How Do You Manage Change During Enterprise AI Transformation?

The owners, sizing method, gates and measures that most published research leaves out. A 2026 implementation guide for enterprise leaders.

By Sandeep Joshi, Founder and Managing Director of MASSIVUE, an enterprise AI adoption firm building structured learning infrastructure for APAC enterprises. Published by MASSIVUE. Last reviewed: August 2026.

The short answer

You manage change during enterprise AI transformation by treating it as a permanent operating capability rather than a project. AI has no go-live date, so episodic change management fails. Six things need standing owners: the work itself, the rules, the skills, the deal with employees, the evidence, and the operating cadence.

Executive summary

Enterprise AI transformation has produced a rare consensus. McKinsey, BCG, Deloitte, IBM, Microsoft, Gartner, Mercer, Protiviti, KPMG and Prosci now say broadly the same thing: the constraint is people and work design, not technology. BCG attributes roughly 70 percent of AI value to the people component. Prosci finds 63 percent of AI implementation challenges are human factors. Microsoft finds organisational factors carry more than twice the weight of individual ones.

That consensus is now commoditised. What almost nobody publishes is the layer below it: who owns what permanently, how to size the work, what the gates are between pilot and enterprise rollout, what questions to actually ask employees, where the saved time goes, and what happens to job descriptions.

This guide publishes as much of that layer as the evidence supports. It sets out a six-layer model, a measurement approach designed to resist gaming, a decision rule for the time AI frees up, a resistance diagnostic, a twelve-month roadmap with gate criteria, a worked sizing example, and a section on how Singapore’s institutional machinery changes the work for APAC enterprises.

Two cautions up front. Several of the most-quoted numbers in this field do not survive checking, and we name the ones we dropped. And where a recommendation is our design judgement rather than a research finding, we say so at the point of use.

Key takeaways

1. AI is the first enterprise change with no go-live date. Traditional change management is built around a cutover. AI capability changes continuously and employees decide task by task where it fits. A programme that closes cannot manage a change that does not end.
2. The workforce is already at change saturation. Gartner, writing in Harvard Business Review, reported that willingness to support enterprise change fell from 74 percent in 2016 to 43 percent in 2022. AI arrives on top of that, not instead of it.
3. Chief executives already know adoption is thin, and still believe the workforce is ready. IBM’s 2026 CEO Study found that CEOs themselves report only 25 percent of their workforce uses AI regularly, while 86 percent of the same CEOs believe employees have the skills to work with AI.
4. Usage dashboards create an incentive to perform usage. If adoption is measured by prompt volume, prompt volume is what you will get. Measure work outcomes and demonstrated proficiency, and treat telemetry as a diagnostic only.
5. The unanswered question drives much of the resistance. Gartner found only 7 percent of organisations give guidance on how to use time saved by AI, from a survey of 114 HR leaders. BCG found 66 percent of regular frontline AI users receive limited or no guidance on redirecting that time. Silence is read as a redundancy plan.
6. Training is outpacing redesign. Deloitte found education was the most common talent adjustment organisations made because of AI, at 53 percent, while career path redesign trailed at 33 percent. Protiviti found AI penetration of 57 percent in IT against 11 percent in HR.
7. Managers are the binding constraint. Gartner found 46 percent of managers experimenting with AI against 26 percent of employees, and just 14 percent of managers reporting that they face no challenges driving effective AI use across their team.
8. Singapore enterprises operate inside institutional machinery that global playbooks ignore. Job-redesign funding, tripartite consultation norms, sector skills standards and voluntary AI governance frameworks all change how the work is funded, sequenced and consulted.

What is AI change management?

AI change management is the discipline of redesigning work, rules, skills, incentives and measurement so that an enterprise actually uses artificial intelligence in daily operations. It is distinct from AI governance, which controls risk, and from AI deployment, which delivers systems. It answers a different question: will the work genuinely change, and will it stay changed.

The distinction matters commercially because the three are usually funded separately and sequenced badly. Deployment gets a budget and a date. Governance gets a policy and a committee. Change gets a communications plan and a training catalogue, and is expected to close when the programme closes.

That sequencing produces the pattern the research keeps finding. McKinsey’s State of AI, published November 2025, found 88 percent of organisations use AI in at least one business function, that around a third have begun to scale it, and that 39 percent of respondents attribute any level of earnings impact to AI, most of them under 5 percent. Gartner, in a survey of 782 infrastructure and operations leaders fielded in late 2025 and published in April 2026, found that only 28 percent of AI use cases in infrastructure and operations fully succeed and meet return expectations, while 20 percent fail outright.

A useful working definition, then: AI change management is the work of making sure the organisation’s operating reality catches up with its AI capability, and keeps catching up.

AI change management compared with adjacent disciplines

How is AI change management different from traditional change management?

Traditional change management is built around a cutover date. You prepare people, you launch, you embed, you close the programme. AI has no cutover. Capability changes continuously, the end state is unknown, and employees decide individually where AI fits into their own work. That inversion breaks the standard playbook in five specific ways.

1. There is no go-live date. An ERP replacement has a date when the old system is switched off. An AI rollout has a date when a licence is provisioned, which is not the same thing. Prosci puts the mechanism precisely: AI “requires individuals to determine where AI fits into their daily work.” Nobody can be pushed across a threshold that does not exist.

2. The future state is genuinely unknown. In a conventional transformation you can show people the target operating model. In an AI transformation, the model capability in eighteen months is not known to anyone, including the vendors. Change leaders are asked to build commitment to a destination they cannot describe. Pretending otherwise costs credibility quickly.

3. The change is continuous, and the workforce is already saturated. Gartner, writing in Harvard Business Review in May 2023, reported that the average employee experienced ten planned enterprise changes in 2022, up from two in 2016, and that willingness to support enterprise change fell from 74 percent to 43 percent over the same period. Note that this data is attributed to Gartner within an HBR article and no underlying study is separately published. Wiley’s 2025 survey of 1,685 North American workers found 67 percent expect still more change, and the study’s lead, Dr Tracey Carney, describes a “cascade crisis” of repeated disruption before recovery. AI change is not a fresh start. It lands on an already depleted base.

4. Resistance is quieter and harder to see. In a system cutover, non-adoption is visible because the old path is closed. With AI, non-adoption looks like normal work. Reporting by Aftermath in July 2026, based on interviews with workers who remained anonymous because they were still employed, described employees generating deliberately inflated prompt volume to satisfy adoption dashboards while changing nothing about their actual work, with one describing being designated an “AI Champion” as a result. This is journalism with anonymous sources rather than survey evidence, and should be read as an illustrative account rather than a measured prevalence. KPMG’s quarterly pulse separately reported employee resistance rising from 5 percent to 20 percent between consecutive quarters in 2026, with 53 percent of respondents attributing it to trust and ethical considerations rather than capability gaps. A four-fold move in a single quarter of a pulse survey should be treated as directional, not precise.

5. The commercial question arrives earlier. Boards ask about AI return within one budget cycle. BCG’s February 2026 analysis found only about 5 percent of organisations have reaped substantial financial gains from AI, defined as increases to revenue or cash flow along with significant process and workflow improvements. Change leaders therefore have to show progress on a horizon shorter than the change itself.

