AI use-case selection depends on reallocating validation capacity, not a fixed ranking

The MAP function in NIST AI RMF 1.0 calls for defining business value, expected benefits and costs, application scope, and viable non-AI alternatives in context. [1] It does not prove profitability, but it supports describing a candidate by whose decision or outcome should change rather than by a function label such as summarization, search, or agent.

Begin the candidate sheet with the current decision, the person facing the problem, the desired result, the present process, and a non-AI improvement option. The hypothesis is not that AI can perform a task, but that an AI-enabled process can outperform the current or non-AI alternative. If the purpose is unclear, return to problem framing instead of advancing to an accuracy demo.

For example, ‘automate meeting minutes’ leaves open whether the problem is writing time, missing decisions, or delayed handoffs. ‘Deliver confirmed decisions, owners, and due dates to stakeholders before the next workflow begins’ creates a common outcome for comparing AI summarization, a better template, and a change in meeting practice.

• DECISION: State one decision or confirmed outcome to change
• OWNER: Identify who receives the benefit and owns quality
• BASELINE: Record time, quality, waiting, and failure in the current process
• ALTERNATIVE: Put a non-AI improvement option on the same candidate sheet

Record expected value, testability, harm, reversibility, and learning value separately. A candidate with high expected value but a multiyear outcome and irreversible failure may be a poor first test. A moderate-value candidate that can be compared quickly and teaches a shared data or evaluation capability can be a useful entry point.

The NIST Generative AI Profile separates risks across data, value chains, misinformation, information security, privacy, human interaction, and other groups. [2] The twelve groups do not apply with equal weight to every case, but they prevent a benefit-only ranking from dropping use-case-specific failure modes.

Avoid turning the five dimensions into a weighted decimal ranking. Doing so can offset a severe rights risk with a large benefit or fill an unmeasurable value with a convenient number. Show each dimension with its evidence, define noncompensable floors, and separate uncertainty that additional work can reduce.

• VALUE: Whose outcome may change, and on what time horizon?
• TESTABILITY: Can a comparator, observable outcome, and feasible duration be created?
• HARM: What is the impact on errors, rights, information, and continuity, and can it be remedied?
• REVERSIBILITY: Can the use case be narrowed, stopped, deleted, and returned to manual work?
• LEARNING: What reusable data, evaluation, or operating knowledge will the test produce?

In Dillon et al.’s experiment across 66 firms and 7,137 people, access to generative AI reduced email time among users without a detectable change in work volume or composition. [3] A broad rollout can reveal usage while failing to reveal which process changed and how it reached an outcome.

Instead of dividing candidates into enterprise rollout or rejection, use stages for problem confirmation, feasibility, comparative effect, and operating durability. First test whether required data can be accessed safely and whether representative work meets a quality floor. Only then move to comparison against the established process and reproduction across units.

Let each stage answer one question. Trying to prove satisfaction, financial ROI, and feasibility in the same test obscures the stop condition. Specify the smallest sample, duration, cost, and pass, hold, or stop rule for each stage, and remove measurements that cannot change the next allocation decision.

• FRAME: Confirm the problem, population, and non-AI alternative
• FEASIBILITY: Test data, quality floors, authority, and manual fallback
• EFFECT: Compare with the existing process through to a confirmed outcome
• OPERATE: Test updates, exceptions, monitoring, and continuing cost

Learning value is not a favorable impression. It is the extent to which a result can change the next choice. If the organization will proceed with a full rollout regardless of whether the result is positive or negative, the test does not improve the decision. A small test can be valuable when it removes a shared assumption across several candidates.

Suppose ten candidates depend on the same internal knowledge base. Before running ten demos, test source authority, permissions, freshness, and retrieval on one representative workflow. Failure redirects work to the foundation; success reduces uncertainty for several candidates. It still does not prove quality for every application.

Describe learning value through the number of shared assumptions, the possibility of reversing a decision, and the time at which the result becomes actionable. Even when quantified, state which candidate decision it changes. A longer study is not automatically more valuable; its result must arrive in time for the choice.

Use cases are not independent cards. They compete for the same expert reviewers, data foundation, legal assessment, and model contract, while one test can change the cost or risk of another. The portfolio sheet should connect shared resources, sequencing dependencies, and simultaneous capacity limits.

Starting every high-scoring candidate at once can overload scarce reviewers and lower test quality. Keep deferred candidates with a hold reason, required evidence, and review date. Show not only the cost of starting a candidate, but also the tests delayed during that period as an opportunity cost.

Divide capacity into explore, validate, and operate. Exploration holds a few uncertain but informative candidates; validation holds candidates with comparable outcomes; operations funds monitoring and change for accepted systems. Reserve capacity for stopping and retiring work, not only for new starts.

• SHARED: Connect common data, evaluation, people, and contract constraints
• SEQUENCE: Mark candidates that depend on earlier evidence
• CAPACITY: Set a simultaneous limit for expert review and monitoring
• OPPORTUNITY COST: Record the candidates and learning delayed by each selection

NIST AI RMF describes using measurement to prioritize risk response and deciding whether development or deployment should proceed given available resources and methods. [1] [2] This portfolio model transfers that logic to investment allocation; NIST does not certify these five dimensions or this review format.

At monthly or evidence-stage reviews, do not merely adjust rankings. Decide whether new evidence moves a candidate to scale, hold, redesign, or stop. Include business, workflow, data, quality, risk, and affected-user perspectives. A high score is not sufficient; reviewers need the evidence that changed and the uncertainty that remains.

A good portfolio is not a large candidate inventory. It places small resources on valuable questions, adds resources where evidence strengthens, returns candidates that do not hold, and transfers what was learned. Treat stopping as a completed decision, recover the budget, and reallocate both resources and knowledge.

• EXPAND: Add resources when the evidence stage passes and residual risk and operating capacity are acceptable
• HOLD: Attach a deadline and next question when decision-changing evidence is still missing
• REDESIGN: Return to problem framing when the objective remains but comparison, data, or operation fails
• STOP: End candidates that miss a floor or lose to a non-AI alternative, and recover resources and learning