AI adoption guide · New Zealand & Australia
Why AI rollouts stall, and how to make adoption stick
Many organisations have bought AI licences. Far fewer have changed how work actually gets done. This guide sets out what the evidence says about why AI rollouts stall, and what New Zealand and Australian leaders can do about it.
The adoption gap
- 88% of organisations use AI regularly in at least one business function McKinsey, State of AI 2025
- 39% report any impact on earnings from AI, and most of those say it’s under 5% McKinsey, State of AI 2025
- 1 in 3 Australian public servants with Copilot access used it daily, across 7,600+ staff in 60+ agencies Digital Transformation Agency, 2024
Whether you’re rolling out Microsoft Copilot, ChatGPT or Claude, the tool is rarely the problem. AI rollouts stall when the technology changes faster than the organisation around it: the workflows, the data, who owns the outcome, what people are measured on, and what happens to their roles.
In short
- There is rarely one cause. Most stalled rollouts show several at once.
- Redesign the work, not just the tools, and measure against a baseline.
- Each ADKAR® outcome needs its own response: more training won’t fix a problem that’s really about role security or incentives.
- Decide in advance what would make you stop, and scale only what works.
The causes
Why AI rollouts stall
There is rarely one cause. Most stalled rollouts show several of these at once, and they make each other worse.
-
The work stays the same
AI is added on top of existing processes, approvals and handoffs, so people get yesterday’s process with an extra step. The organisations getting real value are the ones redesigning workflows, not just adding tools.
-
No baseline, no clear value
Use cases are chosen because they’re exciting, not because there’s a measurable cost, delay or error rate to improve. Without a baseline, nobody can prove the investment is working, and funding quietly dries up.
-
Nobody owns the outcome
AI sits between IT, the business, security, legal and HR. When no named business leader owns the result, decisions stall and problems go unresolved.
-
No credible reason for people
Licences are announced, but nobody explains why, or what it means for each role. Worries about jobs, status and expertise go unspoken and slow adoption down.
-
Generic training
One product demo for everyone, rather than practical sessions built around what finance, project or HR teams actually do, including how to check AI output.
-
Data and permissions aren’t ready
AI tools surface whatever people can already access, so weak information management becomes a risk the moment AI is switched on.
-
Incentives reward the old way
People are asked to work differently but are still measured on the old activity. Few people sustain a behaviour their performance measures don’t reward.
-
No one measures, and nothing stops
Usage and value aren’t tracked after launch, and pilots that aren’t working are never shut down.
Is resistance the real problem? Less than many leaders assume. A preregistered experiment with 2,000 US and UK managers found that framing AI around job losses made them noticeably less willing to adopt and advocate for it, while information about productivity benefits made no significant difference on average. [6] That’s evidence about intentions, not proof that resistance explains every stalled rollout.
The practical lesson: when people hesitate, clarify what happens to their role, decisions and incentives before adding more communication.
The evidence
What New Zealand and Australian evidence shows
Some of the most useful evidence comes from public-sector trials, which publish what happened in detail.
- Broad access, moderate habit
Australian Public Service
More than 7,600 staff across 60+ agencies trialled Microsoft 365 Copilot in the first half of 2024. Sentiment was generally positive, but only about a third used it daily. The evaluation pointed to integration, data security and information management issues, uncertainty about acceptable use, and the need for training on agency-specific use cases. [2]
- Routine yes, complex no
Australian Treasury
Almost everyone in Treasury’s trial used Copilot at some point, but only 22% used it four to five times a week or more. Most found it useful for basic administrative tasks. Its value for more complex work was limited by the product itself and by privacy and security restrictions. [3]
- Agents ahead of governance
Australian business
Deloitte’s 2026 survey found 69% of Australian organisations using AI agents, but only 22% with a highly advanced model for governing them. Just 30% said AI was deeply transforming how their business works. [4]
- Model worked, workflow didn’t
Medicare compliance
The federal health department trialled an AI model to flag possible Medicare non-compliance. It flagged more cases than staff could investigate, and the trial was stopped. The Australian National Audit Office found the department had made no assessment of the benefits and costs of that decision. [5]
Lesson: AI has to be judged across the whole workflow, not on what the model can do.
What works
How to make AI adoption stick
The organisations that succeed treat AI adoption as one programme covering technology, the way work is organised, and change management, rather than three separate workstreams. In practice, that runs through four phases.
-
Before launch
Choose a small number of real processes with a measurable baseline. Name the business owner. Be honest about how roles will change. Agree the rules for acceptable use, data and human oversight, and decide in advance what would make you stop.
-
Launch
Redesign the workflow, not just the task: what the AI does, what people do, and who checks the output. Train people on their real work, by role. Leaders use the tools visibly and make it safe to raise concerns.
-
The first 90 days
Give people time to practise on live work. Put a capable user in each team to help others. Fix the data and permission problems that surface. Managers coach rather than simply approve.
-
Embedding
Measure against the baseline. Scale what works and stop what doesn’t. Update processes, role descriptions and performance measures so the new way of working becomes the normal one.
Governance should arrive early enough to set data boundaries and accountability, but not so heavily that it stops useful experiments. Too much control can push people towards unofficial tools as easily as too little creates risk. For board-level questions, see our AI governance guidance.
Change management
Applying the Prosci® ADKAR® Model to AI rollouts
Prosci’s ADKAR Model is a widely used change management model built around five outcomes a person needs for a change to stick. As a Prosci-certified practitioner, Glen finds it especially useful for AI, because AI rollouts often succeed on one outcome and stall on another. Here’s how we apply it, in our own words.
-
Awareness
People understand why the organisation is adopting AI, and what it means for their own work.