Side-by-side comparison

Does ADKAR or Kotter still work for AI change?

Yes, for the individual transition. Established change models remain the best available tools for diagnosing where a specific person has stalled and what to do about it. What none of them handles natively is the absence of a go-live date. Use an established model for the person, and add standing structural ownership for the organisation.

The practical answer for most enterprises is not to choose. Run ADKAR at the level of the person and the team, and run standing ownership at the level of the organisation. Neither substitutes for the other.

What actually goes wrong during enterprise AI transformation?

Four failures recur across the published evidence: leaders overestimate readiness; training substitutes for redesign; managers are skipped; and the question employees most want answered is left unanswered. None of these is a technology problem, and three of the four turn on decisions rather than budget.

Failure one: leaders overestimate readiness

IBM’s 2026 CEO Study, covering 2,000 chief executives across 33 geographies, found that CEOs themselves report only 25 percent of their workforce uses AI regularly as part of the job, while 86 percent of the same CEOs believe employees have the skills to collaborate with AI. Both figures come from the same population, so this is not a leader-versus-employee measurement gap. It is a single group holding two beliefs at once: adoption is thin, and the workforce is ready anyway. If both are true, the gap must be something other than skills, which is exactly what the rest of the evidence suggests.

Prosci’s study of 1,107 professionals found executives report higher trust in AI than frontline workers, at +1.09 against +0.33 on a scale running from minus two to plus two. Executive confidence tends to be calibrated on executive experience. Leaders who use AI daily in relatively unstructured work assume the rest of the organisation is having the same experience. Frontline workers in structured, audited, high-consequence roles are not.

Failure two: training substituting for redesign

Deloitte’s State of AI in the Enterprise 2026, based on 3,235 director-to-C-suite leaders across 24 countries surveyed in August and September 2025, found that education was the most common way companies adjusted their talent strategies because of AI, at 53 percent, followed by upskilling at 48 percent, talent acquisition at 36 percent and career path redesign at 33 percent. Reading that ranking plainly: the interventions that change what a job is came last. Deloitte separately found only 30 percent of organisations redesigning key processes.

Protiviti’s July 2026 pulse quantified the same imbalance from another angle. AI penetration was 57 percent in IT against 11 percent in HR. Only 13 percent of chief HR officers strongly agree their organisations are ready for role redesign, against 36 percent of IT leaders, and only 14 percent strongly agree on readiness for workforce learning, against 46 percent.

Training is easier to buy than redesign. It also produces a completion certificate, which resembles progress. It is not the same as changing what a job consists of.

Failure three: skipping the managers

Gartner’s March 2026 human resources research found 46 percent of managers experimenting with AI to improve their work against only 26 percent of employees, and that just 14 percent of managers said they do not face any challenges in driving effective use of AI across their team. Gartner’s own framing is that HR has focused on empowering employees to explore and has overlooked the role of the manager.

Microsoft’s 2026 Work Trend Index, based on 20,000 full-time employed or self-employed knowledge workers who use AI at work across ten markets, with fieldwork between 18 February and 7 April 2026, quantified what manager behaviour is worth. Where managers actively modelled AI use, employees reported a 17-point lift in reported AI value, a 22-point lift in critical thinking about their AI use, and a 30-point lift in trust in agentic AI. Microsoft also found only one in four AI users say their leadership is clearly and consistently aligned on AI.

Wiley found 52 percent of managers feel responsible for helping teams navigate change while fewer than half, at 45 percent, feel adequately supported by senior leadership. Managers are asked to carry the change and given the least support to do it.

What to do about it: a manager enablement checklist

This checklist is our design judgement, built from the evidence above rather than derived from a study. Work through it for every people manager in scope, in this order.

1. The manager can state the time-redeployment rule in their own words before they brief anyone. If they cannot, they will improvise, and what they improvise will be worse than the truth.
2. The manager completes assessed proficiency before their team does. Not a briefing, not a demo. The same assessment, on their own work.
3. A role-change conversation is scripted and scheduled for each direct report. Scheduled, with a date. An unscheduled conversation is one that does not happen.
4. The manager has a working escalation route for “this made the work worse”. If there is no route, the signal never reaches you, and you scale something that is not working.
5. Manager modelling is observed, not self-reported. Ask the team, not the manager. Microsoft’s 17, 22 and 30-point lifts are the payoff for getting this right.

How should leaders handle resistance to AI?

Diagnose the type before responding, because the four common types need opposite responses and treating all of them as a communications problem addresses none of them. Job fear needs a decision from the sponsor. Capability gaps need protected practice time. Quality objections often need listening to, because they may be correct. Values-based refusal needs an honest role-fit conversation.

The four-type diagnostic below is our own typology, offered as a working tool rather than an established taxonomy. It does not map one-to-one onto Prosci’s or anyone else’s resistance model.

The strong-performer case

The hardest version of this is a senior performer who refuses outright and whom you cannot afford to lose. The sequence that works is to separate the three possible objects of refusal before deciding anything. If the refusal is about this tool, that is a product problem and you may be able to solve it. If it is about this task, it is often a quality objection wearing a refusal costume, and it deserves testing. Only if it is about the principle is it a role-fit question, and then it should be handled as one, with process and with time, not as a disciplinary matter.

Making it a performance conversation first is the common error. It converts a solvable product or quality issue into an employment dispute, and it teaches everyone watching that quality objections are career-limiting. That is the fastest way to lose the signal you most need.

The AI Change Stack: six layers that need standing owners

Because AI change has no end date, the useful question is not “what are the steps” but “what needs a permanent owner.” Six layers do. Each has an accountable executive role, a measure, a failure mode, and an evidence base. Together they form what MASSIVUE calls the AI Change Stack.

This is not a sequence. All six run continuously and interact. A gap in any one layer will stall the others, which is why organisations that invest heavily in a single layer, usually skills, see so little movement.

Layer 1: Work

What changes. The task composition of specific roles. Not the tools, the tasks.

Who owns it. The line business owner for the affected function, supported by process owners. This cannot be delegated to a programme office or to IT.

What the enterprise must do. Decompose the target roles to task level. For each task, make one of four calls: automate, augment, eliminate, or leave alone. Write the resulting task map down. McKinsey describes the same discipline as mapping what a key role’s week looks like today and what it should look like tomorrow, “not at the level of strategy, but at the level of decisions.”

How progress is measured. Percentage of in-scope roles with a signed-off task map, and percentage of those maps reflected in the actual job description.

What can go wrong. Buying tools and hoping workflow follows. Deloitte found only 30 percent of organisations redesigning key processes. McKinsey’s State of AI found high performers nearly three times as likely as others to fundamentally redesign workflows.

Evidence. BCG attributes roughly 70 percent of AI value to the people component, against 10 percent for the algorithms themselves and 20 percent for the technology required to implement them. BCG presents this as a pattern observed across its case work involving hundreds of companies, not as a study finding with a stated sample, and it should be weighed accordingly.

Layer 2: Rules

What changes. What people are permitted to do with AI, at what level of autonomy, with what disclosure and what escalation path.

Who owns it. The chief risk officer or general counsel, working with the AI governance body. In financial services this is usually already a board-level accountability.

