Practical guide

How an operations excellence leader can build the case for applied AI capability evidence

An operations excellence leader needs a business case that compares credible ways to improve a bounded process. It should show the operational constraint, alternatives, resources and risks, missing evidence, and the smallest reversible decision that can produce better evidence. It should not begin with a tool purchase or end with an adoption number.

By Two Prune

An operations excellence leader needs a business case that compares credible ways to improve a bounded process. It should show the operational constraint, alternatives, resources and risks, missing evidence, and the smallest reversible decision that can produce better evidence. It should not begin with a tool purchase or end with an adoption number.

Use the CASE business-case method

CASE creates a decision record with four parts: Constraint, Alternatives, Supportable value, and Evidence-gated execution.

### C — Constraint and decision ownership

Define one end-to-end process, the user or customer outcome, the unit of analysis, and the decision required. Reconstruct the current process from system data, sampled cases, frontline observation, user research, quality review, resource records, and exception handling. Separate touch time, waiting, review, correction, escalation, and work outside the visible system.

The leader's authority varies by organization. They may own the improvement method and evidence packet while a process owner owns the outcome, finance reviews the resource case, technology owns the configured system, risk functions review their domains, and an executive sponsor authorizes investment. Do not infer budget authority from the role title.

Name the decision owner, process owner, operations excellence lead, frontline representatives, technology and data owners, security and privacy reviewers, finance partner, and person authorized to stop a test. Record whether the operations excellence leader is owner, reviewer, adviser, or collaborator for each decision.

NIST's Baldrige operations guidance encourages organizations to design work as processes, identify requirements, incorporate customer and supplier input, measure performance, find defect sources, and improve systematically. It is general performance-improvement guidance, not evidence that AI is the right intervention for a particular process. NIST Baldrige: Operations

### A — Alternatives before advocacy

Build a credible longlist before selecting a pilot:

For every option, state its mechanism: exactly how it could change the constraint. Record dependencies, implementation work, new operating work, affected people, failure modes, reversibility, control burden, and the information needed to appraise it. Do not weaken the non-AI comparator to make the favored option look attractive.

The UK government's Green Book describes appraisal as comparing the costs, benefits, and risks of options for achieving defined objectives. It also calls for proportional effort and a structured evidence basis for decision makers. It applies directly to specified public-sector proposals; here, its options discipline is used as a general reasoning aid rather than a claim that a private proposal follows public-spending rules. HM Treasury: The Green Book 2026

### S — Supportable value and full resource case

Create a value chain rather than a headline benefit:

capability input → changed work behavior → process change → operational outcome → organizational value

For each link, state the evidence available, the assumption, a plausible counter-explanation, and the owner of the next measurement. Keep capability, usage, control operation, process quality, and business outcome as separate constructs.

Show costs and resources across the lifecycle: case design, source preparation, tool configuration, licenses, integration, security and privacy review, participant time, reviewer calibration, supervision, monitoring, exception handling, incident response, model or vendor change, support, and exit. Include displaced work: time saved in drafting may reappear as review, correction, or escalation.

Use ranges or scenarios when the evidence cannot support a point estimate. Label unmeasured benefits, do not monetize them by guesswork, and show which conclusion changes under an adverse assumption. A financially polished spreadsheet is not a stronger case if its causal links are untested.

The GOV.UK Service Manual recommends choosing measures that fit the service scope and combining performance data with user research and other evidence. Although written for public services, it supports using a mixed evidence set and measuring the complete service rather than a single digital step. GOV.UK: Measuring the success of your service

### E — Evidence-gated execution and exit

Choose the smallest test that can change the decision. A capability exercise should require realistic work: framing an ambiguous case, finding authoritative sources, using AI in an inspectable way, detecting contradictions, verifying outputs, handling exceptions, revising, and delivering a useful artifact. Preserve the output and enough process evidence to explain it.

Write the gate before the test:

Set an exit route: data export and deletion, alternative operating process, contract boundary, responsible owner, and trigger. Reversibility is part of the value proposition, not an afterthought.

