Iron Goo
Iron Goo robot checks an open ledger beside survey sheets and a laptop chart in an office at dusk.

AI change management that gets tools used

Atamyrat Hangeldiyev
Atamyrat Hangeldiyev
Systems Architect
January 22, 2026
On this page
AI & Automation

AI change management is the work of getting people to adopt changed tasks, tools and responsibilities. It should not start by assuming employees fear replacement: in Gallup's February 2026 survey, 46% of non-users with AI available preferred their current way of working; 10% cited concern about their role being replaced. Gallup's reasons-for-non-use analysis supports a more useful starting point: ask what makes the existing way preferable, then fix the task fit and trust problems.

For an owner who has already bought licenses, that distinction changes the next expense. A motivational presentation cannot connect a tool to your invoicing system or settle which customer data employees may enter. Those are decisions you can make and verify.

Common claimWhat the evidence saysWhat to do instead
1. Most staff already use AIPew, September 2025: 21% of workers self-reported doing at least some work with AI.Read your own usage log.
2. Job fear explains resistanceGallup, February 2026: 46% of non-users with access preferred current methods.Ask about habit, privacy, ethics and task fit.
3. AI is already eliminating jobs everywhereDanish study, March 2026 revision: earnings and hours effects above 2% ruled out.Explain the mixed evidence and your own commitments.
4. One training session fixes adoptionUK DBT pilot, October to December 2024: satisfaction associated with at least 2 hours of self-led practice.Protect practice time on real tasks.
5. Access produces useDanish study: take-up about 41% without initiatives, 93% with all three.Integrate the tool, support use and publish rules.
6. Satisfaction proves adoptionUK DBT pilot: 72% satisfied among diary respondents; 30% mean daily activity on working days.Separate sentiment, usage and work results.
7. Tasks disappear and roles shrinkDanish study: entirely new tasks reported by about 8% of users without initiatives, roughly 18% with initiatives.Assign review, ownership and exceptions.
8. Reward heavy users and prohibit other toolsMicrosoft 2024 survey: 78% of AI users brought their own tools.Reward outcomes and provide an approved route.
9. A 70% failure rate demands a frameworkHughes, 2011: no valid, reliable evidence supported that figure.Use frameworks as diagnostic checklists.
46%
Prefer the current way of working
18%
Expect their job could be eliminated
72% / 30%
Satisfied / daily active
19% slower
Measured developer task completion

1. “Most staff already use AI”: the denominator changes the answer

The claim: Your employees already use AI elsewhere, so workplace adoption should require little effort. A large survey percentage usually supplies the reassurance.

What the evidence says: Pew surveyed 5,010 employed US workers from September 2 to 8, 2025: 21% said at least some of their work was done with AI, while 65% reported little or none. Gallup's February 4 to 19, 2026 survey of 23,717 employed US adults used frequency: 13% reported daily use, 28% a few times weekly or more, and 50% at least a few times yearly. These are self-reports with different thresholds.

The Microsoft and LinkedIn Work Trend Index 2024 reported 75% using AI at work. This vendor survey, conducted by Edelman Data & Intelligence from February 15 to March 28, 2024, covered 31,000 full-time knowledge workers across 31 markets. Knowledge workers are a selected population, and the vendor sells workplace AI.

The Census Business Trends and Outlook Survey reported 24.6% of US employer businesses using AI in any business function during September 21 to October 4, 2026. That measures businesses over two weeks, not individual workers. None of these percentages estimates adoption in your company.

Three measuring tubes show different AI adoption levels for workers, knowledge workers and businesses.

The apparent disagreement comes from the population and the definition of use. A business using one automation and an employee using a chatbot daily answer different questions.

What to do instead: Export usage for your own licensed population and name the reporting window. Record enabled users, active users and days of activity, then check which applications the report includes. Keep people without licenses outside that denominator; report company-wide access separately.

2. “They fear losing their jobs”: current methods lead the stated reasons

The claim: Employee resistance to AI is mainly anxiety about replacement. Reassure everyone about jobs and the resistance should recede.