What the enterprise must do. Publish a usable acceptable-use position that a non-specialist can apply in under a minute. Define autonomy tiers, for example: suggest only, human approves each action, human approves in batch, autonomous within limits. Say what must be disclosed to customers and colleagues. Say what happens when someone gets it wrong in good faith, because if that is unclear, people will stop telling you.

How progress is measured. Percentage of known AI use cases mapped to a defined autonomy tier. Visibility of unsanctioned tool use, which is a health indicator rather than a disciplinary one.

What can go wrong. A policy vacuum, then shadow AI. Wiley found 33 percent of employees were not sure whether their company even had AI policies. Deloitte found only 21 percent of organisations had mature governance for AI agents.

Is shadow AI a discipline problem?

No. In almost all cases it is a product problem. Gartner’s survey of 12,004 employees and managers across 40 countries found that 88 percent of employees with enterprise AI access also use personal AI tools for business tasks, often to save time. Access alone does not displace shadow use. Only useful rules and genuinely better sanctioned tools do.

Before treating it as a violation, diagnose why the sanctioned path is losing.

Disciplinary responses without this diagnosis do one reliable thing: they move shadow use further out of sight.

Layer 3: Skills

What changes. Demonstrated capability on real work, differentiated by role, refreshed continuously.

Who owns it. The CHRO for the system, the line leader for the outcome. If learning and development owns the outcome, the outcome becomes course completions.

What the enterprise must do. Build role-differentiated pathways rather than one general literacy course. Protect practice time, in the diary, on work that matters. Assess on a real work artefact, not a quiz.

How progress is measured. Proficiency demonstrated on a role-relevant task, assessed by someone qualified to judge the work. Course completion is an input, not an outcome.

What can go wrong. A single generic curriculum that plateaus. This is the one layer where MASSIVUE has published a first-party position rather than a reading of someone else’s research. In our department-by-department analysis of AI training, we set out that after three years of building AI adoption programmes across more than 40 APAC enterprises, a single curriculum serving every function is the reason we see AI initiatives plateau at 15 to 20 percent utilisation, measured as Weekly Active Utilisation: the share of licensed users who actively use the tool in a given week. That is our own observation across our client base, not a survey finding, and it should be weighed as such.

Evidence. BCG’s 2025 AI at Work study found regular usage is sharply higher for employees who receive at least five hours of training and have access to in-person training and coaching. The two conditions appear together in BCG’s finding and should not be separated: hours alone is not the reported driver. BCG’s 2026 edition found 72 percent of employees say skill expectations have shifted while only 36 percent feel they have received adequate upskilling.

Layer 4: The Deal

What changes. The explicit understanding between the organisation and its people about what AI is for, what happens to saved time, and what happens to roles.

Who owns it. The chief executive, with the CHRO. It cannot be delegated, because the only reason anyone believes it is that the person who could change it said it.

What the enterprise must do. Publish a time-redeployment rule before the rollout, not after. Say what is decided and what is genuinely not yet decided. Prosci’s guidance here is sound: name the fear rather than ignore it, and position employees as participants in shaping how AI is used rather than recipients of a decision.

How progress is measured. A published rule that a randomly selected employee can state accurately. Trust measured directly and trended, not assumed.

What can go wrong. Silence, then quiet non-compliance. Prosci found 29 percent of employees worry about job displacement or role ambiguity during AI implementations.

Evidence. BCG found 66 percent of regular frontline AI users receive limited or no guidance on redirecting saved time. Gartner, from a survey of 114 HR leaders, found just 7 percent of organisations provide guidelines on how to use time saved by AI. BCG also found only 28 percent of employees see a strong connection between what leaders say and what the organisation actually does. Mercer’s Global Talent Trends 2026 measures a related but distinct construct: 62 percent of employees feel leaders underestimate AI’s emotional and psychological impact, while only 19 percent of HR leaders address these impacts in their digital strategy. Mercer is not a third measurement of the guidance gap; it is adjacent evidence that the silence has a cost.

A worked example: what a chief executive can say before rollout

The article’s central claim is that the CEO must name the time decision. Here is what that can sound like. Adapt the specifics; keep the structure.

“We are introducing AI across the claims and servicing teams starting in October. Here is what I can tell you, and here is what I cannot.

What is decided. The tasks this changes first are triage, first-draft correspondence and document summarisation. It does not change adjudication or any decision that affects a customer’s outcome, and it will not, without me telling you first.

What happens to the time. We have a backlog of work we have not been able to serve for two years. The time this frees goes there first. I am not planning a headcount reduction on the back of this programme, and if that position ever changes, you will hear it from me before you hear it anywhere else.

What is not decided. We do not yet know how far this goes or how quickly. Anyone who tells you they know the eighteen-month picture is guessing.

What I am asking. Use it, and tell us when it makes the work worse. There is a route for that and nobody’s standing suffers for using it. The fastest way to make this go badly is for people to stay quiet about what is not working.”

What not to say. Avoid “no one will lose their job because of this” unless you can guarantee it, because you will be quoted on it. Avoid “AI will free you up for higher-value work” without naming the work, because it is heard as a euphemism. Avoid announcing a decision that has not actually been made in order to end the conversation, because it will be tested within a quarter.

This is a template built from the structure above, not a transcript. Adapt every specific to your own position, and do not say any part of it that is not true in your organisation.

Layer 5: Evidence

What changes. What counts as proof that the transformation is working.

Who owns it. The transformation office jointly with finance. Finance matters because an unaudited benefit claim will not survive a board cycle.

What the enterprise must do. Baseline before you deploy. Measure work outcomes at the level of the process, demonstrated proficiency at the level of the person, and usage telemetry only as a diagnostic. The detail is in the next section.

How progress is measured. Whether a specific, pre-registered outcome metric moved, and whether the movement survives scrutiny by someone who did not run the programme.

What can go wrong. Measuring adoption instead of value, and thereby creating an incentive to perform adoption. BCG’s instruction to chief executives is direct: “Change the scoreboard: measure value, not adoption.” Gartner, in a May 2026 release, makes a similar point, warning that most leaders are mistaking basic access or adoption metrics for transformation.

Evidence. Gartner’s finding that 28 percent of AI use cases in infrastructure and operations fully succeed against return expectations indicates the measurement bar is real and that most of those use cases are not clearing it. Note the scope: this is an infrastructure and operations population, not a whole-enterprise one.

Layer 6: Cadence

What changes. Who keeps doing all of the above after the programme ends.

Who owns it. A named standing role, not a temporary programme director.

What the enterprise must do. Fix the intervals and the forums. A practical minimum, and this is our design judgement rather than an evidence-derived schedule: monthly review of the task maps against current model capability; quarterly review of the time-redeployment rule and the trust measure; semi-annual review of the autonomy tiers and the skills pathways. Put each one into an existing forum that already has authority, rather than creating a new committee that can be ignored.

How progress is measured. The honest test is whether the loop still ran in quarter five, after the launch energy has gone.

What can go wrong. The programme closes, the change reverts, and eighteen months later a new programme is chartered to fix the same problem. In our view this is the most common cause of reversion, though we know of no study that has tested predictors of AI-change reversion specifically.

