Practical guide

Building the business case for applied AI capability evidence as an organizational development leader

An organizational development leader can build a credible business case by connecting one changing workflow to one decision about work design, then pricing the smallest evidence programme that reduces that decision's uncertainty. The case should not promise generic productivity, treat training attendance as capability, or assume the organizational development function owns technology, risk,.

By Two Prune

An organizational development leader can build a credible business case by connecting one changing workflow to one decision about work design, then pricing the smallest evidence programme that reduces that decision's uncertainty. The case should not promise generic productivity, treat training attendance as capability, or assume the organizational development function owns technology, risk, finance, or operations. Use a CHANGE case: context, hypothesis, alternatives, needed evidence, governance, and economic boundary. The output is a decision packet for an authorized sponsor, not a sales narrative for assessment volume.

Context: start with the work-design decision

Name the workflow, affected people, current outcome, trigger for change, source of the problem, and decision due. Describe what AI changes in tasks, handoffs, judgment, review, exceptions, and accountability. Do not frame “low AI fluency” as the problem until the underlying work has been diagnosed.

The CIPD's organization development and design standards describe work design, operating models, diagnostics, evidence, behavior, culture, systems, and metrics as connected concerns. They also recognize evidence from interviews, observation, focus groups, surveys, mapping, and organizational data. These professional standards support a systemic diagnosis; they do not establish that a particular AI programme will improve performance. CIPD: Organisation development and design

Write the decision as a sentence: “Should the customer-operations director fund a bounded test of AI-assisted complaint summarization, with specialists retaining category and escalation decisions?” Name the organizational development leader as owner, influencer, reviewer, or collaborator for each part. The title alone does not establish authority.

Sources: [1]

Hypothesis: show the causal chain and its breakpoints

Build a chain that can be challenged:

Beside every link, list a rival explanation. Poor assessment performance could reflect unclear instructions, inaccessible materials, unfamiliar tooling, insufficient opportunity, or a flawed task—not capability alone. Strong assessment performance does not prove live transfer, customer impact, or safe deployment.

  • a defined task or handoff is changing;
  • people need specified observable capabilities and support conditions;
  • a realistic assessment can reveal whether those capabilities appear in a controlled simulation;
  • the result can inform training, work design, or a pilot decision; and
  • later operational evidence, collected separately, can test whether the redesigned workflow performs as intended.

Alternatives: compare four legitimate choices

Every case should compare at least:

Score alternatives against the actual decision: evidentiary fit, participant burden, design effort, accessibility, privacy, review capacity, reversibility, and what the method cannot show. A business case is weak if assessment is the only option permitted to look useful.

The CIPD's evidence-based practice guidance draws on organizational data, scientific literature, stakeholder views, and practitioner expertise. Use those streams separately and record disagreement rather than collapsing them into one confidence score. CIPD: Evidence-based practice

  • continue the current workflow and measurement;
  • improve the process without an assessment;
  • use a lighter evidence method such as observation, work samples, or a manager review; and
  • run a bounded applied assessment linked to the target workflow.

Sources: [2]

Needed evidence: define the minimum decision packet

Specify evidence before estimating cost:

| Evidence layer | Example | What it can support | What it cannot prove | | --- | --- | --- | --- | | Work diagnosis | process map, exception sample, interviews | where work and controls may change | individual capability | | Applied assessment | observed source use, AI interaction, verification, deliverable | performance in the stated simulation | live operational outcomes | | Learning evidence | practice completion, feedback, repeated task | progress under the learning conditions | durable transfer without follow-up | | Operational evidence | quality, rework, review effort, escalations, workload | behavior of a bounded live workflow | broad organization-wide impact | | Stakeholder evidence | worker, manager, customer, risk-owner feedback | experienced effects and missing issues | causal attribution by itself |

Include denominators, missingness, prohibited failures, comparison conditions, and review authority. State how accessibility needs and reasonable participation conditions will be handled. Do not use assessment evidence to make an automated employment decision.

Governance: attach owners and stop conditions

Create a responsibility table for the business sponsor, workflow owner, organizational development lead, learning lead, technology owner, privacy and security reviewers, accessibility reviewer, finance partner, participant-support owner, and evidence analyst. For each, record decision rights, deliverable, and escalation route.

