Bring People With You: The AI Adoption and Change Management Playbook
A WorkplaceAI Strategy Playbook
Most AI programs are launched as technology projects and stall as people projects. Access is not adoption, and adoption is not value. Here is how to move a workforce from a license to a changed way of working.
Buying the tool is the easy part. Getting several hundred people to change how they do their jobs is the hard part, and it is the one most AI programs underfund.
This playbook treats AI rollout as what it is: A change management problem with a technology attached. It is written for the leader who owns the outcome, whether that sits in operations, HR, IT, or a business unit.
It pairs with the AI ROI Measurement Playbook. That one proves whether the change paid off. This one makes the change happen.
Use is rising everywhere. Support for the people doing the using is not keeping pace. Several recent studies show the same gap from different angles.
| Source | What it found |
|---|---|
| PwC Global Workforce Hopes and Fears Survey 2026, 49,364 workers in 48 countries, fielded May and June 2026 | 64% used AI at work in the past 12 months, up 10 points. Daily generative AI use rose from 14% to 22%. But 51% said they have access to the learning resources they need, down from 59%. Among daily users, 68% felt confident about job security, versus 57% of infrequent users. |
| Gallup, U.S. workers, 43,262 responses, February and May 2026 | Where the employer has a clear plan for AI, 43% of workers were engaged, versus 28% where there is none. Where the manager actively supports AI use, 48% were engaged, versus 30%. Frequent users were far more likely to rate AI’s productivity impact at the top level than infrequent users (24% versus 5%). |
| BCG AI at Work 2025, more than 10,600 employees in 11 countries and regions | Over three-quarters of leaders and managers use generative AI several times a week, but regular frontline use has stalled at 51%. Only about one-third of employees say they have been properly trained. Positive sentiment toward generative AI rises from 15% to 55% with strong leadership support, which only about a quarter of frontline employees report. |
| Grant Thornton 2026 AI Impact Survey, 950 business leaders, February and March 2026 | 34% of finance leaders say training is the most underfunded AI investment area. Only 6% of executives rate change leadership and workforce enablement as an essential skill. Frontline employees and middle managers are named as the groups needing the most support. |
| WRITER and Workplace Intelligence, 2,400 respondents (vendor-published) | 29% of employees admit to undermining their company’s AI strategy, and 35% say their manager is an AI champion. These are self-reported figures from an AI vendor’s own survey, so treat them as directional. |
Two things stand out. The workers closest to the work get the least support, and the factors that predict good outcomes are mostly managerial: A clear plan, an engaged manager, and time to learn. None of them is a software feature.
Most rollouts begin with the tool. A license is purchased, an announcement goes out, and a training session is scheduled. Adoption is then judged by logins.
This works backward. People do not change behavior because a tool exists. They change when a specific piece of their work becomes easier, safer, or more rewarding to do a new way, and when the people around them make that safe.
The playbook rests on three questions, asked in order.
Skip the first question and training becomes generic. Skip the second and people stall. Skip the third and the old habits return within a quarter.
The eight stages below build on each other. Each one answers a question that, left unanswered, will surface later as resistance.
You do not need to run them as a strict sequence. You do need all eight, and the first three before anything is announced.
Begin with one workflow, not the whole company. Write down what it looks like today and what it should look like after AI is in place.
The last item matters more than it looks. A change framed only as a cost saving tends to be received as a threat. A change that also removes a tedious task is easier to adopt.
If you cannot fill in that list, the rollout is not ready. Vague instructions to “use AI more” produce scattered, shallow use that is hard to measure and easy to abandon.
A workforce does not adopt as one body. Different groups need different handling, and treating them identically wastes effort on the eager and loses the cautious.
| Group | What they are like | What they need |
|---|---|---|
| Early adopters | Already experimenting, often on their own. Quick to find uses. | Official channels, recognition, and a role in helping others. Left alone, they build unmanaged workarounds. |
| Pragmatists | Willing if it clearly helps with a real task. The largest group. | Role-specific examples, protected practice time, and proof that colleagues use it well. |
| Skeptics | Doubt the quality, the point, or the intent. Often experienced and respected. | Honest answers about limits, a say in the review rules, and evidence from their own work. |
| Exposed roles | Worry the work they do is the work being automated. | A clear statement about the future of the role, a path to new skills, and no surprises. |
| Managers | Responsible for outcomes but often not trained or briefed first. | Preparation ahead of their teams, a script for hard questions, and clarity on what to measure. |
Frontline staff and middle managers deserve extra attention. Grant Thornton’s survey named both as the groups needing the most support, and BCG found frontline regular use stuck at 51% while most leaders used AI weekly. Leaders often judge adoption by their own experience, which is not their team’s.
