Leading the Change to AI-Enabled Operations

What must change so the new operation can work?

A team can understand an AI initiative, complete its training and still have good reasons to keep the old routine. Checking outputs may take longer than expected. Another department may reject the new handoff. A manager may ask for judgment while rewarding volume. The change reaches the employee as an additional obligation.

Leaders need to ask: what must change in the work, management routines and employee experience so this operation is used appropriately and delivers sustained outcomes? Individual readiness matters. So do workload, incentives, trust, authority, support and the consequences of speaking up.

A randomized field experiment across 66 firms found that individual access to an integrated AI assistant changed some work patterns, including email time, without a detected change in the quantity or composition of tasks. That is a bounded finding about individual assistance, not autonomous operations. It gives leaders reason to examine the surrounding work when interpreting tool adoption. [1]

The framework

Apply this method to one AI-enabled operation: a defined flow of work with an accepted outcome, accountable owners and agreed authority. Begin with its proposed design, current performance, affected roles and governing conditions. Produce an adoption decision record connecting the transition to evidence and the next commitment.

Redesigning Work Around Outcomes establishes what the operation should become. Governing AI-Enabled Operations establishes how it is authorized and reviewed. This framework establishes how people move into that operation, work within its boundaries, and sustain or challenge it. Findings can reopen either design or authorization.

The five activities below are a repeatable decision aid. Revisit them when the work, evidence or employee experience changes. They do not form a maturity ladder or grant permission to operate.

Five activities for leading the change

1. Understand the consequences with the people affected

Walk through representative work with users, reviewers, receiving teams and support functions. Record what each role gains, loses or retains: tasks, discretion, responsibility, status and opportunity. Include informal coordination and checking that may disappear from a process diagram but still protect the outcome.

Ask people to demonstrate friction and test the proposed route. A concern about quality, fairness, workload or job implications deserves investigation. Give each concern an owner, response date and route for challenge outside the immediate reporting line where needed. Explain how feedback will be used and who can see it; a promise of participation means little if objections damage performance ratings.

OECD surveys in manufacturing and finance associated worker consultation and training with better reported outcomes. They also recorded concerns about job loss. These associations support taking employee experience seriously; they do not prove that consultation causes success. [2]

Output: an agreed consequence map, disputed assumptions and changes to investigate before commitment.

2. Secure the commitments that make the change credible

The Sponsor's job is to resolve commitments beyond the operating team's authority. Specify the mandate: funding, protected transition capacity, incompatible targets, or obligations across departments. Record what will be provided, by whom and when. If the Sponsor lacks the necessary authority, name the executive or forum that holds it.

Other responsibilities remain explicit. The Outcome owner defines acceptance and owns the result. The Operating owner organizes delivery, support and barrier resolution. Managers allocate work and coach. The Knowledge owner addresses source problems; the Technology owner addresses system reliability; the Control lead addresses controls within the assigned mandate. Sponsor status does not automatically confer policy-exception or risk-acceptance authority.

Examine which measures make the old routine attractive: tickets closed before receiving acceptance, utilization that leaves no practice time, or rewards for avoiding escalation. Agree changes with the owners of those measures. Communicate what is changing and why, what remains uncertain, who decides, and when employees will hear the next answer. Show which feedback changed the proposal.

Output: funded commitments and revised expectations that managers can actually honor.

3. Prepare people and test a bounded transition

Define the trial's eligible work, teams, volume, duration and exclusions. Specify support coverage, response times, fallback conditions and who may pause the route. Rehearse unreliable outputs, missing context, human override and escalation. Someone must receive and resolve the exception.

Use AI-Era Workforce Capability to assess demonstrated judgment for real tasks and retained responsibilities. Preserve its distinction between performance with support and performance for the agreed scope. A learning gap needs practice; missing permissions, unusable evidence or unavailable reviewers need operating changes. OECD's 2026 skills synthesis likewise emphasizes data interpretation, managerial and human skills alongside digital capability. [3]

Make the additional work explicit. Budget preparation, coaching, review, support and parallel processing by role, in hours and cost, with an end or review date. State which normal assignments will reduce. Temporary parallel work needs a defined purpose, such as comparing handoff quality; open-ended duplication can become the permanent operating model.

Discuss redistribution, progression and employment implications with the relevant people and HR decision owners. Identify available roles, selection criteria, support and unresolved decisions. Do not promise that everyone will automatically move to more rewarding work. Preserve opportunities to practice the judgment people remain responsible for.

Output: a bounded trial with demonstrated responsibilities, paid-for support and an executable fallback.

4. Make the routine workable every day

During transition, managers review incoming demand against actual user, reviewer and receiving-team capacity. They rebalance assignments, check exception aging and intervene before support queues become hidden overtime. They sample accepted and rejected work, ask people to explain consequential choices, and coach judgment against the information available at the time.

Review barriers with the Operating owner at an agreed cadence. Each entry needs an observed case, consequence, accountable owner, action, response date and evidence of resolution. Escalate funding or target conflicts to the Sponsor; route source, technology and control problems to their owners. A closed ticket is insufficient if the employee still cannot complete the work.