Evidence. Gartner’s March 2026 change management research found that organisations which continuously or regularly adapt change plans based on employee responses are four times more likely to achieve change success.

The stack at a glance

How do you measure AI adoption without creating an incentive to fake it?

Measure three things separately and never mix them: work outcomes owned by the line, demonstrated proficiency assessed on real artefacts, and usage telemetry used only as a diagnostic. Set targets on outcomes, set thresholds on proficiency, and set no target on telemetry. The moment a usage number becomes a target, it stops being evidence.

This is an application of a well-established principle, that a measure which becomes a target ceases to be a good measure. The AI-specific illustration comes from Aftermath’s July 2026 reporting, in which workers described producing inflated prompt volume to satisfy adoption dashboards. That is anecdotal rather than measured, and we present it as an illustration of a mechanism rather than as evidence of how widespread the behaviour is. The design principle stands on its own regardless.

The three-tier measurement model

Tier 1: Work outcomes. Owned by the line business owner. These are the numbers that were already on the operating scorecard before AI arrived: cycle time, error rate, throughput per person, cost to serve, first-contact resolution, days to close. Choose two or three per process, baseline them before deployment, and pre-register the expected direction and rough magnitude. Pre-registration matters because it prevents retrospective selection of whichever metric happened to move.

Tier 2: Demonstrated proficiency. Owned by the CHRO with the line. Not course completion and not a self-rating. A person is proficient when they can complete a task from their actual role to the required standard using AI, assessed by someone competent to judge the output. This is where a training investment is proved or disproved.

Tier 3: Usage telemetry. Owned by the platform team and treated as a diagnostic. Licence utilisation, weekly active use, feature depth. Useful for answering “where is nothing happening at all” and “which teams may need help”. Poor as a target and actively harmful as a performance measure.

Six design rules that make the measure harder to game

These are design recommendations, not research findings.

1. Do not set a usage target. If a usage number is needed for a board pack, present it as context alongside an outcome number, never as the headline.
2. Separate the person who reports the number from the person whose performance it reflects. Self-reported adoption is not evidence.
3. Baseline before deployment. A post-hoc baseline is an estimate written by someone who wants the number to look good.
4. Pre-register the metric and the expected direction. Write it down before you start.
5. Sample real work. Periodically pull actual outputs and review quality. Telemetry cannot tell you whether the work got better.
6. Ask about non-use without penalty. Create a route for a team to say “we tried this and it made the work worse” that does not read as failure. Suppressing that signal is how organisations scale something that does not work.

What to put in the board pack at 90 days

Ninety days is too early for a defensible profit-and-loss claim and too late to say nothing. The answer is to report movement in a pre-registered operational metric, and to be explicit about what you are not yet claiming. Under our own gate design, no financial benefit claim should be made before Gate 3.

A usable formulation for the pack: “One pre-registered outcome metric, claims cycle time, moved from X to Y in the pilot population over the quarter. We are not claiming enterprise financial impact, and will not until the gain has held for two quarters and finance has reviewed it.”

The MASSIVUE AI Change Readiness Check: 13 items you can field this week

The MASSIVUE AI Change Readiness Check is a 13-item structured practitioner assessment that measures whether the five conditions people need in order to adopt AI are actually present: clarity about the change, permission to act, capability to do the work, manager support, and trust in what leadership has said. It is designed to be fielded in-house, before deployment and at intervals afterwards, without a consultant or a licence.

What it is, and what it is not. This is a structured practitioner assessment, published by MASSIVUE. It has not been tested for construct validity, reliability or factor structure, and it is not a validated psychometric instrument. Do not use it for individual assessment, performance management or selection. Use it as a diagnostic conversation with numbers attached.

Why it exists. Microsoft names an Organization versus Employee AI Readiness Index, Gartner names a True ROI Index, and Prosci and Protiviti each publish diagnostic constructs. As far as we have been able to establish, none of them publishes a scoreable item set a practitioner can field themselves. That is the gap this fills.

How to use it. Field it anonymously, by team, before deployment and then at six-month intervals. Use a five-point agreement scale, from strongly disagree to strongly agree. Report the distribution rather than only the mean, because the shape of disagreement carries the signal: a team splitting 50-50 on trust is a different problem from a team clustered in the middle, and both average out the same way.

Clarity

1. I can explain in one sentence why my organisation is adopting AI.
2. I know which of my specific tasks AI is intended to change.
3. I know what my organisation intends to do with the time AI saves.

Permission

4. I know what I am allowed to use AI for in my role.
5. I know what I must disclose when I use AI.
6. If I made a good-faith mistake using AI, I would report it.

Capability

7. I have had enough hands-on practice to use AI on my real work.
8. Someone I can ask for help is available when I get stuck.
9. I have protected time to build these skills.

Manager

10. My manager uses AI in their own work in a way I can see.
11. My manager has discussed with me how my role will change.

Trust

12. I believe what leadership has said about AI and jobs here.
13. I would tell a colleague honestly what I think about this programme.

How to read the result. Each of the five groups tells you which layer of the AI Change Stack is failing, which is what makes the instrument actionable rather than merely descriptive.

Item 13 is the integrity check. Where scores are high across items 1 to 12 and low on item 13, treat the other twelve answers as unreliable rather than reassuring. A population that will not speak candidly to a colleague has not spoken candidly to you either.

One caution on repeat fielding. Do not attach a target to any item, and do not report results by named team to that team’s leadership as a performance measure. Both convert the instrument into the same problem as a usage dashboard, and the same behaviour follows.

MASSIVUE publishes this instrument openly. You are welcome to use, adapt and field it inside your own organisation with attribution.

What happens to the time AI saves?

Decide it before you deploy, publish the decision, and apply it consistently. Four destinations cover most cases: reinvest in unmet demand, reinvest in quality and depth, return it to the employee as capacity, or extract it as cost. Which is right depends on whether your binding constraint is demand or cost. What is never right is refusing to say.

This is the largest gap we found in the published research. BCG found 66 percent of regular frontline AI users receive limited or no guidance. Gartner found just 7 percent of organisations publish guidelines, and found HR leaders and managers disagreeing about the answer: 55 percent of HR leaders would like a freed hour spent on special projects outside the core job, against 28 percent of managers who would prioritise that. Both organisations measure the gap. Neither publishes a mechanism.

The four destinations, and when each is right

Three practical points. First, an organisation can legitimately choose different destinations for different functions, but it must say so and say why. Inconsistency that is explained is survivable. Inconsistency that is discovered is not.

Second, these four are the common cases rather than a logically exhaustive set. Redistributing freed capacity unevenly across a team to relieve a specific bottleneck, for instance, sits across two of them.

Third, if the honest answer is the fourth row, say it early. Employees are already assuming it. Prosci found 29 percent worrying about job displacement or role ambiguity. Mercer found AI job-loss anxiety rising from 28 percent in 2024 to 40 percent in 2026. In Singapore, ADP Research’s People at Work 2026, based on fieldwork in July and August 2025 with more than 39,000 working adults across 36 markets, found only 15 percent of Singapore workers strongly agree their job is safe from elimination, against 22 percent globally. Note that ADP’s question covers job elimination generally, framed against advances in AI and changing demographics, rather than AI displacement alone. A workforce that already fears the answer is not protected by not being told.