Every change carries an implied deal between the company and its people. If you do not state it, employees will write their own version, usually a worse one.
A good message is specific, repeated, and consistent from the top to the team lead. It answers the questions people are asking quietly.
Do not promise what you cannot guarantee. PwC’s survey found that daily AI users were more confident about job security than infrequent users, 68% versus 57%. Familiarity appears to reduce fear. But that is a reason to give people hands-on experience, not a reason to dismiss concerns before they have it.
Back the message with a short written policy. The Agent Governance Playbook covers ownership and boundaries for the agent side of this.
Generic tool training teaches where the buttons are. People need to learn how their own work changes. Those are different lessons.
PwC’s result is a warning: Use grew while access to learning resources fell from 59% to 51%. BCG found only about a third of employees say they have been properly trained, and that regular use was sharply higher among those who received at least five hours of training plus coaching. Grant Thornton found training is the area finance leaders most often call underfunded.
Include the review skill. Using AI well means knowing how to spot a confident wrong answer. Teach the checks, and show real examples of errors in your own domain.
Then keep teaching. A tool will change within months. A one-time course is out of date before the next quarter ends, so plan a light recurring rhythm, such as a monthly show-and-tell of what worked.
Of all the levers here, the manager is the strongest. Gallup’s 2026 data shows it plainly.
Among AI users in organizations that have adopted it, 33% of those whose manager actively supports AI use gave the top productivity rating, compared with 9% of those without that support. A clear organizational plan mattered too, 21% versus 12%. When frequent use, a clear plan, and manager support all came together, half gave the top rating.
Gallup’s tables carry some labeling inconsistencies, so read the exact percentages as indicative. The direction is unmistakable.
Middle managers sit between a leadership mandate and a team that feels it. Give them the information and cover to translate one to the other, or they become the place where the program quietly stalls.
Peers persuade better than announcements. A small group of respected practitioners, drawn from each team, can show what works and report what does not.
The feedback matters as much as the advocacy. Champions are your early warning. They hear the complaints and see the workarounds before the dashboards do.
Treat what they bring back as data. If three teams report the same failure, that is a defect in the rollout, not a flaw in the people.
If people respond to incentives, a rollout that ignores them is asking the workforce to act against its own interests. Check these four before blaming resistance.
| Question | Why it matters |
|---|---|
| Does saving time earn more work, or recognition? | If efficiency is rewarded with a bigger queue, people learn not to be efficient. Decide where freed time goes and say so. |
| Is it safe to use AI and make a mistake? | If an AI-assisted error is punished more than an unaided one, people avoid the tool or hide their use. |
| Is it safe to admit you use it? | If staff fear being seen as replaceable, they will use it quietly, outside the rules, or not at all. |
| Do performance reviews reflect the new way of working? | If goals and reviews stay unchanged, the new behavior is optional, and optional things lose. |
Be careful with coercion. The WRITER survey, which is vendor-published, reports that most executives plan layoffs for people who will not adopt and warns that those who are not AI proficient will not be promoted. Whatever you think of those intentions, threats alone tend to produce compliance theater: Visible usage, shallow results, and hidden workarounds.
A better approach pairs a clear expectation with real support. Set what is expected by when, provide the training and time to meet it, and recognize people whose work improves.
Also revisit job design where AI genuinely changes the work. Update descriptions, skills expectations, and career paths, and make sure the people affected see the plan. Surprises do more harm than bad news delivered early.
Logins show access. They do not show change. A useful adoption measure asks whether the work is actually done differently, and whether it is done better.
| Measure | What it tells you |
|---|---|
| Active use by team | Where adoption is real and where it is thin. Look at weekly users, not accounts. |
| Depth of use | Whether people use it for substantive tasks or only simple ones. |
| Task coverage | What share of the target workflow is now done the new way. |
| Quality and rework | Whether the output is good enough, and how much correction it needs. |
| Sentiment and confidence | Whether people feel equipped, and whether the picture differs by manager or team. |
| Outcomes | Cycle time, error rate, cost, or customer measures, tied back to the baseline. |
Take the outcome measures from the AI ROI Measurement Playbook, so adoption and value are tracked from the same baseline. Adoption without outcomes is activity. Outcomes without adoption data cannot be explained.
