Starting question
The project thesis must survive a chain of evidence.
Who uses the system, which activity changes and what observable outcome follows? If these elements are not distinct, model accuracy, user counts or commercial interest do not establish project value.
Technical and business due diligence meet in dependencies: the data, suppliers, people, permissions and costs that make performance possible.
Operational relations
Five steps that must remain connected.
Thesis
What relationship between user, problem and outcome is claimed?
Evidence
Which data or observable behaviour supports the claim?
Dependencies
What must remain available, authorised and sustainable?
Risk
Which failure changes value, safety or the right to operate?
Milestone
Which uncertainty will the next proof remove?
Boundaries and responsibility
Accuracy, benchmarks and ROI only have meaning in context.
A score may depend on an unrepresentative dataset; savings may omit human checking and operating costs; growth may be supported by free trials or manual work. Every number needs its method and boundary.
Evaluation must also distinguish owned, licensed, API-accessed and hypothetical components. Current availability cannot be inferred from a roadmap.
Practical object
AI Project Evidence Canvas
Complete the AI Project Evidence Canvas. Each field exposes a relationship to verify before extending the system.
User
Who adopts, uses, pays or experiences the effect?
Problem
Which work or decision matters enough?
Outcome
What changes in an observable way?
Evidence
Which data, tests or real use supports the thesis?
System
Which parts are owned, external or still manual?
Dependencies
Which data, suppliers, people and permissions are critical?
Risks
Which errors, constraints or concentrations change value?
Milestone
Which next proof reduces the decisive uncertainty?
First test
A short test should produce knowledge, not merely an output.
- 01Reconstruct the thesis
Write user, problem, activity and outcome without using the product name.
- 02Request primary evidence
Inspect data, examples, protocol and failure cases rather than the summary alone.
- 03Decompose the system
Separate model, data, software, people, services and external components.
- 04Stress one dependency
Simulate higher cost, supplier change, poorer data or loss of a key person.
- 05Define the milestone
Connect resource and time to removing a verifiable uncertainty.
Public sources
References for verification and further work.
- NIST AI Test, Evaluation, Validation and Verification.
- NIST AI 600-1, Generative AI Profile.
- UK Government, Guidelines for AI procurement.
These sources support initial design. Legal, professional, ethical and organisational requirements depend on the case and the responsible functions.
Frequently asked questions
Investor or project evaluator: questions to clarify before the project.
What evidence should an AI project provide?
Representative examples, evaluation method, baseline, failures, human intervention, dependencies and behaviour in context. A demo is one piece of evidence, not the full case.
How should an agentic system be evaluated?
Observe actions, tools, state, permissions, stopping and recovery—not only the quality of generated text.
Is a benchmark sufficient?
No. Its dataset, metric, comparison, representativeness and relationship to real work must be known.
How can economic benefit be verified?
State the baseline, period, included costs, human checking, infrastructure and assumptions. Without them, value remains a claim to test.
What makes a milestone useful?
It removes a decisive uncertainty through an observable result rather than merely recording activity.