Before you build the AI, price the problem
Most AI business cases fail the first question a CFO asks, and most AI specifications leave the builder inventing the answers nobody wrote down. These are the frameworks we use to stop both.
They are free, and they are the same ones we run inside engagements. The judgement is the part we get hired for. The method is not a secret.
The six frameworks
Each one stands on its own. Together they are the spine of how an AI build gets costed, specified, tested and proven.
They run in the order a build actually happens. Price the problem first: if nobody can put a number on what the pain costs, the project is an opinion, not a business case. Then decide what kind of value the build returns, because capacity released is not the same as headcount removed, and the CFO knows the difference. Put one payback number in front of the argument and let the assumptions show. Only then write the specification, and write it so the model proposes while the system decides. Prove the prototype against a pass mark agreed before anyone builds, and put the guardrails in the system, not in the prompt.
The PRD and TRD templates carry the same discipline into the documents your builders work from. Inside an engagement the six sit in the Build act of the KeyDelta Method, after operators have fixed what the AI would otherwise amplify. That is the order AI implementation runs in here.
A problem without a price is an opinion
The test that separates a cost you can defend from one you cannot: can you point at the invoice, the contract, or the payroll line?
Twelve people saving twenty minutes a day is not a headcount
The five value types, and the question that decides whether returned hours are money or just relief.
One payback number hides the argument
State payback at three levels of evidence, then name the single figure that breaks the case if it is wrong.
The model proposes. The system decides.
The architecture patterns that hold, stated at the level of shape rather than product, so they travel to any stack.
A prototype without a pass mark always passes
Eight things to settle before anyone opens a laptop, including the arm that detects a reviewer who rubber-stamps.
A guardrail in a prompt is a wish
Enforcement points, hostile inputs, and why three worked examples in a document are not a test set.
Want someone in the room while you run it?
The frameworks are the easy part. The hard part is asking the exception question twice, and telling a CFO their number is soft. That is the work we do.
Book a 30-minute call