What does AI change management cost, and who do you need?

The only published budget ratio we could locate is McKinsey’s, from its August 2026 guidance on closing the agentic adoption gap: a 1:3:5 split across technology, process redesign and capability building, with the observation that most organisations invert it. Treat it as the best available anchor rather than a benchmark, because McKinsey publishes no derivation for it.

The honest position is that credible, independently derived cost and staffing benchmarks for AI change work do not currently exist in public. Rather than invent them, here is a method for deriving your own, followed by an illustrative worked example.

Sizing method

Step 1. Count the affected population, not the licensed population. People whose task composition changes, not people who received an account. These are different numbers and the second is usually much larger.

Step 2. Segment by change depth. Three tiers is normally enough:

• Light: AI assists existing tasks and the job is recognisably the same.
• Moderate: significant tasks are automated or restructured, and the job description needs rewriting.
• Deep: the role’s purpose changes, or the role merges with or splits from another.

Step 3. Assign support intensity by tier. Light populations need clear rules, good self-serve enablement, and a briefed manager. Moderate populations need role-level task mapping, assessed proficiency and a rewritten job description. Deep populations need individual consultation, a career path conversation and, in many jurisdictions, formal process.

Step 4. Size the coaching layer against the deep and moderate populations only. The only published staffing data point we found is McKinsey’s description of a Fortune 500 technology company placing AI coaches into engineering teams at roughly one coach per five to ten employees, joining sprint ceremonies and retrospectives. That is a single example in a single function and should not be generalised to a whole enterprise. It is a reasonable starting ratio for a deep-change population and clearly too rich for a light one.

Step 5. Budget the standing cadence separately from the programme. This is the line item organisations forget. The programme has an end date. Layer six does not.

A worked sizing example

The example below is illustrative arithmetic derived from the method above. It is not a benchmark, and the segmentation percentages are assumptions you should replace with your own analysis.

Assume a 5,000-person enterprise, of which 1,200 are materially affected in year one.

Coaches are normally internal super-users on a part-time secondment rather than new hires, so the figure above is a workload estimate, not a headcount request. At, say, 40 percent time allocation, 33.8 coach FTE is roughly 85 people contributing part of their week.

Then the standing roles:

The point of the second table is not the numbers. It is the right-hand column. AI change budgets are frequently built as if every line ended at go-live, and the recurring cost of the Cadence layer is the one that tends not to reach the business case at all.

To be explicit about what is and is not evidence here: the 1:3:5 ratio and the one-coach-per-five-to-ten figure are McKinsey’s, cited above. The segmentation percentages, the coach ratios applied to the moderate and light tiers, and every FTE figure in the second table are our assumptions, included so the method produces something you can argue with. Replace them with your own population analysis before taking any of it to a budget conversation.

Who you need

On the sponsor role specifically, Prosci’s research found leadership commitment showed the largest gap of any factor evaluated between organisations that were struggling with AI adoption and those that were succeeding, at a difference of 1.65 on a scale from minus two to plus two. On transparency and clarity around AI strategy, Prosci reports high performers at +1.29 against low performers at minus 0.54.

A twelve-month roadmap with gate criteria

Run four phases with explicit gates between them. The gates matter more than the phases, because a common failure is scaling something that has not been proved. Each gate is a decision to proceed, hold or stop, made by the executive sponsor on evidence rather than on schedule pressure. The specific thresholds below are our design judgement, not empirically derived, and should be calibrated to your risk appetite.

Phase 1, days 0 to 30: Baseline and decide

Gate 1. Do not proceed unless: the time-redeployment rule is decided and published; outcome metrics are baselined and pre-registered (that is, written down before deployment with an expected direction); and the executive sponsor has personally communicated both. If the sponsor will not make the redeployment decision, the programme is not ready, and starting anyway converts an unmade decision into a rumour.

Phase 2, days 31 to 90: Redesign and pilot

Gate 2. Do not scale unless: at least one pre-registered outcome metric (a metric baselined and declared before deployment) has moved in the expected direction; assessed proficiency has been demonstrated by a majority of the pilot population; and the pilot managers can describe the change in their own words. Movement in usage telemetry alone does not clear this gate.

Phase 3, days 91 to 180: Scale the pattern

Gate 3. Do not proceed to enterprise rollout unless: job descriptions have caught up with the actual work; the readiness survey has improved on the Clarity and Trust item groups specifically, since those are the two that measure whether the Deal has landed; and the outcome gains from Phase 2 have held for at least one further quarter rather than decaying. This is also the earliest point at which a financial benefit claim should be made.

Phase 4, days 181 to 365: Install the cadence

Gate 4, the real one. Twelve months in, ask whether the loops would still run if the change lead left tomorrow. If the answer is no, the capability has not been installed and what exists is a programme wearing the costume of an operating model.

On timing generally: in our experience a narrow, well-chosen scope can show movement in a pre-registered operational metric within 90 days, and enterprise coverage is usually a twelve-month conversation rather than a quarterly one. We are not aware of published data establishing a typical duration, so treat these as planning assumptions rather than benchmarks.

How does AI change management differ in Singapore and APAC?

The people work is broadly the same. What differs is the institutional context. Singapore enterprises operate inside tripartite consultation norms, publicly funded job-redesign instruments, sectoral skills standards and voluntary AI governance frameworks that global playbooks do not account for. That context changes how change is funded, sequenced and consulted, and it creates obligations a US-designed programme plan will simply miss.

The adoption profile is unusual

Singapore’s enterprise adoption is high by global standards and highly uneven internally. IMDA’s Singapore Digital Economy Report 2025, covering reference year 2024, reported AI adoption among non-SMEs rising from 44 percent to 62.5 percent, with SMEs tripling from 4.2 percent to 14.5 percent. At the same time, a Ministry of Manpower survey published in April 2026 found that 71.5 percent of firms had not begun adopting AI in their operations, that 18.9 percent were redesigning roles, and that 13.9 percent were creating new AI-related jobs.

Those figures are not contradictory. They describe a market where large enterprises have moved substantially and the broad base of firms has not. They also show adoption running well ahead of role redesign, which is the same imbalance the global research identifies, visible here in national statistics.

Confidence is below the global average even where adoption is above it

ADP Research’s People at Work 2026 found only 15 percent of Singapore workers strongly agree their job is safe from elimination, against a global figure of 22 percent. The fieldwork was conducted in July and August 2025, so treat it as a lagging indicator.

This produces a tension worth naming, because it shapes what change leaders should prioritise. Singapore sits above the global average on enterprise AI adoption and below it on job-security confidence. In that combination, Layer 4 of the stack, the Deal, is not a communications nicety. It is the binding constraint.

How do you fund AI job redesign in Singapore?

The main instrument is the SkillsFuture Workforce Development Grant for job redesign, administered by Workforce Singapore. It supports up to 70 percent of qualifying costs for SMEs and up to 50 percent for non-SMEs, capped at S$150,000 per enterprise. Note carefully that this is support against qualifying costs, not a cash grant of S$150,000.

The practical sequencing point is the one most organisations get backwards: check eligibility before you finalise the redesign scope, not after. The scope that qualifies for support is often broader than the scope an organisation would self-fund, which means scoping first and checking funding second leaves money on the table and produces a narrower redesign than the business actually needs.