What the evidence says: Gallup's February 2026 survey asked non-users and infrequent users in organizations making AI available why they did not use it more. Respondents could select several reasons, so the columns must never be added together or presented as shares of all employees.

Self-reported reason, Gallup February 2026Non-usersInfrequent users
Prefer current ways of working46%36%
Data privacy, security and compliance concerns43%38%
Ethical opposition to AI43%25%
Believe AI cannot assist their work39%22%
Concern about their role being replaced by AI10%10%

Job anxiety is real. In the same February survey, 18% of US employees considered elimination of their job through AI or automation within five years very or somewhat likely; among employees in AI-adopting organizations, it was 23%. Those questions measure expectations about employment, not reasons for skipping a provided tool.

Pew's February 25, 2025 publication drew on 5,273 employed US adults surveyed October 7 to 13, 2024. It found 52% worried about AI's future workplace impact, 36% hopeful and 32% expecting fewer job opportunities for themselves. Worry and non-use are related questions, but they are not interchangeable measurements.

What to do instead: Ask each non-user which of the first four reasons applies, allowing several answers. For current-method preference, watch the existing workflow and count the extra steps the AI route requires. For privacy or ethical concerns, identify the specific objection and make a written decision; for poor task fit, test one real task with that employee.

Do not label a documented limitation “resistance.” If the tool needs copying between systems or produces output someone must substantially repair, the employee may be making a sensible choice.

3. “AI is eliminating jobs everywhere”: the evidence differs by population

The claim: AI is already producing mass job losses, or, at the opposite extreme, nobody needs to worry about employment. Neither statement gives employees a dependable basis for planning.

What the evidence says: Humlum and Vestergaard's Still Waters, Rapid Currents, issued in May 2025 with a March 2026 revision, linked surveys covering 25,000 Danish workers across 7,000 workplaces and 11 exposed occupations to administrative records. It found precise null effects on earnings and recorded hours, ruling out effects larger than 2% two years after ChatGPT's launch. Adopters themselves reported average earnings impacts of only 0% to 0.5%, alongside reorganized tasks. About 6% of employers explicitly prohibited chatbot use.

The Census working paper CES 26-25, April 2026, reports that 2.0% of AI-using firms reported AI-related employment decreases. About two-thirds of firms reporting a task effect used AI only to augment tasks. The paper has not received the review given to official Census publications, and these are firm reports, not an independently identified causal employment effect.

The Stanford Digital Economy Lab's August 2026 revision used ADP payroll data through June 2026. Employment among 22 to 25 year olds in highly exposed occupations was about 19% below the level implied by keeping pace with similarly aged workers in less-exposed occupations. That is a gap against a comparison trend, not a 19% employment fall.

Stanford reports that adjustment appears to come mainly through reduced hiring, concentrated where AI automates tasks. The authors describe the patterns as descriptive, not causal, and report no widespread economy-wide displacement. Danish incumbent-worker records and US early-career hiring patterns can both be true.

What to do instead: Tell staff that the measured evidence includes changed tasks with limited earnings and hours effects, plus a concerning entry-level hiring signal in automatable work. Then state what your company controls: staffing decisions for this rollout, reassignment arrangements and the period covered by any commitment. Promise only what the owner has actually approved, with a review date.

4. “A training session fixes adoption”: practice and relevance need attention

The claim: Buy the tool, book a class and treat attendance as readiness. Anyone still struggling needs another class.

What the evidence says: The UK Department for Business and Trade Copilot evaluation covered a 1,000-license pilot from October to December 2024, with approximately 70% volunteers and 30% randomly selected participants. At least two hours of self-led training and practice was associated with greater satisfaction and willingness to recommend Copilot. Formal departmental classes showed a satisfaction association only at five or more hours.

Participants often described formal sessions as introductory and insufficiently relevant to their roles. The measured outcomes were satisfaction and recommendation, not output or productivity. The results do not establish a universal training dose.

Microsoft's 2024 knowledge-worker survey found that 39% of people using AI at work had received company training. In the Danish study, training as the only employer initiative was associated with about 69% take-up, compared with about 41% without initiatives. These Danish comparisons bundle employer initiatives with who chooses to use the tools; they are associations.

