From Sanskriti Khandelwal | Product & Market Analysis
Why Employees Quietly Sabotage Your AI Rollout, and What Actually Fixes It
On this page
54% of employees bypassed their company's AI tools in the past 30 days and completed the task by hand. Almost none of them filed an objection. That is the real problem in AI change management. It is not a debate you can win in a town hall, it is a private decision made hundreds of times a day at individual desks, and your adoption dashboard will not show it.
Key takeaways
- Avoidance is the majority behaviour, not the fringe. WalkMe's 2026 survey of 3,750 executives and workers across 14 countries found 54% bypassed company AI tools in the previous 30 days and 33% had not used AI at all.
- Active sabotage is real but smaller than the quiet version. 29% of knowledge workers told Writer and Workplace Intelligence they had deliberately undermined their employer's AI strategy, rising to 44% among Gen Z respondents.
- The trust gap is 52 points wide. 61% of executives trust AI for complex, business-critical decisions. Among workers the figure is 9%, which means the two groups are not disagreeing about a tool, they are describing different jobs.
- Visible peer use beats mandates. In Microsoft's own rollout study, engineers whose skip-level peers had adopted a coding agent were 216% more likely to try it, and manager use raised sustained adoption too.
Why do employees resist AI tools at work? Because the tool threatens something they hold: the interesting part of the job, the authority to decide, or the standing that came from being the person who knew. Resistance is rarely announced. It shows up as quiet avoidance, unsanctioned tools used in private, and superficial compliance that keeps the old process running underneath the new one.
What passive resistance to an AI rollout actually looks like
Open sabotage is easy to manage because it is visible. Somebody objects in a meeting, and you can answer them. The behaviours that decide the outcome of an AI rollout are the ones nobody raises.
They are also well documented now. Three separate 2026 surveys, fielded independently, describe the same set of moves.
The four behaviours that show up in the survey data
The first is bypass. The employee opens the approved tool, closes it, and does the task the old way. WalkMe measured this at 54% in a 30-day window.
The second is substitution. Rather than refuse AI, the employee uses a different model on a personal account. 45% of workers reported using unsanctioned tools in the same 30 days. That is not resistance to AI. It is resistance to your AI, and it carries its own exposure, which is covered in the analysis of what shadow AI costs when a breach lands.
The third is performance. In a June 2026 survey of 1,005 employed US workers by Software Finder, 13% admitted faking AI use outright, and only 6% believed their manager understood how the tools were actually being used.
The fourth is sabotage, which is the one that gets the headlines and is the least common of the four. It still reached 29% in the Writer study, which is a large number for a behaviour respondents had to confess to.
None of this reaches your steering committee, because every one of these behaviours is individually deniable. Nobody has to argue against the strategy. They simply do not change what they do on a Tuesday. That is why the programme reports green while the outcome drifts, and there is no escalation path for a decision that was never announced.
Your adoption dashboard is measuring the wrong thing
Most AI programmes report on seats provisioned, logins, and a monthly active user count. All three can rise while the work does not change at all.
A login is not a task completed. A monthly active user may have opened the tool once, on the day the mandate email arrived. If your reporting cannot distinguish that person from a daily user, it is not measuring adoption.
I would take the licence chart out of the steering deck entirely. It answers a procurement question, not an adoption question, and its presence lets everyone avoid the harder one.
| Metric on the deck | What it proves | What it hides |
|---|---|---|
| Seats provisioned | Finance signed a contract | Whether anyone opened the tool |
| Monthly active users | Somebody logged in once in 30 days | The difference between a daily user and a compliance login |
| Prompts submitted | The tool received input | Whether the output was used or quietly discarded |
| Training completion | A video played to the end | Whether anyone can do the task unaided afterwards |
| Satisfaction score | People will answer a survey politely | The 54% who bypass the tool and say nothing |
The honest replacement metrics are narrower and harder to game: weekly task completions inside the tool, per role, against a baseline count of how that task was completed before. Vendor-reported utilisation benchmarks circulating in 2026 are mostly unaudited, so treat any single number as directional.
Gallup's February 2026 survey of 23,717 employed US adults gives the clearest external check. 50% used AI in their role at least a few times a year, but only 13% used it daily. Only about 1 in 10 strongly agreed AI had fundamentally changed how their work gets done. That is a wide funnel with a very narrow end. The same distance between headline adoption and measurable change turns up in the review of AI deployments that returned nothing.
