The AI Value Canvas is a practical framework for deciding whether an AI project is worth pursuing — and for designing it responsibly before code is written.
It brings business, technical, risk, and operational thinking into one conversation. The aim is not to make every idea look viable. It is to make the assumptions visible early enough to improve, pause, or stop the project before more time and money are committed.
Ten elements, four domains
The Canvas organizes ten elements into four domains. Each domain answers one question; each element gives the team something specific to work out.
| Domain | Key question | Elements |
|---|---|---|
| Customer Understanding (Desirability) | Do users want this? | Customer/User Segments · Jobs to Be Done · Pains & Gains |
| AI Opportunity & Solution (Feasibility) | Can we build it this way? | AI Opportunity Fit · Solution Concept / Interaction Flow |
| Data, Model & Risk (Sustainability) | Can we do it responsibly? | Data & Model Considerations · Risks & Guardrails |
| Value & Cost (Viability) | Does it make economic sense? | Value & Metrics of Success · Cost Drivers · Build / Buy / Partner |
The elements are connected. A weak answer in one domain should change the decisions in the others. For example, if users will not trust automated recommendations, the Solution Concept may need an approval step. If the needed data is inaccessible, the value estimate should not be treated as achievable. If token costs rise sharply at production volume, an agentic workflow may no longer make economic sense.
The project context that sits around the Canvas
The ten elements describe the opportunity. Four metadata fields describe the environment in which it has to succeed:
- Strategic Intent — the organizational objective this project supports. It should be specific enough to test, not a general ambition such as “become AI-enabled.”
- Project Owner — the executive sponsor, project lead, technical lead, and business owner accountable for adoption and value realization.
- Horizon — whether this is an Experiment, Pilot, or Scale initiative. Each stage needs different evidence, investment, and success criteria.
- Organizational Readiness — leadership commitment, change capacity and skills, and data infrastructure maturity. Score each from 1 to 5. Before comparing projects, agree what 1, 3, and 5 look like for each field rather than relying on a general impression that the organization is “ready.”
These fields are designed to test whether a technically promising idea has the sponsorship, operating ownership, and organizational capacity needed for adoption.
Work from left to right — then revisit
Start with the user, not the technology. Work through the Canvas in this sequence:
- Understand the customer. Define the primary user, the job they are trying to get done, and the pains or gains that matter in their current context.
- Test the AI opportunity. Ask whether AI is necessary. Compare it with rules, a simpler model, a process change, or hiring a person. Then design how people will interact with the output.
- Assess data, model, and risk. Confirm that the necessary data exists and can be used. Choose the simplest model approach that meets the need. Name the risks, owners, and guardrails before launch.
- Test value and cost. Define a baseline, a measurable target, and the full cost of building and running the solution. Decide whether to build, buy, or partner.
This is not a one-way process. A Canvas is a living document. Revisit it when you learn something material: after user research, a data-quality assessment, an Experiment, a Pilot, a vendor change, a major model update, or a production incident.
What v3.0 adds
Version 3.0 adds the following framework changes.
Organizational Readiness
Organizational Readiness records leadership commitment, change capacity and skills, and data infrastructure maturity for discussion in the Canvas. Teams score each field from 1 to 5. Treat a low score as a signal to change the Horizon, reduce scope, or invest in readiness first.
Build / Buy / Partner
The Canvas makes the delivery choice explicit. Assess building, buying, and partnering against proprietary data and workflow differentiation, product maturity, internal and external capability, cost, time-to-value, data exposure, vendor dependency, and post-launch skills.
Agentic AI, with a higher bar for proof
For this Canvas, an agentic system is one in which a large language model (LLM) dynamically directs a multi-step process and tool use to pursue a task. A single-turn chatbot that only returns a response is not treated as agentic here.
These systems are useful when the steps genuinely vary by task. They add failure modes: errors can compound across steps, an agent can misuse a tool, and retries or long chains can increase costs (Anthropic, “Building Effective AI Agents,” 2024; OWASP, “AI Agent Security Cheat Sheet,” 2025). Retrieval-augmented generation (RAG) incorporates retrieved information into response generation (Lewis et al., “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks,” 2020). Before choosing an agentic approach, test whether a single model call, RAG, or a fixed workflow meets the task at lower complexity and acceptable cost and risk.
Evaluate agentic systems against defined task-success criteria using typical and deliberately misleading test inputs. For high-impact actions, use human approval, cost limits, action logs, and permissions limited to each tool’s necessary access (Anthropic, “Building Effective AI Agents,” 2024; OWASP, “AI Agent Security Cheat Sheet,” 2025).
A pre-mortem and an exit plan
The Risks & Guardrails element now includes a pre-mortem: “It is 12 months from now and this project failed. What happened?” The team uses the answers to identify assumptions, risks, and guardrails that are missing from the current Canvas.
The Canvas also asks what happens if a vendor changes its pricing, deprecates a model, or no longer meets the organization’s needs. A credible project has a dependency register, a migration path, and an understanding of what switching would cost.
Score anchors, not impression scores
A score of 3 should mean roughly the same thing across different teams. Before comparing projects, define what 1, 3, and 5 look like for each scored field. Use these anchors to make trade-offs visible and to avoid comparing one team’s optimism with another team’s evidence.
What a completed Canvas should help you decide
A useful Canvas does not promise a “go” decision. It gives the team enough shared evidence to choose one of four next moves:
- Proceed — the problem, approach, data, risks, and economics are sufficiently understood for the current Horizon.
- Test an assumption — a specific uncertainty needs an Experiment before the project can progress.
- Rescope — the opportunity may be viable at a smaller scope, with a simpler model, or with a different interaction design.
- Stop or defer — the project is not currently justified. Record why, so the decision can be revisited if the data, technology, or strategic context changes.
The goal is clear thinking before commitment. A Canvas is successful when it helps a team pursue the right project with appropriate safeguards — or avoid spending on the wrong one.