Reinforce useful contributions in assignments, recognition and performance conversations: sound checking, timely challenge, helpful handoffs and appropriate intervention. Explain what changed after feedback and why a suggestion was declined or deferred. Trust grows through visible responses and dependable boundaries.

Retire an old checklist, report or duplicate route only when its purpose is covered, affected owners accept the evidence, and remaining obligations are met. Remove the corresponding expectation and workload. Keep the agreed fallback usable. In continuing operation, retain these routines at a frequency suited to consequence and variation, with additional review after material changes.

Output: a supported operating routine, resolved barriers and an explicit decision about obsolete work.

5. Review the evidence and revise the commitment

Bring operating results and employee experience into the existing service review. Examine quality, acceptance, workload distribution, overrides, justified nonuse, unresolved objections and cost. Include ordinary demand peaks and absences; separate exceptional trial assistance from the support budget available after transition.

• Expand: extend a specified scope when outcomes, capacity and controls support it. Recheck the receiving teams' commitments.
• Continue: retain scope with explicit conditions, support and another review date.
• Narrow: remove unsuitable cases, actions or teams while keeping a viable bounded operation.
• Redesign: change the work, responsibilities, information or support and test again.
• Stop: withdraw the route when its case no longer holds and move affected work to an authorized alternative.

The mandated decision owner records the rationale, implementation owners, deadline and next review. For Redesign, state whether current work continues, narrows or pauses meanwhile. Confirm that changed instructions, access, assignments and support took effect. Expanding eligible work does not automatically expand AI authority.

Measure appropriate use against eligible work

Define a work unit, eligibility rule and version before measuring. Specify the teams or sites, measurement period, data sources, exclusions and exclusion counts. Include eligible work completed through another route; tool logs alone cannot supply that denominator.

Appropriate use in eligible work

Numerator: eligible work units that used the specified AI route within its authority and completed the required human checks.
Denominator: all work units meeting the agreed eligibility definition in the same scope and period.

Count a unit once, even if it involves several prompts or attempts. An appropriate human override after using AI can still qualify. Record justified nonuse, fallback and escalation separately, together with the reason; sound judgment should not become an employee failure. Do not silently remove these cases from the denominator. Identify inappropriate use, including use outside the eligible scope. Date changes to eligibility and explain their effect on comparisons.

Hypothetical arithmetic: in one four-week cohort, 100 cases are eligible. Eighty appropriately use the AI route, 12 take justified alternative routes, and eight have unresolved use barriers. Appropriate use is 80/100, or 80%. This is neither a target nor evidence that 80 outcomes were acceptable. The alternatives and barriers require explanation, and outcomes require separate follow-up.

Keep five kinds of evidence distinct:

1. Access and preparation: availability, awareness and training attendance show opportunity to participate.
2. Appropriate use: eligible-work handling shows whether the agreed route and responsibilities are practiced.
3. Accepted outcomes: owner acceptance, quality, failures, exceptions and later corrections show what the operation delivers. Track open cases and the follow-up window.
4. Total effort: include users, reviewers, receivers and support, plus repeated attempts, correction and recovery. Show whose burden changed.
5. Sustained operation: repeat the review under ordinary funded staffing, demand and support. A heavily assisted trial does not establish sustainability.

Cost per accepted outcome = total recurring cost of running the capability in a period ÷ outcomes accepted by the owner in that period. Include failed attempts, review, correction, support and recovery; keep one-time build investment separate. Match scope and period, and do not calculate the metric when no outcomes are accepted. Identify temporary parallel-work and support costs explicitly. Released capacity becomes a financial saving only when a corresponding expenditure is actually reduced; record useful redeployment separately.

Interpret low adoption before intervening

Inspect a sample of eligible work across roles and routes. If the task is unsuitable, revisit design or eligibility. If outputs are unreliable, address sources or the capability. If checks cannot fit within capacity, change staffing, demand or scope. If measures penalize the new contribution, change those measures. If restrictions are justified, retain them and reconsider the operating proposition.

Use targeted practice when evidence identifies a capability gap. Further communication helps when people lack a clear explanation or current instructions. Neither remedies an operation that imposes unacceptable workload or requires unjustified trust. Maximum usage is not the objective.

Two contrasting transitions

Both settings below are hypothetical. Their commitments, figures and review findings illustrate decisions, not client results or benchmarks. They are different operating choices, not stages toward greater autonomy.

A maintenance-planning assistant: retaining human judgment

Assume one manufacturer trials an assistant that assembles equipment history and recommends priorities for routine maintenance planning. Its arrangement is Recommend. A qualified planner retains scheduling and technical judgment; safety-critical decisions remain outside the assistant's remit.

Planners and technicians test representative cases. They identify a consequence: planners spend less time assembling records but more time checking conflicting history. Technicians may lose influence if their observations disappear from the summary. The proposed route therefore retains technician input and visible sources. The manager rewards a defensible plan rather than the number of recommendations accepted.

Assume a four-week trial budget of 16 planner hours, eight technician hours and eight support hours for additional checking and practice: 32 hours at an assumed loaded rate of $60, or $1,920. These are transition labor assumptions only; system charges and normal operating costs remain separately recorded. The Sponsor protects that capacity by reducing other assignments. Parallel comparison of plans ends or is expressly renewed at week four.