The NIST AI Risk Management Framework calls for defined roles, documentation, human oversight, context mapping, measurement, monitoring, and treatment of third-party AI risk. It is voluntary guidance and does not certify the proposed programme. Use it to test whether ownership and evidence routes are explicit. NIST AI RMF Core

Predeclare stop conditions, such as exposure of restricted data, an inaccessible task, a material scoring defect, insufficient reviewer agreement, a prohibited control failure, or evidence that the simulation does not represent the intended work. A pause is a valid outcome, not an embarrassment to hide in an average.

Sources: [3]

Economic boundary: price the decision, not an imagined transformation

Separate one-time design costs from recurring delivery and review costs. Include workflow research, synthetic material production, subject-matter review, accessibility work, privacy and security review, participant support, platform configuration, scoring review, reporting, remediation, analysis, and follow-up measurement. Show internal time as well as vendor spend.

Do not book benefits before the corresponding evidence exists. At the business-case stage, describe a benefit hypothesis and the observation needed to test it. Report a range and a break-even condition rather than a false-precision return on investment.

For example:

> The programme is worth testing if evidence from 24 representative simulations can resolve whether the team needs verification practice, work redesign, or both before a live shadow pilot. It is not justified by a claim that assessment itself saves handling time.

Worked CHANGE case: complaint-summary review

A synthetic service organization wants specialists to review AI-drafted complaint summaries. Managers report uneven quality, but dashboard usage is high. The organizational development leader diagnoses a work-design question: which evidence must specialists inspect before accepting a summary, and where should product-safety cues escalate?

| CHANGE element | Recorded decision | | --- | --- | | Context | One complaint-summary workflow; safety classification stays with specialists | | Hypothesis | A simulation with conflicting product notes can reveal source checking, uncertainty, and escalation behavior | | Alternatives | manager observation is cheaper but misses hidden source use; a knowledge quiz lacks work performance; process repair alone does not reveal individual workflow behavior | | Needed evidence | 24 simulations, hard-case sample, source and revision logs, final artifact, reviewer agreement, missingness report | | Governance | operations owns the workflow; organizational development owns diagnosis and facilitation; product safety owns escalation rules; privacy reviews materials; finance reviews cost | | Economic boundary | 38–52k for design, delivery, review, and analysis; no productivity benefit booked |

The gate has four outcomes: stop because the task is invalid; repair instructions or supports and rerun; proceed to targeted practice; or run a limited shadow pilot if capability evidence and control behavior are adequate. Operational quality, review effort, rework, worker experience, and customer effects are measured later and are not attributed to the assessment alone.

Objections the case must answer

“We already have adoption analytics.” Usage can show activity, not whether people inspect evidence, catch an unsupported claim, or escalate correctly.

“Managers can tell who is good.” Observation may contribute contextual evidence, but opportunity, task difficulty, and unseen work vary. Use a common simulation when the decision requires comparable observed behavior.

“Why not train everyone?” A broad course may be appropriate, but it should not substitute for defining the changed task and evidence need. Assessment can identify a bounded uncertainty; it does not automatically justify individualized intervention.

“Does this prove a return?” No. The case funds a decision-learning step. Claims about operational or financial outcomes require later evidence with a credible comparison and stated confounds.

Reporting boundary

Report the exact workflow, task version, participant opportunity, support, sample, missing cases, reviewer process, confidence limitations, and permissible decisions. Keep critical failures visible. Label percentiles unavailable or directional when benchmark evidence is immature.

Two Prune can provide evidence from realistic AI-enabled work simulations, including source use, AI interaction, verification, revision, escalation, and the final deliverable. It does not prove organization-wide capability, establish a programme outcome, redesign the customer's operating model, or make an employment decision.

Sources and limits

The CHANGE method, costs, organization, and example are original and synthetic. This page provides a decision method, not evidence of Two Prune customer outcomes.

  • CIPD: Organisation development and design — professional role, work-design, diagnostic, and systems context.
  • CIPD: Evidence-based practice — multiple evidence streams and problem diagnosis.
  • NIST AI RMF Core — roles, context, oversight, documentation, measurement, and monitoring prompts.

Sources: [1] [2] [3]

Sources

  1. 1.CIPD: Organisation development and design · www.cipd.org
  2. 2.CIPD: Evidence-based practice · www.cipd.org
  3. 3.NIST AI RMF Core · National Institute of Standards and Technology