What to do instead: Schedule protected practice using each person's actual tasks and permitted data. Two hours is a defensible starting allocation drawn from the UK pilot, not a guarantee of competence. Have the employee complete a task, check the output and explain when they would reject it.

Write a short playbook per role: suitable tasks, inputs, review checks and escalation. Use skills planning for the AI transition to decide which capabilities need developing beyond tool operation. Attendance records tell you who attended; task completion tells you what they can do.

5. “Availability produces use”: integration and manager behavior are associated with adoption

The claim: Providing a license removes the barrier. The remaining decision belongs entirely to the employee.

What the evidence says: Within organizations making AI available, Gallup's February 2026 survey found the following frequent-use splits. “Frequent” meant a few times a week or more; the comparison was between employees strongly agreeing with each statement and those not strongly agreeing.

Workplace conditionStrongly agreeDo not strongly agree
AI integrates well with existing systems and processes88%55%
Manager actively supports AI use78%44%
Organization supports experimentation72%44%
Organization has clear AI policies68%47%

These are correlations, not effects from controlled interventions. Frequent users may also judge their systems and managers more favorably.

The Danish study found take-up of about 41% without employer initiatives, 76% with encouragement alone and 93% with encouragement, an enterprise chatbot and training together. These descriptive associations also include differences in who chooses to use AI. In the UK DBT pilot, users were anecdotally less likely to use Copilot when colleagues and line managers were hesitant.

What to do instead: Connect the tool to the system where the task already happens, and have the manager demonstrate a useful task with its review step. Publish a one-page policy naming approved tools, permitted data and the person who resolves uncertainty. Check whether following the policy is practical during normal work.

Involve the people doing the work in mapping it and testing the automation in shadow mode before replacement; that is the approach available through operations automation work. A human should verify the shadow output while the existing process remains authoritative. Set the replacement date after the agreed acceptance checks pass.

6. “Satisfaction proves adoption”: read the log and the work result separately

The claim: Employees like the tool, so the rollout has succeeded. A positive survey becomes the adoption report.

What the evidence says: In the UK DBT pilot of October to December 2024, 72% of diary respondents were satisfied or very satisfied, with a Net Promoter Score of 31. The diary response rate was 32%, so those answers did not represent every license holder. Usage records described a different aspect of the rollout.

Two valid measurements from the UK DBT pilot
What respondents said

Among diary respondents, 72% were satisfied or very satisfied and Net Promoter Score was 31. The response rate was 32%.

What the usage log recorded

Mean daily activity was 30% of license holders on working days, or 21% across calendar days. Mean weekly activity was 64%.

Daily activity peaked at 50% on October 8, 2024. Across the pilot, the average was about 1.14 Copilot actions per user per working day, and the evaluation found no evidence that time savings led to improved productivity. Satisfaction, activity and business improvement require separate measurements.

METR's July 2025 developer study randomized AI access across 246 tasks completed by 16 experienced open-source developers. They expected 24% faster work and afterward felt 20% faster, but measured completion was 19% slower. METR now flags those results as outdated and points to a February 2026 update; the original study used early-2025 tools and cannot be generalized beyond its small software-development sample.

The Danish study adds a self-reported measure: adopters reported saving about 3% of work hours. That is useful evidence about perceived savings, but it is not a measured company productivity gain.

A high satisfaction score sits beside a sparse daily usage log for the same AI rollout.

A happy survey and a sparse activity log can describe the same rollout. Read adoption from recorded use, then check whether the work improved.

What to do instead: Open the Microsoft 365 admin center at Reports > Usage > Microsoft Copilot > Copilot, then select the Usage tab. The Copilot usage report documentation, updated August 18, 2026, specifies 7, 28, 90 and 180-day windows, with data typically available within about 48 hours. Export the CSV and retain the window dates.

Read Enabled Users, Active Users, Active users rate and Last activity date. Opening the Copilot pane without submitting a prompt does not count as active. Active Days counts only days with prompts submitted to Microsoft Copilot Chat; read the per-app last-activity-date columns for Word, Excel, PowerPoint and Outlook too.

