Practical guide

How a digital transformation leader can build the case for applied AI capability evidence

A digital transformation leader should build the business case around a service decision and evidence gap, not around buying more assessment or training. The case needs alternatives, current-state evidence, a bounded capability question, operating dependencies, cost and burden ranges, safeguards, and a staged decision with exit conditions.

By Two Prune

A digital transformation leader should build the business case around a service decision and evidence gap, not around buying more assessment or training. The case needs alternatives, current-state evidence, a bounded capability question, operating dependencies, cost and burden ranges, safeguards, and a staged decision with exit conditions.

Classify authority and the decision

Classify the digital transformation leader as owner, influencer, reviewer, or collaborator for each decision. Service owners may own outcomes; learning may own support; security and privacy own their controls; finance and procurement own investment and contracts. Do not assume the transformation title controls every budget or production release.

The UK Government Digital, Data and Technology Playbook emphasizes outcomes, user needs, multidisciplinary teams, iterative delivery, commercial planning, and lifecycle management. HM Treasury’s Green Book requires credible options and proportionate appraisal of costs, benefits, risks, and uncertainty. These sources support a staged business case but do not establish that Two Prune or any AI programme produces business outcomes. UK Government: Digital, Data and Technology Playbook HM Treasury: The Green Book

Sources: [1] [2]

Build the case with VALUE

### V — Value decision and current baseline

Define the service or workflow decision, affected users, current performance, decision date, and accountable owner. State which evidence is missing and why existing measures cannot answer the decision.

Separate activity metrics such as licences, logins, prompts, course completion, and prototypes from work evidence and service outcomes. Describe current quality, rework, delay, incidents, accessibility, user experience, and review burden where available. Do not invent a baseline to make the case complete.

### A — Alternatives and additionality

Compare no change, process redesign without AI, improved access or guidance, targeted learning, manager observation, work-sample review, a realistic task-based assessment, and combined approaches where relevant.

Explain what the proposed evidence adds beyond existing data. A new assessment is not additional if current work samples already answer the question at lower burden. Conversely, usage data may not reveal whether people can detect unsupported claims, verify calculations, or manage uncertainty.

### L — Logic, limitations, and learning path

Write a logic chain without pretending it proves causation:

List competing explanations and design the next test to distinguish them. State sampling, task fidelity, review, accessibility, missingness, and interpretation limits.

### U — Unit economics, operating burden, and uncertainty

Estimate ranges for design, materials, participant time, reviewer preparation, scoring, adjudication, support, accessibility, data handling, reporting, and follow-up. Show assumptions, owners, time horizon, and sensitivity. Separate fixed setup from per-participant cost.

Do not claim ROI without a credible benefit measure, counterfactual, and attribution design. The first case may be feasibility and decision usefulness rather than financial return.

### E — Evidence gates, execution, and exit

Stage the request: discovery, design, bounded pilot, repair, broader test, transfer study, or stop. Pre-register thresholds for evidence usefulness, reviewer agreement, burden, safeguards, technical reliability, and interpretation.

Use explicit dispositions: proceed, proceed with conditions, revise, investigate another cause, hold, stop, or retire. Assign every stage an owner, evidence packet, budget range, and review date.

  • A defined workflow requires observable AI-enabled practices.
  • Current evidence cannot distinguish capability from access or adoption.
  • A bounded realistic task can sample specified practices under controlled conditions.
  • Results may inform learning, support, workflow, or measurement decisions.
  • Workplace transfer and operational impact require separate evidence.

Business-case evidence map

| Decision question | Existing evidence | Gap | Proposed evidence | Boundary | Owner | | --- | --- | --- | --- | --- | --- | | exact choice | usage, learning, work sample, service metric | unresolved alternative | bounded method | what it cannot establish | named role |

Attach the outcome definition, authority map, alternatives, current baseline, task brief, operating model, cost ranges, safeguards, risk and exception register, thresholds, and next evidence gate.

Worked case: service-request triage

A synthetic organization wants to expand an AI assistant for internal service-request triage. Eighty percent of target users have activated the tool and most completed training. Leaders infer that scale will improve resolution time.

VALUE reframes the decision: should the service owner fund a controlled capability-and-workflow test before broader deployment? Current evidence shows access and participation, not correct routing, escalation, verification, or service impact.

Alternatives include improve the request taxonomy, revise the form, add manager observation, review sampled work, run a realistic task, or combine taxonomy repair with a bounded task. Existing operations data show frequent reassignment between access incidents and account provisioning, suggesting the taxonomy may be a competing cause.

The proposed test uses approved synthetic requests with ambiguous categories, conflicting clues, and source-date changes. It records source review, AI interactions, corrections, escalation, final routing, and explanation. Reviewers measure consequential error patterns, correction burden, agreement, accessibility issues, and fallback operation.

Costs are presented as ranges for task design, taxonomy expertise, 40 participant-hours, reviewer calibration, double-scored sample, adjudication, technical support, and reporting. No productivity benefit is booked. A sensitivity shows that reviewer burden doubles if the rubric combines routing accuracy and explanation quality.

The funding request is design and bounded pilot, not full rollout. Gates require approved data handling, frozen taxonomy, accessible task, reviewer agreement, interpretable failure patterns, workable burden, and a decision-useful report. Stop conditions include unresolved authority, unreliable evidence capture, disproportionate burden, or duplication of existing work evidence.

The business case records an objection that simulated work may not transfer. It agrees and assigns a later sampled-work gate only if the pilot is useful. The artifact contains the current-state evidence, authority map, alternatives, logic, cost ranges, task and review plan, safeguards, thresholds, decision, and next date.

The case does not claim productivity, ROI, workforce readiness, or operational improvement. It supports a decision to collect better evidence.

Failure and hold conditions

Hold the case when the service decision is unclear, transformation activity substitutes for outcomes, the evidence gap is not demonstrated, alternatives are token, costs omit review and support, benefits are invented, transfer is assumed, authority is missing, or thresholds are written after results.

Two Prune provides realistic AI-enabled work assessments and evidence about observed task performance. It may supply one evidence layer in a staged transformation case; it does not run the transformation, prove business impact, or make employment decisions.

Sources

UK Government: Digital, Data and Technology Playbook

  • HM Treasury: The Green Book

Sources: [1] [2]