Adoption decays. After the launch energy fades, people drift back to old habits, new hires miss the training, and the tool changes underneath them.
Resistance is information. Before treating it as a problem to overcome, find out what it is telling you.
| What you hear | What may be behind it | A useful response |
|---|---|---|
| “It gets things wrong.” | A fair concern about quality, or a bad early experience. | Share known failure patterns, teach the checks, and let skeptics shape the review rules. |
| “I don’t have time to learn it.” | Workload, with no protected time for learning. | Block time on the calendar and remove something else. |
| “This is going to replace me.” | Real fear, often unaddressed by leadership. | Give an honest, specific statement about the role, and a path to new skills. |
| “I already do this faster.” | Expertise that the generic training did not respect. | Pair them with a champion role, and use their method to improve the guidance. |
| Silence, and quiet non-use. | Unclear expectations, or fear of looking incompetent. | Ask managers to hold one-to-one conversations and offer private practice. |
| Heavy use outside approved tools. | Official tools are too limited, or the rules are unclear. | Learn why, offer a safe alternative, and fix the policy gap. |
Some resistance is justified. If a team says the tool produces unreliable output for a task, investigate before pushing harder. They may have found a limit you should know about.
The following example uses invented figures to show how the stages fit together. It is not drawn from a real company.
A 180-person customer support organization licenses an AI drafting assistant. Leadership sends an announcement and a recorded training video. After eight weeks, 70% of agents have logged in, but only 22% use it weekly. Edits to the drafts are heavy, and two team leads say privately that they do not trust it.
Twelve weeks later, weekly use is 64% and the edit rate has fallen as agents learn which tickets suit the tool and which do not. Two ticket types turned out to be poor candidates and were removed from the plan.
The numbers are made up, but the sequence is realistic. The tool did not change. The surrounding conditions did.
| Failure | What it looks like | Fix |
|---|---|---|
| The announcement is the plan | A launch email and a video, then a login report. | Define the work change and train inside the job. |
| Training is generic | Everyone gets the same course, and no one changes a task. | Build role-based practice on real work. |
| Managers are briefed after their teams | Leads face questions they cannot answer. | Brief and equip managers a week ahead. |
| Usage is the only metric | A dashboard shows logins while the work stays the same. | Track depth, coverage, quality, and outcomes. |
| Efficiency is punished | Faster work brings a bigger queue. | Decide where freed time goes, and say it. |
| Fear is ignored | Rumors fill the silence, and strong people leave. | State the deal honestly and give a path to new skills. |
| Champions are volunteers without time | The network fades within a quarter. | Give them protected hours and a defined role. |
| The rollout is declared done | Use decays, and new hires are missed. | Plan refreshers, onboarding, and quarterly reviews. |
Before announcing a rollout, check that each item is in place. A gap here is a risk you can still fix cheaply.
Resist the urge to extend too early. A rollout that works for one workflow and one team is a pattern you can copy. One that has not been proven in a single place cannot be scaled safely.
The surveys point in one direction. Companies are buying and deploying AI faster than they are preparing people to use it, and the gap shows up in thin results and uneasy workforces.
That is a leadership problem with a practical answer. Define the work, tell people the truth, train them in context, prepare their managers, reward the new behavior, and measure whether the work changed.
Start with one workflow and do it properly. The people who see it work will carry the next one.
The ROI Measurement Playbook shows how to prove the change paid off, and the Agent Governance Playbook covers ownership and controls. Our CIO survey report covers why leaders feel the pressure.
AI ROI Measurement Playbook → Agent Governance Playbook → CIO Survey Report → Employee Lifecycle Playbook → Automation ROI Estimator →Sources: PwC, Global Workforce Hopes and Fears Survey 2026 (49,364 workers, 48 countries, fielded May and June 2026), as reported by The Gleaner on October 14, 2026. Gallup, Employee Engagement Remains Flat as AI Adoption Accelerates (43,262 U.S. workers, February and May 2026). BCG, AI at Work 2025: Momentum Builds, but Gaps Remain (more than 10,600 employees, 11 countries and regions, June 2025). Grant Thornton, 2026 AI Impact Survey (950 business leaders, February 23 to March 18, 2026). WRITER and Workplace Intelligence, 2026 enterprise AI adoption survey (2,400 respondents), published by WRITER, an AI vendor; figures are self-reported and the field dates are not stated. The walkthrough uses invented figures for illustration only. Survey figures are as published by each source and were not independently verified.