Why employees resist AI, and why "training" is the wrong first answer
The instinct is to explain resistance as a skills problem. Sometimes it is. More often, the employee understands the tool well enough and has decided against it.
Das Narayandas and Shunyuan Zhang at Harvard Business School frame this as an identity problem rather than a capability one. In their account, employees resist because the tool threatens who they are at work, and they name three distinct threats. This framing is a working paper argument, not a measured effect size, and I am treating it as a useful lens rather than settled evidence.
Role compression
Automation takes the interesting parts of a job first, because those are the parts with a clear output. What remains is review, exception handling and cleanup.
The employee is not imagining the downgrade. Their day genuinely got less interesting, and nobody discussed it with them beforehand.
Control shift
When a model proposes the answer, the human becomes an approver. Approval feels like less authorship than authorship, even when the final decision is identical.
This is why an override that exists on paper but is socially discouraged does not help. If overriding the model is treated as friction, control has moved and everyone knows it. The design pattern that actually preserves authority is covered in the piece on human-in-the-loop architecture.
Span erosion
Managers face their own version. If the team can ask a model instead of asking them, their influence narrows.
That matters more than it sounds, because managers are the single strongest lever you have on adoption. A manager with an unspoken reason to slow the rollout will slow it, without ever saying so.
There is a fourth reason the identity framing does not cover, and it is economic. Apollo Global Management's Torsten Slok and Sania Edlich, using Anthropic usage data across 321 occupations, reported that real wage growth for AI-exposed workers slowed by 6.7 percentage points after 2023. That analysis is observational and directional, and it is Apollo's data, not a controlled study. If workers suspect the upside of their own productivity gain lands somewhere else, quiet resistance is a rational response rather than an irrational one.
The training gap sitting underneath the resistance
Resistance and incompetence are different problems that produce the same dashboard. Both are usually present at once.
WalkMe put the cost of that at 51 working days lost per employee per year to software and AI friction. Close to half of the lost time was attributed to missing guidance rather than tools breaking. The methodology behind that figure is not published in full, so treat it as the vendor's own estimate with an obvious commercial interest attached.
Confidence is rising faster than competence
WalkMe's separate pulse survey of 2,037 US workers, run by Propeller Insights and released in August 2026, found something more uncomfortable than lying. The share who admitted pretending to know AI in a meeting fell from 45.2% to 28.3% year on year, and passing off AI work as their own fell from 48.7% to 32.5%.
Those look like wins. Then read the next two numbers. 90% of workers said they were confident using AI, while only 24.6% got the result they wanted on the first attempt. Half reported spending more time on an AI-assisted task than doing it manually would have taken.
Deception fell because it stopped being necessary. People now sincerely believe they have the skill. That is harder to fix than a lie, because the person has no reason to ask for help.
What compliance without capability produces
The employee who complies but cannot use the tool well does not simply waste their own time. They export the cost to whoever receives the work.
BetterUp Labs and Stanford's Social Media Lab named this "workslop" in Harvard Business Review in September 2025: AI output that looks finished and is not. In their survey of 1,004 full-time US desk workers, 40% said they had received workslop in the previous month, and each incident took 1 hour 51 minutes on average to resolve.
BetterUp estimated the drag at $186 per employee per month, or over $9 million a year in a 10,000-person organisation. That estimate is built on self-reported time and salary assumptions, so it is an order of magnitude rather than an audited figure. Managers reported receiving workslop more often than individual contributors, at 54% against 38.5%.
Here is the part that should change how you read your own adoption numbers. Every hour spent fixing workslop is recorded as productive AI usage on the sender's side and as ordinary work on the receiver's side. The tool shows a gain. The organisation shows a loss. Neither system is lying.
What actually moves adoption, measured rather than asserted
The strongest evidence available on this question in 2026 does not come from a survey. It comes from Microsoft studying its own rollout of command line coding agents across tens of thousands of engineers.
The paper by Murphy-Hill, Butler and Savelieva found that first use spread primarily through social networks. Retention tracked how much the engineer coded rather than who they were, and adopters merged roughly 24% more pull requests over a four-month window. The authors are explicit that a merged pull request is a proxy for output and not for value delivered.