Confirm current eligibility criteria, qualifying-cost definitions and application steps directly with Workforce Singapore before relying on them, because programme parameters change.

What the proposed MAS AI risk management guidelines could mean for change and capability work

Status first, because it matters. The Monetary Authority of Singapore issued a Consultation Paper on Proposed Guidelines on Artificial Intelligence Risk Management for Financial Institutions on 13 November 2025. The consultation closed on 31 January 2026, and the paper proposes a 12-month transition period following issuance of any final guidelines. These are proposed guidelines at consultation stage, not regulation in force. At the last review of this article in August 2026, we could not locate a MAS response to feedback or a final issued text. That is an absence of evidence rather than proof that none exists, and MAS’s own publications page is the only authoritative place to check.

Because the paper itself is not accessible to automated retrieval, the summary below is drawn from professional advisers’ published readings of the consultation rather than from the MAS text directly. Treat it as a planning prompt, not as a statement of obligation, and read the primary document before acting on any of it.

If the guidelines issue broadly in the shape the consultation describes, the consequences for the people side of an AI programme are direct. Reported requirement areas map onto change and capability work as follows.

Scope, as reported, is all financial institutions, with implementation commensurate with the size and nature of each institution’s activities.

What to verify before you rely on any of this. Open the MAS consultation page directly and check four things: whether final guidelines have been issued; if so, their issue date and the transition end date; whether the five requirement areas above survived consultation in this form; and the exact wording of the competence and change management provisions, since those are the two that determine what your capability evidence has to look like. The document is the Consultation Paper on Proposed Guidelines on Artificial Intelligence Risk Management for Financial Institutions, at mas.gov.sg, with the accompanying media release at mas.gov.sg/news.

For financial institutions specifically, the NTUC and IBF commitment announced on 25 June 2026 to equip up to 100,000 finance professionals in Singapore with AI skills over three years is a useful capability-planning anchor. Note that “up to 100,000” is a stated aspiration and a ceiling rather than a committed number.

The wider institutional machinery

What this means practically

For an APAC enterprise handed a globally designed AI change playbook, three things genuinely have to change rather than simply being translated.

• Bring the consultation conversation forward. In a tripartite environment, consultation is an input to the design, not a communications output at the end of it.
• Check funding eligibility before designing the job-redesign scope. The scope that qualifies is often broader than the one organisations self-fund.
• Align capability building to recognised sector standards where they exist. That is what makes the investment portable for the employee, which in turn is what makes the Deal credible.

Common mistakes

1. Running it as a programme. A programme has an end date. The change does not. Nobody owning the loops after the programme closes is, in our view, the most common route to reversion.

2. Buying training instead of redesigning work. Deloitte’s finding that education was the most common adjustment while career path redesign trailed describes a widespread pattern. Training a person to use a tool inside an unchanged process produces a faster version of the same output.

3. Setting a usage target. It converts your measurement system into a compliance exercise, and once that starts your data is worth very little.

4. Skipping the managers. With 46 percent of managers experimenting against 26 percent of employees, and just 14 percent reporting no challenges in driving effective use across their team, the manager layer is simultaneously ahead on usage and short on support to lead it. Those are two different problems and both need addressing.

5. Refusing to answer the jobs question. Silence is not neutral. It is interpreted, and it is interpreted badly.

6. Treating shadow AI as a discipline issue. With 88 percent of employees who have enterprise access also using personal tools for business tasks, shadow use is a signal that the sanctioned path is worse. Fix the path.

7. Launching AI change on top of unrecovered change. Check what else the affected teams are absorbing. The saturation evidence and Wiley’s “cascade crisis” framing both point the same way: capacity to absorb change is finite and currently stretched.

8. Designing only for the enthusiasts. The population that determines whether this works is not the people already using AI daily. It is the substantial group who will comply if required and are not persuaded. Design for them.

9. Quoting statistics that do not survive checking. Several of the most-repeated figures in this field trace to preliminary work with convenience samples, or to secondary reporting with no published study behind it. Using them in a board paper is a credibility risk. We name ours in the next section.

Limitations of this guide

Stating what this guide cannot tell you is part of making it useful.

• No independently derived cost or staffing benchmarks exist publicly. The only published ratio we located, McKinsey’s 1:3:5, has no published derivation. The only published coaching ratio is one example from one function at one company. The worked sizing example in this guide is illustrative arithmetic built on assumed segmentation, not a benchmark.
• The MASSIVUE AI Change Readiness Check is a structured practitioner assessment, not a validated psychometric scale. It has not been tested for construct validity, reliability or factor structure. Use it as a diagnostic conversation with numbers attached, not for individual assessment, and improve it with your own data.
• Several original elements are our design judgement, not research findings. Specifically: the six-layer stack, the four-type resistance diagnostic, the three-tier measurement model and its six design rules, the four time-redeployment destinations, the gate thresholds, the manager checklist and the cadence intervals. Each is labelled as such at the point of use. They are reasoned from the evidence cited, but none has been empirically tested.
• Two widely quoted statistics were excluded after checking, and we name them. First, the frequently repeated claim that roughly 95 percent of enterprise AI pilots fail to deliver measurable profit-and-loss impact. It originates in a July 2025 preliminary, non-peer-reviewed report from a group affiliated with the MIT Media Lab, drawing on 153 survey responses recruited at industry conferences, and its derivation for that specific figure has not been shown. Second, a widely circulated pair of Prosci sponsorship scores rendered as “+1.65 versus minus 1.50”. The 1.65 is the size of the gap between struggling and succeeding organisations, not a score for either group, and we could not locate a minus 1.50 figure in Prosci’s published material. Where we do cite Prosci’s paired scores, we use the pair Prosci actually publishes, +1.29 against minus 0.54 on transparency and clarity.
• Several sources are lagging. ADP’s 2026 report uses mid-2025 fieldwork. Deloitte’s 2026 State of AI uses August and September 2025 fieldwork. Gartner’s March 2026 release draws on July 2025 surveys. AI sentiment moves faster than survey publication cycles.
• Small sub-samples are flagged at first use. Gartner’s 7 percent time-savings figure comes from a survey of 114 HR leaders, which is a thin base for a headline number.
• Scope qualifiers matter. Gartner’s 28 percent success figure applies to AI use cases in infrastructure and operations, not to all enterprise AI. BCG’s 42 percent and 66 percent apply to regular frontline AI users, not to all employees. IBM’s 86 percent and 25 percent both come from chief executives, not from employees.
• The Singapore regulatory position is live, and one section rests on second-hand sources. The MAS guidelines discussed here are proposed guidelines at consultation stage, not regulation in force. We could not locate a final issued text at the last review of this article, and MAS’s own consultation paper is not accessible to automated retrieval, so that section is drawn from professional advisers’ published readings rather than from the MAS text directly. Read the primary document before acting on any of it.
• This guide is written by a firm that sells AI transformation consulting and capability building. That is a commercial interest and you should read it accordingly. We have tried to make it useful whether or not you engage anyone, and to cite sources you can check independently.

Frequently asked questions

What is AI change management in one sentence? It is the discipline of redesigning work, rules, skills, incentives and measurement so an enterprise actually uses AI in daily operations and keeps using it as the technology changes.