  • keep the current process and manage the known constraint;
  • repair policy, training, source quality, intake, staffing, or handoffs;
  • use deterministic rules, workflow automation, or conventional analytics;
  • test AI assistance for a narrowly defined step with human control;
  • combine the strongest non-AI and AI elements.
  • expand when the bounded capability, quality, control, and resource conditions are met;
  • repair and retest when the mechanism remains plausible but materials, configuration, measurement, supervision, or capability fail;
  • narrow when evidence supports only a smaller task, source type, population, or risk class;
  • hold when missing evidence prevents a sound decision;
  • stop when a prohibited condition occurs or the option no longer addresses the constraint.

Sources: [1] [2] [3]

Copyable option-adjusted business case

| Decision field | Current process | Process repair | Rules/automation | AI-assisted test | | --- | --- | --- | --- | --- | | Mechanism | | | | | | Evidence available | | | | | | Missing evidence | | | | | | Quality effect | | | | | | Review and exception work | | | | | | Data/control dependencies | | | | | | One-time resources | | | | | | Ongoing resources | | | | | | Affected people | | | | | | Failure and stop condition | | | | | | Reversibility | | | | | | Next evidence gate | | | | |

Attach a decision log with the selected option, rejected options, assumptions, reviewer disagreements, conditions, owner, review date, and evidence location.

Worked example: quality-incident root-cause summaries

A manufacturer has a recurring delay between a quality incident being logged and a cross-functional review receiving a usable root-cause summary. The initial proposal is to use generative AI to draft the summary from incident notes. The operations excellence leader maps 60 recent cases and finds that late summaries are associated with missing machine identifiers, inconsistent defect codes, photographs stored outside the incident system, and unclear ownership for requesting supplier evidence.

The business case compares four routes. The current process remains the baseline. A process-repair option adds required identifiers, a supplier-evidence owner, and a photo link. A rules option validates required fields and routes incidents by defect family. An AI-assisted option extracts evidence from the complete case file and drafts a source-linked chronology, while an authorized engineer retains the root-cause judgment.

The team implements required identifiers first because that repair benefits every option. It then uses a synthetic 24-case exercise for the rules and AI-assisted routes. Cases include a reused defect code, an image that contradicts the text note, a supplier certificate issued after the incident, and a maintenance event recorded in a separate source. Participants must identify the controlling evidence, state uncertainty, and escalate incomplete cases.

The first AI-assisted round creates readable chronologies, but several participants accept the stale defect code and one treats sequence as causation. Reviewer time also increases because source anchors are inconsistent. The disposition is repair and retest: enforce source anchors, add a causal-claim check, calibrate reviewers, and keep the field-validation improvement regardless of the AI decision.

The case does not record invented savings. It reports the resources used, completeness and contradiction-detection results, reviewer effort, missing evidence, and expected next measurement. Expansion requires reliable source traceability, correct escalation, acceptable review effort, and evidence from a controlled live comparison that downstream investigation quality does not decline.

Objections and reporting boundaries

“People already use AI” shows activity, not capability or value. “The draft is faster” ignores downstream review and rework. “The pilot scored well” does not establish live process impact. “The tool vendor owns the risk” does not remove process, data, or outcome accountability. “The estimate must be precise” can conceal uncertainty when the underlying evidence is immature.

Report the process, population, configuration, comparison, observation window, missing records, exceptions, reviewer disagreement, costs included, costs excluded, and next gate. Preserve null or adverse evidence. Do not generalize from one workflow to organization-wide capability or from a controlled simulation to realized business value.

Two Prune boundary

Two Prune can run realistic work simulations and provide evidence about observed performance in context framing, evidence navigation, AI orchestration, verification and risk control, judgment and synthesis, and communication. It does not create or approve a customer's business case, redesign operations, establish ROI, operate controls, assign budgets, or make employment decisions. Customers retain responsibility for process choices, systems, data, investment decisions, and use of assessment evidence.

Sources

  1. 1.NIST Baldrige: Operations · National Institute of Standards and Technology
  2. 2.HM Treasury: The Green Book 2026 · UK Government
  3. 3.GOV.UK: Measuring the success of your service · UK Government