Managers and visible peers, in that order
The peer effects are the part worth acting on. An engineer whose manager used the tool had 82% higher odds of trying it and 22% higher odds of sustained use. Where more than a quarter of an engineer's skip-level peers had adopted, the odds of trying rose 216%, and retention odds rose 66%.
If I had one lever to spend on an AI rollout, I would spend it on managers rather than on a training portal. Not because training is worthless, but because a manager who visibly uses the tool changes the default, and a training module does not.
Gallup's earlier work found that fewer than 1 in 3 employees strongly agreed their manager actively supported AI use, and that a majority reported no clear guidelines on when to use the tools. Both are cheap to fix and neither requires new software. Sequence matters more than volume here. Training delivered after the tool arrives is remedial, and remedial training carries an implicit accusation. Training delivered before the tool arrives is preparation, and it lands differently for exactly the identity reasons above.
A rollout sequence that survives contact with a workforce
This is the order I would run, and it deliberately puts the uncomfortable conversation first.
| Stage | What you actually do | Failure signal |
|---|---|---|
| 1. Name the headcount question | State in writing whether this rollout is expected to reduce roles, and over what period | Nobody asks a question in the session, which means they got their answer elsewhere |
| 2. Recharter the role | Write down which judgements stay human and become more visible, not just which tasks go | The new role description is a list of removals |
| 3. Baseline the task | Count how the task is completed today, per role, before anything changes | You cannot produce the before number six weeks later |
| 4. Start with managers | Managers use the tool for their own work for a full cycle before their teams get it | Managers hold the licence and never open it |
| 5. Make use visible | Share real prompts, real failures and real time saved inside the team, not in a newsletter | The only examples circulating come from the vendor |
| 6. Keep the override real | Log overrides, review them, and never treat an override as a compliance exception | Override rate falls to near zero within a month |
Stage 6 is the one most programmes get wrong. An override rate of zero is not adoption maturity, it is usually evidence that disagreeing with the model became socially expensive.
Stage 1 is the stage people skip, and skipping it is the single most common own goal in AI change management. If you cannot honestly say headcount is safe, say what you can commit to instead: notice periods, redeployment, or a retraining budget with a number attached. 60% of companies in the Writer survey said they planned to let go of employees who would not adopt AI. Employees are not paranoid about this. They are reading the same coverage you are, and the wider pattern is set out in the analysis of AI-attributed layoffs.
One more structural point. If your performance criteria still reward the old process, every incentive in the building argues against the rollout. Rewriting what "good work" means is slower than buying licences and matters considerably more, which is the subject of the piece on performance review criteria after AI.
Where this argument is weakest
Three things in this post would not survive a hostile review, and you should know what they are before you quote any of it.
Almost all of it is self-reported
The 54%, the 29% and the 13% are people describing their own behaviour to a survey. Self-reports of undermining your employer are subject to social desirability bias in both directions. Some people will not admit it. Some will overstate it because the question invites a grievance.
The Microsoft study and the Gallup panel are the two strongest sources here precisely because they observe behaviour or sample at scale. Where they disagree with a vendor survey, believe them.
Several sources sell the cure
WalkMe sells digital adoption software. Writer sells enterprise AI. BetterUp sells coaching. Each of them measured a problem their product addresses, and each figure is a company announcement rather than peer-reviewed work.
I have cited them because they are the only public data at this sample size, and because the findings are broadly consistent across firms with different products to sell. That consistency is the reason to take the direction seriously and the size of any single number less so.
Sometimes the resistance is correct
This is the case against the whole framing. If half of workers report that an AI-assisted task took longer than doing it manually, some proportion of "resistance" is accurate professional judgement about a tool that is not ready.
Gartner's prediction that at least 30% of generative AI projects would be abandoned after proof of concept listed poor data quality and unclear business value alongside cost. None of those are employee failures. A change management programme that treats every objection as a psychological barrier will successfully drive adoption of a tool that should have been cancelled. The readiness tests that catch that earlier are in the guide to production readiness for AI pilots.
Frequently asked questions
Why do employees resist AI tools at work?
Because the tool threatens role identity more than it threatens the task. Harvard Business School researchers describe three specific threats. Role compression removes the interesting work and leaves cleanup. Control shift moves decisions to the model, so the employee becomes an approver. Span erosion narrows a manager's influence because the team asks the model instead. Job security fears sit underneath all three.