How is AI change management different from traditional change management? Traditional change management is organised around a cutover date. AI has no cutover. The future state is unknown, capability changes continuously, and employees decide task by task where AI fits. That means standing ownership rather than a programme with a close date.

Why do most AI transformations stall? Because the technology is delivered and the work is not redesigned. Deloitte found education was the most common organisational response, at 53 percent, while career path redesign trailed at 33 percent. A tool inside an unchanged process produces a faster version of the same output.

Who should own AI change management? The CHRO owns the system and the line business owner owns the outcome. The CEO owns one thing that cannot be delegated: the statement of what AI is for and what happens to the time it saves.

How do you prepare employees for AI? Tell them which of their specific tasks will change, tell them what happens to the time saved, give them role-relevant practice on real work rather than a generic course, protect the practice time in the diary, and make sure their manager visibly uses AI too.

How do you measure AI adoption? On three separate tiers: work outcomes owned by the line and baselined before deployment, demonstrated proficiency assessed on real work artefacts, and usage telemetry treated only as a diagnostic. Do not set a target on the third tier.

Can employees game AI adoption metrics? Yes. Reporting in 2026 described workers producing inflated prompt volume to satisfy adoption dashboards while changing nothing about how they work. That account is anecdotal rather than measured, but the mechanism is well understood: any measure that rewards volume of AI activity will be met with volume of AI activity.

How long does AI change management take? In our experience a narrow, well-chosen scope can move a pre-registered operational metric within 90 days, and enterprise coverage is usually a twelve-month conversation. We are not aware of published data establishing a typical duration, so treat those as planning assumptions. The standing cadence does not end, and that is the design rather than a planning failure.

What is a realistic AI adoption rate to expect? There is no defensible universal benchmark and anyone offering one is guessing. The useful question is not what percentage of people used the tool, but whether a pre-registered work outcome moved. Set your own baseline and compare against yourself.

How should leaders handle resistance to AI? Diagnose the type first. Job fear needs a decision from the sponsor. Capability gaps need protected practice time and in-the-moment help. Quality objections need testing, because they may be correct. Values-based refusal needs an honest role-fit conversation. Treating all four as a communications problem addresses none of them.

What should we do when two strong performers refuse to use AI outright? Establish first whether the refusal is about the tool, the task, or the principle. A quality objection from a strong performer is information about your deployment, not an obstacle. Only a refusal on principle is a genuine role-fit conversation, and it should be handled with proper process rather than as a compliance matter. Making it a performance conversation first converts a solvable problem into an employment dispute.

What skills do employees actually need during AI transformation? Role-specific ones. The general capabilities that transfer are judgement about when not to use AI, the ability to verify output in your own domain, and clear task decomposition. Everything else is specific to what the person actually does.

What is the difference between AI adoption and AI transformation? Adoption means people use the tool. Transformation means the work itself is different, the job description reflects it, and an operating metric moved. You can have high adoption and no transformation, which is a common and expensive outcome.

Does our AI change programme need a Chief AI Officer? Not necessarily, and the role does not resolve the change question by itself. IBM found 76 percent of organisations reported having a Chief AI Officer in 2026, up from 26 percent a year earlier. What matters is whether someone with authority owns the decision on jobs and time, which is usually a CEO and CHRO responsibility rather than a technology one.

How do we manage AI change when our teams are already exhausted by change? Check the concurrent change load per team before you schedule anything, sequence rather than parallelise, and be prepared to say “not yet” for specific teams. Willingness to support enterprise change has fallen sharply over the last decade and absorption capacity is finite.

What do we do if our AI mandate has already backfired? Stop measuring usage immediately, because that is a likely source of the damage. Acknowledge specifically what went wrong rather than relaunching under a new name. Make the jobs and time decision that was probably never made. Then rebuild with a narrow scope where you can show a real outcome, and let the evidence do the persuading.

What should an AI change readiness assessment include? Five things, at minimum: whether people can state why the organisation is adopting AI and which of their tasks it changes; whether they know what they are permitted to do and what they must disclose; whether they have had real practice and reachable help; whether their manager visibly uses AI and has discussed their role; and whether they believe what leadership has said about jobs. The MASSIVUE AI Change Readiness Check in this guide is a 13-item version you can field anonymously by team, before deployment and at six-month intervals. It is a structured practitioner assessment rather than a validated psychometric instrument, and it should not be used for individual performance assessment.

How do we prove our AI training actually worked? Baseline a work outcome before the training, assess proficiency on a real artefact after it, and compare the trained population against a comparable untrained one where you can. Course completion rates prove attendance and nothing more.

Is shadow AI a discipline problem? Almost never. Gartner found 88 percent of employees with enterprise AI access also use personal AI tools for business tasks. That is a signal that the sanctioned path is slower, less capable, or unclear on permission. Diagnose which, then fix it. Disciplinary responses mainly move the behaviour out of sight.

How does AI change management differ in Singapore? The people work is the same. The context is not. Job-redesign funding, tripartite consultation norms, sector skills standards and voluntary AI governance frameworks shape how the work is funded and sequenced, and financial institutions face an evolving regulatory position that will affect governance and capability evidence.

Is there government funding in Singapore for AI-driven job redesign? Yes. The SkillsFuture Workforce Development Grant for job redesign, administered by Workforce Singapore, supports up to 70 percent of qualifying costs for SMEs and up to 50 percent for non-SMEs, capped at S$150,000 per enterprise. It is support against qualifying costs, not a cash grant. Check eligibility before fixing your redesign scope, and confirm current parameters with Workforce Singapore.

Does the IMDA Model AI Governance Framework for Agentic AI apply to us? It is voluntary guidance rather than legislation, published in January 2026 and maintained as a living document. It is the natural reference point for setting autonomy tiers in Singapore deployments, and adopting it is a reasonable default even where nothing compels it.

What do the MAS AI risk management guidelines require on the people side? As at August 2026 the guidelines were still in consultation and not finalised, with a 12-month transition period proposed after issuance. The expected shape is board-level oversight, a maintained AI inventory, human oversight and override, and competence proportionate to risk. In practice that turns capability building and the change cadence into documented controls rather than good practice. Work from the final MAS text when it issues.

Should we use ADKAR, Kotter, or something else for AI change? Use an established model for the individual transition and add standing structural ownership for the organisation. ADKAR maps well to diagnosing where a person has stalled. What none of them handles natively is the absence of a go-live date, which is an organisational design problem rather than a personal transition problem.

What is the first thing we should do on Monday? Find out whether anyone has decided what happens to the time AI saves. If nobody has, that decision, not the technology roadmap, is your first deliverable.

Related MASSIVUE resources

• What Is an AI Operating Model? (August 2026) sets out how decision rights over AI are allocated, which is the authority structure this change work depends on.
• Why Generic AI Training Fails (April 2026) goes department by department on the skills layer.
• Why Enterprise AI Pilots Stall Before Production (June 2026) covers the delivery-side failure modes that sit alongside the people-side ones described here.
• End-to-End Change Management in Singapore: A 2026 Playbook (July 2026) covers general change management practice and the Singapore regulatory layer in more depth.
• Which Companies Offer AI Upskilling Programs in Singapore? (July 2026) compares named providers and current funding routes.
• Agentic AI Governance covers the control side of the rules layer.
• Protum™, MASSIVUE’s AI operating model framework, builds the six business capabilities that make AI and human work coordinate.
• AI Workforce Transformation and Enterprise Transformation are the service lines behind this guide.
• The MASSIVUE Academy offers an AI Change Management microcredential covering upskilling and reskilling, within a wider set of AI and transformation certifications.