Consider a hypothetical 45-person company with 30 enabled licenses over 28 days. Assume the readout concerns Copilot Chat, and define habitual use as at least 10 active days, an explicit local threshold assuming roughly 20 working days. This is not Microsoft's definition of habitual use.

Hypothetical export inputPeople
Enabled licenses30
Users with at least one Copilot Chat active day21
At least 10 active days9
Between 1 and 9 active days12
Zero active days9
Hypothetical 28-day adoption readout

copilot_chat_active_day_share = users_with_at_least_one_copilot_chat_active_day / enabled_users = 21 / 30 = 70% habitual_share = people_with_10_or_more_active_days_in_28 / licences = 9 / 30 = 30%

Microsoft's report's own Active users rate counts activity in any Copilot app and may be higher than this Copilot Chat share.

The 70% headline includes 12 people who used Chat on only one to nine days. Check per-app activity before treating the nine zero-day records as non-use; if those are also empty, ask those nine people about Gallup's four leading reasons. The 10-day threshold is a management assumption, not proof that someone's work needs daily AI.

Do not turn the export into a leaderboard

Microsoft hides names by default in usage reports. Identify people only for a stated support purpose: using the log to rank or punish employees gives them a reason to hide use.

For an automation, count the old path. In a second hypothetical readout, 400 invoices arrive during a month, 310 pass through the automation and 90 are keyed by hand: adoption is 310 / 400 = 77.5%. Read the 90 by supplier, person and reason, distinguishing unreadable formats from distrust or habit.

Use a separate business-system measure for whether the work improved. Measuring AI automation ROI starts where activity counts end: compare the business result with its pre-launch baseline.

QuestionSourceMetricCadence
Is it used?Usage logActive users rate and active days per person per 28 daysWeekly
Why not?15-minute conversations with non-users using Gallup's reasonsReasons for non-useOnce around week four
Does it help the work?Business systemOutput per person-hour or cycle time against the pre-launch baselineMonthly
Is the old way still running?Old spreadsheet edit history and old inbox volumeCount of items done the old wayWeekly

7. “Tasks disappear, so roles shrink”: new work needs an owner

The claim: Removing production steps simply leaves each person with less work. Review and maintenance will somehow fit around existing responsibilities.

What the evidence says: The Danish study found about 8% of chatbot users taking on entirely new tasks even without employer initiatives, rising to roughly 18% where employers ran chatbot initiatives. Among new AI-related tasks, integration accounted for about 26%; quality review together with ethics and compliance accounted for about 35%. Users also self-reported reallocating saved time to other job tasks at a rate of 85%.

The NBER customer-support study, published in the Quarterly Journal of Economics in 2025, covered 5,179 agents. AI assistance increased issues resolved per hour by 14% on average and 34% for less-experienced and lower-skilled agents, with near-zero gains for the most skilled. Attrition fell, particularly among newer agents, and agents retained discretion to ignore suggestions.

That evidence concerns a particular support setting, not every role. It does show why a single productivity assumption applied across seniority levels is a poor staffing calculation.

What to do instead: Write down the changed role before launch, using organization and role design for AI adoption to separate production from review and ownership. Name the person who reviews and releases output, the owner of examples and exceptions, and the person working the exception queue. Allocate time to those duties.

Set the date the old way stops, subject to explicit acceptance checks. Specify the fallback route for failures so retiring the old process does not leave employees improvising a replacement under pressure.

8. “Reward heavy users and ban other tools”: incentives can hide the work

The claim: Prompt leaderboards encourage adoption, while a temporary public-tool ban contains risk until a policy exists. Both assume employees will respond exactly as management intends.

What the evidence says: In Microsoft's 2024 knowledge-worker survey, 78% of AI users reported bringing their own tools, rising to 80% at small and medium-sized companies. Among AI users, 52% were reluctant to admit using AI for their most important tasks, and 53% worried that doing so made them look replaceable. These self-reports describe incentives for quiet use.

