Companies Slow Workplace AI Rollouts After Errors and Weak Gains

What changed

Companies that rushed to deploy AI are slowing down, tightening access and rebuilding human oversight after errors, security concerns, weak productivity gains and rising costs. Meta has scaled back plans to eliminate staff through AI agents while expanding managerial ranks again; Microsoft is warning about low-quality AI-generated work; and Walmart workers have reported unreasonable demands, errors and false alerts from AI scheduling and task tools.

Why it matters

The first bill for workplace AI is arriving as cleanup. Employees are spending time correcting technical and security mistakes, managers are being asked to improve productivity without sacrificing quality, and companies are paying for tools whose benefits remain unproven.

The numbers explain the change in mood. An Atlanta Federal Reserve working paper found that 90% of executives said their AI efforts had not yet produced productivity gains. An Ernst & Young survey found that 36% of senior AI executives at companies with at least $1bn in annual revenue reported an AI incident involving problems such as data loss, financial damage, brand damage or operational disruption.

That makes “deploy first, govern later” an expensive operating model. Some companies, including Microsoft, McKinsey, Palantir, Booz Allen Hamilton and Nvidia, are limiting employee use or tightening security rules. Meta’s return to managerial hiring points to a broader lesson: when automated systems create more exceptions, someone experienced still has to make sense of them.

The skills gap is making the transition harder. More than half of employees surveyed said they were not getting the AI training they needed, while only 45% of managers said they felt equipped to lead AI-native staff. Access to a tool is not the same as knowing where it belongs, how to check its work or when to stop using it.

The pullback may be temporary or broad, and the report does not identify which deployments will survive. But the direction is clear: AI is being treated less like a software switch and more like a system that needs supervision, limits and proof before it earns a larger role.