Where it often breaks Licences are announced with no business reason, or the reason sounds like a vendor’s pitch.
-
Desire
People choose to take part, not just comply.
Where it often breaks Concerns about jobs, status and expertise go unanswered, so people quietly opt out.
-
Knowledge
People know how to do the redesigned work, including when to use AI and how to check its output.
Where it often breaks Training covers the product’s buttons, not the role’s real tasks.
-
Ability
People can do it reliably in day-to-day conditions.
Where it often breaks No time to practise, and workflows or permissions still block the new way of working.
-
Reinforcement
The new way of working lasts.
Where it often breaks Nobody measures after launch, and performance measures still reward the old process.
The key point for leaders: each outcome needs its own response. Explaining why AI matters won’t create willingness, and more training won’t fix a problem that’s really about role security or incentives.
This guide is written independently by Matrix AI Consulting. It is not a Prosci publication and is not affiliated with or endorsed by Prosci. To learn more about the model, visit Prosci’s ADKAR page.
Metrics
What to measure
Licences issued, training attendance and prompt counts don’t show adoption. A useful scorecard covers whether the work actually improved, whether people keep using AI, and whether the risks are under control.
-
Business outcome
Cycle time, unit cost, backlog, quality of output, customer resolution, against the baseline.
-
Repeat use
Weekly active users in the target group, and the share of the workflow done with AI.
-
Quality
Error rates, rework and how often people override the AI’s output.
-
Risk
Incidents, policy breaches and sensitive-data events.
-
Workload
Time saved, and whether work has simply moved to another team or queue.
-
Capability
Whether people can do the redesigned work and check AI output, not course attendance.
-
Change health
Role clarity, confidence and visible sponsorship.
-
Economics
Total cost per working use case, including licences, support and integration.
-
Role redesign
Tasks removed, decisions that have moved, and where people have been redeployed.
Readiness checklist
12 questions to answer before you scale
Before you scale an AI use case, your leadership team should be able to answer these questions. If you can’t answer several of them, redesign the rollout rather than adding more communication or training.
Value and ownership
- What business outcome will change, and what is the baseline today?
- What result would make us stop rather than scale?
- Which named business leader owns the outcome, not just the technology?
Work and roles
- Which tasks disappear, move or become exceptions?
- Who handles the AI’s output downstream, and do they have the capacity?
- Whose role, decisions or expertise changes, and have we told them honestly?
- Do the people affected have a credible reason to support this change?
Foundations
- Are data permissions and information quality good enough for this use case?
- What can the AI do on its own, and when does a person review or step in?
Making it last
- Can people do the redesigned work and check AI output, or have they only attended a general course?
- Do performance measures and manager behaviour reward the new way of working?
- Can we show real value at 30, 60 and 90 days without relying on enthusiasm alone?
This checklist is Matrix AI’s own planning tool. It is not a Prosci assessment.
FAQ
AI adoption questions
Why do AI rollouts stall?
Rarely for one reason. AI rollouts stall when the technology changes work faster than the organisation changes ownership, workflows, incentives, capability, governance and role value. Common causes include unchanged workflows, weak data and permissions, unclear ownership, generic training and incentives that still reward the old way of working.
Is employee resistance the main reason AI adoption fails?
The evidence doesn’t support that. A preregistered experiment with 2,000 US and UK managers found that framing AI around job losses reduced their willingness to adopt and advocate for it, but that shows an effect on intentions, not that resistance explains every stalled rollout. Treat hesitation as a signal to clarify roles, decision rights and incentives.
How does the ADKAR Model apply to AI adoption?
Prosci’s ADKAR Model describes five outcomes people need for a change to stick: Awareness, Desire, Knowledge, Ability and Reinforcement. With AI they come apart easily: explaining why AI matters doesn’t create willingness, and training on its own doesn’t create sustained use. Looking at each outcome separately helps you see where adoption is stuck and fix the right problem.
Do we need Prosci training to use this guide?
No. The guide is written for leaders and project teams without a change management background. The ADKAR Model is a useful lens, but the causes, phases, measures and checklist all work on their own. If you want support applying them to your rollout, we can help.
What should we measure in an AI rollout?
Measure business outcomes, repeat use, quality, risk, workload, capability, change health, economics and role redesign. Avoid vanity measures such as licences issued, training attendance or prompt counts.
Planning an AI rollout?
Matrix AI doesn’t sell software, so our advice applies whichever tools you choose. We help leadership teams plan, lead and measure AI adoption.
- AI change management : planning and leading an AI rollout
- AI workshops : practical training for teams and leaders
- Copilot workshops : role-based Microsoft 365 Copilot training
Sources
Sources
- McKinsey & Company, The state of AI in 2025: Agents, innovation, and transformation , November 2025
- Digital Transformation Agency, Evaluation of the whole-of-government trial of generative AI , 2024, and executive summary
- The Treasury, Evaluation of a trial of generative AI (Copilot) , 2025
- Deloitte Australia, State of AI in the Enterprise , February 2026
- Australian National Audit Office, Artificial Intelligence and Medicare Benefits Integrity , 2026
- Melbourne Institute / ANU, Managers as Gatekeepers in the Age of AI , 2026
This guide is written independently by Matrix AI Consulting and is not affiliated with or endorsed by Prosci. Prosci and ADKAR are trademarks of Prosci, Inc., registered in the US and other countries.
Is your AI rollout stalling?
Tell us where your rollout is up to and what’s getting in the way. In a 20-minute call we’ll help you work out which part of the adoption system to fix first, with no obligation and no sales pitch. Prefer email? hello@matrixconsulting.ai