
How Is This Different From the AI Readiness Canvas?
They answer different questions. The AI Readiness Canvas, AskAjay's readiness assessment, scores the organisation: whether it is ready for AI across five pillars. The model behind it is explained in the five-pillar AI readiness assessment. The Blueprint scores one use case. Run the readiness assessment once to find the organisation's weakest pillar; run the Blueprint on every use case that asks for money.
Where Do You Start?
Take your three most-discussed AI use cases and score each on all twelve blocks in one working session. Fund the ones with no reds in Foundation or Value Engine first.
Download: AI Use Case Blueprint Worksheet
Get the full 4-pillar, 12-block worksheet with Red/Yellow/Green scoring, the interpretation rules and space for notes, ready to print or save as PDF.
Enter your email to get instant access. You'll also get The Brief, every other Tuesday.
Free. No spam. Unsubscribe anytime.
For Block 11, the AI governance playbook and Minimum Viable Governance set up the governance the block asks for, and the AI GDPR compliance guide covers use cases that process personal data in the EU. For measuring the value in Block 4 after launch, see measuring AI ROI. Founders choosing their first use cases should also read the Founder's Playbook for Responsible AI. To work through a portfolio with me, see advisory.
Sources and Changes
- S&P Global Market Intelligence, 451 Research, "Generative AI experiences rapid adoption, but with mixed outcomes: Highlights from VotE: AI & Machine Learning", 30 May 2025. Survey of 1,006 IT and line-of-business professionals in North America and Europe.
- McKinsey, The state of AI in 2026: On the road to ROI, August 2026. Online survey of 1,719 participants in 97 countries, 4 May to 8 June 2026.
- Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", press release, 25 June 2025.
- Strategyzer, The Business Model Canvas.
- Regulation (EU) 2024/1689 (the AI Act), Annex III points 4(a) and 5(b), and Regulation (EU) 2026/1744 (the Digital Omnibus on AI), Article 1(40), which sets 2 December 2027 for Annex III high-risk systems. Both read in the Publications Office copies of the Official Journal on 4 October 2026.
Changes on 4 October 2026. I renamed the framework the AI Use Case Blueprint so that it is not confused with the AI Readiness Canvas. The URL is unchanged for now, and the blocks, pillars and scoring rules are the same. I cut the piece to a reference with question headings. Removed because no public source or record could be shown: the opening and closing story of a healthcare client (a $4 million budget, a project shelved after eighteen months and later rebuilt for four hospital systems), the root-cause chart attributed to "80+ AI initiative post-mortems" and its percentages, the statement that three initiatives became unprofitable as they scaled, and an anecdote about a computer vision team. The Gartner prediction for 2025 was replaced with the June 2025 prediction for 2027, since its horizon has passed. The McKinsey reference now cites the 2026 survey's figures, and the 451 Research figures are new. The CB Insights callout was removed: its startup-failure figures cover startups in every sector, so they did not support a claim about AI use cases. The worked example is the same contract-review case, now labelled hypothetical, framed as one vendor's case, and with its arithmetic shown. The worksheet's title and glossary were updated to match.
Questions readers ask
Is this the AI Use Case Canvas?
Yes. The framework was published under that name from January 2024 until October 2026. It was renamed the AI Use Case Blueprint so that it is not confused with the AI Readiness Canvas, which assesses a whole organisation. The twelve blocks, the four pillars and the scoring rules did not change.
How do I prioritise AI use cases with it?
Score every candidate on all twelve blocks. Set aside any use case with a red in Foundation, and pause any with a red in Value Engine until demand is proven. Rank the rest by their greens in those two pillars. Before any of them scales, ask for a written plan for each red in Sustainability.
Who should fill it in, and how long does it take?
The people who own the answers: the business owner for the problem, the value and the market, a data lead for Block 2, engineering for Blocks 3 and 8, finance for Block 10, and whoever owns risk and compliance for Block 11. A first pass should fit in one working session, because rough answers are enough to show where the reds are.
Get The Brief
Original thinking on AI strategy and governance. Every other Tuesday.
Senior AI strategist helping leaders make AI real across four continents. Forbes Technology Council member, IEEE Senior Member.
Ajay's views, from 15 years in the field. Not legal or compliance advice. See full disclaimers →
Published by AI Exponent LLC