What percentage of employees resist AI adoption?
It depends which behaviour you count. WalkMe's 2026 survey of 3,750 executives and workers found 54% bypassed company AI tools in the previous 30 days and 33% had not used AI at all. Writer and Workplace Intelligence found 29% of 2,400 knowledge workers admitted actively sabotaging AI strategy. All of these are self-reported, so treat them as directional rather than precise.
How do you measure real AI adoption rather than licence usage?
Count task completions inside the tool, per role, per week, against a baseline you recorded before the rollout. Seats provisioned, logins and monthly active users can all rise while nothing about the work changes. A monthly active user may have opened the tool once on the day the mandate email arrived, and no dashboard built on logins can tell that person apart from a daily user.
Does mandating AI use increase adoption?
Mandates raise the recorded numbers and often not the real ones. They produce compliance logins, faked usage and unsanctioned tools on personal accounts. In Microsoft's study of its own rollout, the strongest predictor of first use was visible peer adoption. Engineers whose manager used the tool had 82% higher odds of trying it, and skip-level peer adoption raised the odds 216%.
What is AI workslop and why does it matter for adoption?
Workslop is AI output that looks finished but lacks the substance to advance the task, a term coined by BetterUp Labs and Stanford's Social Media Lab in Harvard Business Review. In their survey of 1,004 US desk workers, 40% received it in a month and each incident took about 1 hour 51 minutes to resolve. The cost lands on the recipient, so your adoption metrics record it as a gain.
What should HR do first in an AI change management plan?
Answer the headcount question in writing before anything else. Employees decide what a rollout means for them within days, and they will use outside coverage if you do not tell them. If you cannot promise roles are safe, commit to something specific instead: notice periods, redeployment routes, or a retraining budget with an actual number on it.
Where to start this week
Pick your largest AI deployment and do three things, none of which needs a budget approval.
First, ask your reporting team for weekly task completions per role rather than active users. If that query cannot be run, you have found the reason the programme looks healthier than it feels.
Second, check whether the managers in the highest-usage team hold licences and use them. Compare that against the lowest-usage team. In most organisations, the answer is visible in an afternoon and explains more than any survey will.
Third, write one paragraph stating what this rollout means for headcount, get it approved, and send it. If you cannot get that paragraph approved, that is your finding, and every adoption number after it should be read in that light.
Related on rollout risk
Adoption is only half the failure surface. The other half is the pilot itself, covered in the breakdown of how agent pilots fail and in the review of deployments that returned nothing.
References
- WalkMe, State of Digital Adoption 2026, April 2026. 3,750 enterprise executives and employees, 14 countries. Used for bypass, shadow AI, trust gap and lost workdays.
- Fortune, Nearly a third of workers admit to sabotaging their company's AI, 30 July 2026. Used for the Writer and Workplace Intelligence figures, Software Finder figures and the Apollo wage analysis.
- Fortune, Workers are lying less about AI but they have developed a worse habit, 25 August 2026. Used for the WalkMe pulse survey of 2,037 US workers.
- Gallup, Rising AI adoption spurs workforce changes, April 2026. 23,717 employed US adults, fielded 4 to 19 February 2026. Used for frequency of use and transformation figures.
- Harvard Business School Working Knowledge, Why employees resist AI and how companies can win them over, on work by Das Narayandas and Shunyuan Zhang. Used for the three identity threats.
- Niederhoffer, Kellerman, Lee, Liebscher, Rapuano and Hancock, AI-generated workslop is destroying productivity, Harvard Business Review, 22 September 2025. Used with BetterUp Labs data for all workslop figures.
- Murphy-Hill, Butler and Savelieva, Adoption and impact of command-line AI coding agents at Microsoft, arXiv, July 2026. Used for peer effects and the pull request figure.
- Gartner, 30% of generative AI projects will be abandoned after proof of concept, 29 July 2024. Used for abandonment causes.
Weakest thing about this source base: four of the eight references are commercial surveys published by firms that sell adoption, AI or coaching products, and none of the resistance figures are independently audited. The Gallup panel and the Microsoft study are the only sources here that observe behaviour at scale rather than asking people to describe it.
Related reading