Sources

Each figure cited above was checked against the publisher’s own material at the last review of this article. Where a source is preliminary, small-sample, narrowly scoped or lagging, that is stated at the point of use.

• BCG, AI Transformation Is a Workforce Transformation, February 2026. https://www.bcg.com/publications/2026/ai-transformation-is-a-workforce-transformation
• BCG, AI at Work: Why Strategy Matters More Than Tools, June 2026. https://www.bcg.com/publications/2026/ai-at-work-why-strategy-matters-more-than-tools
• BCG, AI at Work 2025: Momentum Builds but Gaps Remain, June 2025. https://www.bcg.com/publications/2025/ai-at-work-momentum-builds-but-gaps-remain
• McKinsey, Agentic AI change management: closing the adoption gap, August 2026. https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/how-to-close-the-agentic-adoption-gap
• McKinsey, The State of AI, November 2025. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
• Deloitte, State of AI in the Enterprise 2026, January 2026 (3,235 leaders, 24 countries, fieldwork August to September 2025). https://www.deloitte.com/us/en/about/press-room/state-of-ai-report-2026.html
• Gartner, AI projects in infrastructure and operations stall ahead of meaningful ROI, published 7 April 2026 (782 I&O leaders, fieldwork November to December 2025). https://www.gartner.com/en/newsroom/press-releases/2026-04-07-gartner-says-artificial-intelligence-projects-in-infrastructure-and-operations-stall-ahead-of-meaningful-roi-returns
• Gartner, 45% of managers report AI has lived up to their expectations, 4 March 2026 (July 2025 surveys of 2,986 employees, 1,973 managers and 114 HR leaders). https://www.gartner.com/en/newsroom/press-releases/2026-3-4-gartner-hr-survey-reveals-45-percent-of-managers-report-ai-has-lived-up-to-their-expectations
• Gartner, Top change management trends for CHROs in the age of AI, 16 March 2026. https://www.gartner.com/en/newsroom/press-releases/2026-3-16-gartner-identifies-top-change-management-trends-for-chros-in-age-of-ai
• Gartner, By 2027, 50% of enterprises without a people-centric AI strategy will lose top AI talent, 13 May 2026 (12,004 employees and managers, 40 countries). https://www.gartner.com/en/newsroom/press-releases/2026-05-13-gartner-predicts-by-2027-50-percent-of-enterprises-without-a-people-centric-ai-strategy-will-lose-their-top-ai-talent
• Cian O Morain and Peter Aykens (Gartner), Employees Are Losing Patience with Change Initiatives, Harvard Business Review, May 2023. The underlying survey is not separately published. https://hbr.org/2023/05/employees-are-losing-patience-with-change-initiatives
• Microsoft, 2026 Work Trend Index, 5 May 2026 (20,000 full-time employed or self-employed knowledge workers who use AI at work, 10 markets, fieldwork 18 February to 7 April 2026). https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization
• IBM Institute for Business Value, 2026 CEO Study: Rewiring the C-suite, 4 May 2026 (2,000 CEOs, 33 geographies). https://newsroom.ibm.com/2026-05-04-ibm-study-ceos-are-reshaping-c-suite-roles-for-the-ai-era
• Prosci, Why Your AI Rollout Stalled, May 2026, updated August 2026 (1,107 professionals). https://www.prosci.com/blog/adkar-for-ai-adoption
• Prosci, Measuring AI success in organization-wide adoption. https://www.prosci.com/blog/measuring-ai-success-in-organization-wide-adoption
• Mercer, Global Talent Trends 2026, January 2026. https://www.mercer.com/about/newsroom/as-organizations-race-to-adopt-ai-in-2026-marsh-s-mercer-says-empower-talent-and-redesign-work-to-achieve-meaningful-gains/
• Protiviti, AI Pulse Survey, Volume 5, July 2026. https://www.protiviti.com/us-en/survey/ai-pulse
• KPMG, AI Quarterly Pulse Survey, Q2 2026. https://kpmg.com/us/en/articles/2025/ai-quarterly-pulse-survey.html
• Wiley, New survey reveals growing risk of change fatigue “cascade crisis”, August 2025 (1,685 respondents, North America). https://newsroom.wiley.com/press-releases/press-release-details/2025/New-Survey-Reveals-Growing-Risk-of-Change-Fatigue-Cascade-Crisis-for-Employees-Navigating-AI/default.aspx
• Ministry of Manpower Singapore, Adoption of Artificial Intelligence Among Firms, April 2026. https://www.mom.gov.sg/newsroom/press-releases/2026/0430-adoption-of-ai-among-firms
• IMDA, Singapore Digital Economy Report 2025, October 2025. https://www.imda.gov.sg/resources/press-releases-factsheets-and-speeches/factsheets/2025/ar-sgde-2025
• IMDA, Model AI Governance Framework for Agentic AI, 22 January 2026. https://www.imda.gov.sg/resources/press-releases-factsheets-and-speeches/press-releases/2026/new-model-ai-governance-framework-for-agentic-ai
• Monetary Authority of Singapore, Consultation Paper on Proposed Guidelines on Artificial Intelligence Risk Management for Financial Institutions, 13 November 2025. https://www.mas.gov.sg/publications/consultations/2025/consultation-paper-on-guidelines-on-artificial-intelligence-risk-management
• Workforce Singapore, SkillsFuture Workforce Development Grant (Job Redesign). https://www.wsg.gov.sg/home/employers-industry-partners/workforce-development-job-redesign/wdg-jr
• NTUC, Tripartite partners announce new jobs council, 29 April 2026. https://www.ntuc.org.sg/uportal/news/Tripartite-partners-announce-new-jobs-council-to-support-workers-and-businesses-in-AI-transition/
• NTUC, AI-Ready SG, 13 February 2026. https://www.ntuc.org.sg/uportal/news/NTUC-launches-AI-Ready-SG-to-AI-enabled-economy/
• NTUC and IBF, Upskilling up to 100,000 finance professionals, 25 June 2026. https://www.ntuc.org.sg/uportal/news/NTUC-IBF-aim-to-upskill-100000-finance-professionals-in-AI-over-three-years/
• ADP Research, People at Work 2026, May 2026 (more than 39,000 working adults, 36 markets, fieldwork July to August 2025). https://sg.adp.com/about-adp/press-centre/adp-research-only-22-percentage-of-workers-globally-feel-their-jobs-are-safe-from-elimination.aspx
• Luke Plunkett, Welcome To The Resistance: Meet The Workers Dodging (And Sabotaging) Their Employer’s AI Mandates, Aftermath, 30 July 2026. Journalism with anonymous sources, cited as illustrative rather than as measured prevalence. https://aftermath.site/ai-resistance-tips-workforce-llm-workers/