Suppose sampled cases show useful summaries for complete records but excessive correction for legacy histories. The Knowledge owner owns source repair. The authorized review chooses Narrow to complete-record cases, retaining manual planning elsewhere, with a new review in two weeks. Appropriate rejection of an AI priority counts as judgment, not disloyalty to the trial. Evidence must show accepted plans, total preparation and review effort, and subsequent corrections before any expansion.

A classroom-equipment service: bounded action across teams

Assume a university trials an operation that classifies routine equipment requests, creates work orders and reserves approved replacement equipment across support, stores and technicians. Act applies only within agreed catalog, access and capacity limits. Nonstandard commitments and safety or security exceptions Escalate; specified safety sign-off is Never Delegate. These are authority arrangements, not maturity levels.

Stores staff reveal that reservations arrive faster than equipment can be prepared. Technicians receive jobs without usable access details. Support's ticket-closure target encourages premature completion. The Sponsor secures a shared restoration measure and cross-team capacity; the Operating owner changes the handoff to require receiving acceptance and current availability. A designated service receiver confirms restoration before the Outcome owner counts an accepted result.

Assume a two-week, single-campus trial adds ten stores hours, six technician hours and four support hours per week: 40 hours at an assumed $50, or $2,000 in transition labor. A manual dispatch comparison runs for the first week, with reconciliation to prevent duplicate reservations. Managers review queue capacity daily. Unreliable availability or loss of required receiving coverage triggers a hold on automated reservations and the agreed human dispatch route.

If the trial exposes repeated unaccepted handoffs, the review chooses Redesign, holding affected actions while the Technology owner repairs availability checks and managers test the receiving routine. A busy academic week under normal funded coverage must then demonstrate accepted restorations, workable queues and supportable total effort. Higher reservation volume alone cannot justify expansion.

Keep one adoption decision record

Extend the operation's existing record. Use the following fields to connect a decision to what people must experience next; the Operating owner maintains it, while commitments and decisions retain their accountable owners.

Established foundations and limits

Prosci's ADKAR model addresses Awareness, Desire, Knowledge, Ability and Reinforcement in individual change; its wider methodology also addresses organizational change. Use that lens to investigate readiness without treating justified objection as a deficit in desire. [4] Kotter's approach contributes collective direction, removal of barriers and management practices that reinforce changed behavior. [5]

This framework applies established change principles to AI-enabled operations. Its specific demands include variable outputs, changing sources and model versions, delegated action, displaced review work and meaningful intervention. NIST's AI Risk Management Framework supports clear human–AI responsibilities and incorporating evaluated feedback into design and implementation. [6]

The five activities, diagram and decision record are RePerspective Labs' synthesis, not a renamed established model or an independently validated method. They complement service-specific design, technical evaluation, employment processes and governance. They cannot decide appropriate authority or workforce consequences without the people and mandates concerned.

Leadership takeaway

Leaders make the change credible through what they fund, stop asking for, reward and respond to. Ask managers and affected teams to show one real piece of work: how it now moves, who carries the checking and exceptions, and what happens when someone challenges the route.

If the operation works only through exceptional effort or suppressed objections, revisit the commitment. Sustained adoption means people can deliver accepted outcomes within supported boundaries—and use evidence to change those boundaries when necessary.

Where to go deeper

• Redesigning Work Around Outcomes — define and test the operation being changed.
• Governing AI-Enabled Operations — establish authorization, accountability and review.
• AI-Era Workforce Capability and Human-in-the-Loop Synergy — demonstrate role readiness and place human judgment.
• AI Adoption: When the Old Way Still Wins — examine incentives that preserve old routines.
• Leading the change and Sponsorship in practice — explore the connected operating questions.

Sources and further reading

Sources support the stated premises; they do not validate the integrated method. Examples and transition figures are hypothetical. Source versions checked October 11, 2026.

1. Dillon, E. W., Jaffe, S., Immorlica, N., and Stanton, C. T. Shifting Work Patterns with Generative AI, working paper, version 4, November 13, 2025. Randomized access experiment involving 7,137 knowledge workers across 66 firms; individual assistance, not autonomous operations or a test of incentives.
2. Lane, M., Williams, M., and Broecke, S. The impact of AI on the workplace: Main findings from the OECD AI surveys of employers and workers, March 27, 2023. Seven-country manufacturing and finance surveys; reported perceptions and associations, not causal evidence about consultation.
3. OECD. AI and skills: What we know so far, June 5, 2026. Synthesis of changing skill needs and training evidence; no universal workforce prescription.
4. Prosci. The Prosci ADKAR® Model, official explanation. Individual change model within the broader Prosci Methodology.
5. Kotter. The 8 Steps for Leading Change, official explanation. Collective change, barrier removal and reinforcement through management practices.
6. NIST. AI Risk Management Framework Core, AI RMF 1.0, January 2023, GOVERN 2, 3.2 and 5. Guidance on roles, human–AI oversight and feedback; not a certification of this framework.