BCG's AI at Work 2025, published June 26, 2025, surveyed more than 10,600 leaders, managers and frontline employees across 11 countries and regions. More than half said they would find alternatives and use AI tools anyway if restricted; that measures stated intention. The Danish study found about 6% of employers explicitly prohibited chatbot use, while adopters reported average earnings impacts of only 0% to 0.5%.

What to do instead: My opinion: reward checked work outcomes, not prompt counts, because employees can increase activity without improving the result. Provide an approved tool with a clear data rule and explain what saved time is for. A ban without a workable alternative risks moving use out of sight.

Ask which external tools people already use without demanding the content of their conversations. Give them a route to disclose a useful workflow and request approval. Make the manager responsible for resolving that request, with a named decision date.

9. “Seventy percent fail, so buy a framework”: use a diagnostic checklist

The claim: A supposedly universal change-failure rate proves you need a formal methodology. Naming the methodology then substitutes for checking whether it helps.

What the evidence says: Mark Hughes's 2011 review examined the published sources behind the 70% figure and found no valid, reliable empirical support. Kotter's March-April 1995 Harvard Business Review article drew on watching more than 100 companies and described eight errors. It supplied no universal failure percentage.

Kotter's current eight steps are urgency, coalition, strategic vision, volunteer participation, removing barriers, short-term wins, sustained acceleration and institutionalizing change. According to its abstract, Appelbaum and colleagues' 2012 review found support for most individual steps but no formal study of the whole model, and suggested its popularity owes more to its usable format than to evidence of results. Harrison and colleagues' 2021 healthcare review included 38 studies, 19 using Kotter, and could not detect whether using a model contributed to success.

ADKAR is Jeff Hiatt of Prosci's individual-level model, set out in his 2006 book: Awareness, Desire, Knowledge, Ability, Reinforcement. The published support we found comes mainly from its vendor's practitioner research. None of that establishes that adopting either named framework causes a successful rollout.

What to do instead: Use ADKAR to identify the missing element for each person. Treat the following mapping as a practical diagnostic, not a validated test; a training session directly addresses Knowledge, while the other symptoms require different decisions.

ElementSymptom you see in that personMatching fix
AwarenessCannot explain why the task is changingState the business problem and its baseline measure.
DesireSees no gain or sees personal riskDiscuss saved time, responsibilities and actual job commitments.
KnowledgeCannot perform their own task with the toolProvide task-specific practice and a written playbook.
AbilityKnows the steps but cannot complete the workflowFix access, integration or the output limitation.
ReinforcementReturns to the old format that managers still acceptAlign manager behavior, the procedure and the retirement date.

My view: compress Kotter for a small company because the coordination work matters more than ceremony. State urgency as hours spent on the task; make the coalition the owner and two people doing it; write the future job in one sentence; invite initial volunteers. Remove workflow barriers, measure one task within 30 days, review usage weekly, and put the accepted process into the role description and operating procedure.

What the staff announcement must answer

Write the answers before announcing the rollout, and have someone doing the work check whether they are specific enough to act on. An unresolved item needs a named decision owner and date.

  • Name the exact task that changes and which work remains unchanged.
  • State what the tool does and what the person still owns, including review and release.
  • Identify who carries the consequence of a wrong automated output and who can stop the process.
  • Specify which data must not enter which tool.
  • Explain what happens to saved time and who assigns the resulting work.
  • State what the owner actually commits to about jobs, how long that commitment lasts and when it will be reviewed.
  • Give the date the old way stops and the acceptance checks required before retirement.
  • Name where to report a wrong output, who receives it and how urgent failures are escalated.
Free ebook

Put your next steps in order

Build on what you have just read with a free ebook on semantic SEO and AI visibility. Get worked examples, decision rules, checklists and a do-this-week action in every chapter, plus a six-week plan for your site.

By signing up and confirming your address, you consent to receive the free ebook plus one email every two weeks with 3 or 4 new Iron Goo videos, guides and articles, and nothing else; unsubscribe any time with one click.

Related in AI & Automation

Ready to move?

Tell us where your business is today. A written assessment lands in your inbox within two business days.

Talk to us