Practical guide

How a Head of AI can build the case for applied AI capability evidence

A Head of AI should build a business case around a bounded portfolio decision, not around a promised productivity percentage or a generic need to “upskill everyone.” The case must compare credible options, identify who owns each decision, price the evidence work, and define what a pilot can and cannot establish. Use the BRIDGE method: Baseline, Required decision, Interventions, Delivery,.

By Two Prune

A Head of AI should build a business case around a bounded portfolio decision, not around a promised productivity percentage or a generic need to “upskill everyone.” The case must compare credible options, identify who owns each decision, price the evidence work, and define what a pilot can and cannot establish. Use the BRIDGE method: Baseline, Required decision, Interventions, Delivery, Governance, and Evidence gates.

B: Establish a decision baseline

Describe the current workflow before proposing an assessment programme. Record the task, participants, source systems, permitted AI assistance, review steps, error consequences, current quality evidence, time or cost evidence, and known missingness. Separate five layers that are often collapsed:

Do not infer capability from activity or business benefit from one assessment score. If current work quality is not measured, say so. The baseline can still identify a decision and an evidence gap without inventing a before-and-after result.

The OECD's employer and worker AI surveys deliberately collect complementary organizational and first-person perspectives. They are useful evidence that different views reveal different information, while their sampled countries, sectors, and self-reported measures do not establish a local organization's capability or return. OECD: AI surveys of employers and workers

  • Access: people can reach an approved tool.
  • Activity: the tool is used.
  • Capability: a person can perform a defined task under stated conditions.
  • Work quality: the resulting work meets a review standard.
  • Outcome: the work changes an operational or financial result.

Sources: [1]

R: Name the required decision and real owners

Write the decision in one sentence: for example, “Should the operations function fund a controlled assessment pilot for two claims-review tasks?” Name the executive sponsor, functional owner, assessment owner, data owner, privacy and security reviewers, finance reviewer, participant representative, and final release authority.

Classify the Head of AI as owner, influencer, reviewer, or collaborator for each part. The Head of AI may own portfolio standards and coordinate the pilot but usually cannot unilaterally define domain quality, approve data use, accept operational risk, or commit another function's budget.

NIST assigns governance and oversight to actors with management, fiduciary, and legal authority, while design, deployment, operation, domain expertise, and test/evaluation/verification/validation involve different actors across the lifecycle. Use that task map to expose missing owners rather than to create a universal organization chart. NIST: Descriptions of AI actor tasks

Sources: [2]

I: Compare interventions, including doing less

Compare at least three plausible options:

| Option | Problem addressed | Evidence gained | Cost categories | Risks and limits | Reversal point | | --- | --- | --- | --- | --- | --- | | maintain current approach | | | | | | | targeted workflow repair | | | | | | | bounded assessment pilot | | | | | | | broader programme | | | | | |

Include non-assessment alternatives. A source-quality problem may need a repository repair. A permission problem may need access redesign. An unclear approval route may need workflow clarification. Assessment is appropriate only when the decision requires evidence of participant performance on a representative task.

HM Treasury's current business-case guidance uses five connected dimensions: strategic, economic, commercial, financial, and management. It is public-sector spending guidance rather than a universal company template. Borrow its discipline by covering case for change, alternatives, delivery route, affordability, and implementation instead of presenting only benefits. HM Treasury: Guidance on developing business cases

Sources: [3]

D: Design a proportionate pilot and cost model

Define the smallest pilot that can answer the decision. Specify tasks, participant group, sampling route, permitted tools, synthetic or approved source materials, rubric, reviewers, adjudication, accessibility provisions, data handling, exclusions, and stopping conditions.

Price the full evidence path: design, source preparation, participant time, reviewer time, adjudication, security and privacy review, analysis, remediation, and decision meetings. Separate one-time setup from recurring operation. Do not count speculative time savings as realized value. Present a range and list the assumptions that move it.

Use a pilot register:

| Hypothesis | Measure | Baseline | Decision rule | Missingness rule | Confound | Owner | | --- | --- | --- | --- | --- | --- | --- | | | | | continue / repair / hold / stop | | | |

NIST's AI RMF Core calls for roles, policies, contextual mapping, measurement, risk prioritization, and resource allocation, while noting that organizations tailor the voluntary framework to their context. This supports pricing governance and evidence work alongside the assessment itself. NIST AI RMF Core

Sources: [4]

G: Define governance, objections, and failure routes

List foreseeable objections before approval: task representativeness, accessibility, reviewer consistency, privacy, data retention, participant communication, weak benchmarks, manager misuse, tool drift, and pressure to convert evidence into an employment recommendation. Assign an owner and response to each.

Set the use boundary in the business case. Results support interpretation of performance in the defined simulation; they do not establish general intelligence, future job performance, causal productivity, or a universal workforce ranking. Require human review and prevent automated employment decisions. Define who may access individual and aggregate evidence and when it is deleted.

Set explicit stop conditions: source materials are not authorized, accessibility needs cannot be met, reviewers cannot apply the rubric consistently, the task does not resemble real work, missingness is being scored as failure, or stakeholders intend to use results outside the approved purpose.

E: Commit to evidence gates, not promised outcomes

Make the approval incremental. A design gate confirms decision, task, sources, owners, and use boundary. A readiness gate confirms materials, permissions, reviewers, and technical controls. A pilot gate confirms enough interpretable evidence to evaluate the approach. A scale gate considers cost, operating capacity, unresolved risks, and whether another method is better.

Report pass, repair, hold, and stop dispositions. Preserve null findings and adverse evidence. A pilot that reveals an invalid task or an impractical review burden can still answer the business-case decision without proving a positive return.

Worked artifact: a synthetic claims-review proposal

A synthetic professional-services firm wants evidence that consultants can verify AI-assisted proposal claims. The first draft requests a company-wide programme and assumes a 15% productivity gain. BRIDGE removes the unsupported outcome claim and compares four options.

The baseline finds approved-tool access and high activity, but no consistent record of source reconciliation. A workflow-only option adds a claim-source table. A coaching option provides worked examples. A bounded assessment pilot uses two synthetic proposal tasks, an approved source pack, observable verification behaviors, two reviewers, and adjudication. The broad programme remains an option only after the decision evidence is stronger.

The Head of AI owns portfolio coordination and evidence standards; the proposal director owns task quality; security and legal reviewers own their domains; finance reviews affordability; an executive sponsor decides whether to proceed. The pilot costs are shown as ranges, and the decision rule is not “scores improve.” It asks whether the task is representative, the rubric is usable, reviewers reach acceptable agreement, failure evidence is actionable, and the operation is affordable.

The case approves a 24-participant pilot with repair and stop gates. It does not approve company-wide rollout or claim productivity. The example is synthetic and supplies no evidence about Two Prune's customer outcomes.

Claims and product boundary

This method structures a business case; it does not prove demand, return on investment, fairness, validity, compliance, or operational improvement. A local decision requires current organizational evidence and authorized reviewers. Two Prune provides realistic work simulations and task-bounded assessment evidence for human interpretation. It does not make employment decisions or promise business outcomes.

Sources

HM Treasury: Guidance on developing business cases

  • NIST AI RMF Core
  • NIST: Descriptions of AI actor tasks
  • OECD: AI surveys of employers and workers

Sources: [3] [4] [2] [1]

Sources

  1. 1.OECD: AI surveys of employers and workers · www.oecd.org
  2. 2.NIST: Descriptions of AI actor tasks · National Institute of Standards and Technology
  3. 3.HM Treasury: Guidance on developing business cases · UK Government
  4. 4.NIST AI RMF Core · National Institute of Standards and Technology