The 29% Are Right
Twenty-nine percent of employees admit to sabotaging their company’s AI strategy.
Not quietly ignoring it. Actively undermining it.
That number comes from Writer’s 2026 enterprise AI survey — 800 executives and 800 employees across large organizations. The same study found that 35% of executives can’t immediately “pull the plug” on a rogue agent. And 36% have no formal supervision plan for the agents they’ve already deployed.
Every consultant who sees these numbers calls it a change management problem.
They’re wrong.
The standard reading
The change management framing goes like this: employees resist what they don’t understand. The solution is communication, training, visible wins, executive sponsorship. Get people on board, and sabotage rates drop.
This framing is wrong because it assumes sabotage is irrational.
It isn’t.
Twenty-nine percent of employees are sabotaging AI because they are, in most cases, making the correct local decision. The agent they were given doesn’t understand what they’re actually trying to do. It acts without legible intent. It opens loops it doesn’t close. It makes commitments on their behalf that they then have to clean up.
When a system creates more coordination work than it eliminates, routing around it is the smart move.
The 29% aren’t wrong about the system. The system is wrong.
What employees are actually doing
When there’s no coordination layer above AI — no system that tracks what agents have committed to, what’s been closed, what still needs attention — someone has to do that tracking.
And right now, that someone is the employee.
They’re running the reconciliation pass in their head. Checking what the agent did against what was actually needed. Intercepting the outputs before they reach stakeholders. Adding the context the agent doesn’t have. Catching the errors before they propagate.
This is not change resistance. This is coordination labor.
The people labeling this as sabotage are looking at the behavior. What they should be looking at is what it’s replacing. Employees didn’t suddenly become hostile to technology. They became the default coordination layer for a system that was deployed without one.
The structure of the problem
McKinsey’s 2026 State of AI Trust report found something that makes the sabotage data legible: trust in autonomous AI agents is declining — not increasing — despite record adoption rates.
More deployment. Less trust.
The pattern makes sense once you drop the change management frame. Trust in a system doesn’t come from familiarity with it. It comes from legibility — your ability to understand what the system is doing, why, and what happens if something goes wrong.
Can I see what the agent committed to? Can I tell if it closed the loop or left it open? Can I intervene before the downstream damage reaches a stakeholder? Can I, as the executive puts it, pull the plug if I need to?
Thirty-six percent of organizations have no formal supervision plan. Thirty-five percent can’t stop a rogue agent quickly. In that environment, the only rational trust response is the one 29% of employees already chose: keep a hand on the wheel, because nobody else has one.
What this means architecturally
There’s a version of this problem that’s solvable with better change management. Clearer communication. More training. Better rollout plans.
That version isn’t the one we have.
What we have is a deployment pattern where agentic AI gets added to existing workflows without anything above it — no layer that tracks obligations, surfaces context, identifies conflicts, or makes agent behavior legible to the humans responsible for the work.
You’ve given people more capability without giving them more visibility.
The result is predictable. Employees compensate. Some compensate by doing extra coordination work. Some compensate by routing around the system entirely. The 29% who sabotage are the ones who’ve decided the compensation cost is higher than the capability gain.
They’re doing the math correctly.
The coordination layer as trust infrastructure
Here’s the reframe that matters.
A coordination layer — a system that knows what agents have committed to, what’s been done, what’s still open, and what the person actually needs next — doesn’t just improve efficiency.
It earns trust.
Not by persuading people that AI is good. Not by running change management programs. By doing the one thing that makes any system trustworthy: making it visible, predictable, and recoverable.
Visible: you can see what the agent did and why. Predictable: the agent acts in a way that’s consistent with what you actually need. Recoverable: when something goes wrong, you can catch it and fix it before it compounds.
None of the agents deployed in 2025 are missing capability. They’re missing the layer that makes their capability safe to use at full autonomy.
That’s not a change management problem.
That’s an architecture problem.
And the 29% figured it out before the industry did.
Writer 2026 Enterprise AI Survey: 800 executives + 800 employees across organizations with 500+ employees. McKinsey State of AI Trust 2026: agentic AI adoption vs. governance gap analysis.
