Companies rushed in on AI. Now some are pulling back

Over the past few years, companies embraced AI through "rapid introduction and broad experimentation." But according to this analysis in The Guardian, some companies that got off to a fast start are now hitting the brakes: imposing restrictions where experimentation was once encouraged, recalling experienced employees and increasing training. This shift is not an isolated phenomenon, but a problem faced jointly by different types of organizations, from Meta and Microsoft to Walmart and McKinsey.

The most direct signal comes from Meta. According to Reuters, after finding that employees were spending increasing amounts of time cleaning up technical and security errors, Meta scaled back its plan to cut jobs and replace employees with more AI agents. At the same time, after cutting staff (especially middle managers) in its 2023 efficiency drive, Meta is once again expanding its manager ranks, a sign coming from employees' LinkedIn posts.

It should be noted that the specific scale of Meta's scaled-back plan and whether it has formally taken effect remain unclear; Meta also did not respond to a request for comment. But the path itself—"cut people first, then fill in with AI agents, then hire managers back again"—has already exposed the gap between the assumption of rapid replacement and real operations.

One direct reason companies are contracting usage is that errors and low-quality work produced by AI are consuming employee time, raising privacy and security risks and driving up costs. Microsoft Executive Vice President Ryan Roslansky publicly complained about the flood of AI slop in daily work: he received a document "clearly generated by AI," and after summarizing it with AI got no new thinking, and warned of a possible "doom loop"—AI does the reading and writing, leaving behind "a sea of sameness."

Walmart's situation is closer to the front line. According to Business Insider, Walmart store and warehouse employees are troubled by unreasonable expectations, errors and false alarms from AI tools that handle task assignment, scheduling and performance monitoring. Employees described robots plotting absurd routes for pickers, making it harder rather than easier to complete orders quickly; sending useless safety alerts; and setting unrealistic schedules that force employees to cut corners to keep up. One manager told Business Insider that these robots assume employees will "do everything perfectly every time," and said, "We wouldn't hold people to impossible standards."

Walmart declined to comment, and the prevalence of the AI tool problems and the scale of affected employees are unknown. But these descriptions point to a common mechanism: when AI is used to evaluate and schedule human work, its errors are not ignored—they are converted into unreasonable pressure on employees.

Jessica Reif, an assistant professor of management at the University of Pennsylvania's Wharton School, explained that AI tends to produce content that "seems plausible" but whose weaknesses or errors do not immediately show, making it more likely that errors spread through the process rather than being immediately corrected, and the problem can snowball. This judgment explains why "cleaning up errors" takes up increasing amounts of employee time—errors do not erupt in a concentrated way, but are dispersed and embedded in the workflow.

Microsoft's practice confirms this. In a blog post, Microsoft Chief Strategy and Transformation Officer Kathleen Hogan said the company initially promoted AI as a traditional technology: deploy tools, provide training, drive adoption, but "access and use do not equal transformation." Microsoft also found that "adding agents to a flawed process leaves the process flawed," and "accelerating one step only creates a longer queue at the next."

In other words, AI's own speed does not automatically translate into organizational efficiency. If the process itself is flawed, AI merely pushes the bottleneck from one place to another, while adding steps that require human review and error correction.

Although global companies have invested billions of dollars over the past few years, the main returns remain to be seen. According to an Atlanta Fed working paper, 90% of executives say their companies' AI efforts have not yet translated into productivity gains. This figure is consistent with companies' intuitive sense: tools are widely deployed, but improvement on the output side is not obvious.

Risks are also accumulating. A survey released by EY last month found that among companies with annual revenue of at least $1 billion, 36% of senior AI executives reported an AI incident that caused problems such as data loss, financial loss, brand damage or operational disruption. This shows that AI brings not only efficiency issues, but also governance and security issues.

It should be pointed out that the specific methods and samples of the Atlanta Fed working paper and the EY survey are not explained in the source, nor does the source provide positive quantitative data on AI productivity gains. Therefore, these figures should be understood as descriptions of the current situation, not as a final judgment on AI's long-term potential.

Facing errors, costs and security concerns, companies are shifting from "encouraging experimentation" to "imposing restrictions." According to Reuters, companies from Microsoft to McKinsey are restricting employee use of AI, while companies such as Palantir, Booz Allen and Nvidia are restricting employee use out of concern for the security of organizational data at AI companies. The source does not explain the specific policy content and scope of enforcement of each company's restrictions; Microsoft and Walmart declined to comment, and Meta, Palantir, Nvidia and Booz Allen did not respond to requests for comment.

At the same time, the training gap is very obvious. According to a survey released by Indeed in July, more than half of employees say they "have not received the AI training they need," while most managers say they have been surpassed by employees on AI, and only 45% of managers feel "capable of effectively leading AI natives."

Johnny C Taylor Jr, president and CEO of the Society for Human Resource Management, said executives feel that "we're building the truck while driving it, and that's the problem," and said that if he could do it over, he would introduce AI in a more cautious and slower way. Companies are moving from one-off webinars for large groups of employees to "continuous, role-specific training." Microsoft found that focusing on employee training—especially pairing early-career engineers with more experienced employees—improved productivity.

The risks of AI do not stay only in internal processes. Neil Elan, an attorney at Stubbs Alderton & Markiles LLP, said he has observed an "ever-growing gap" between how corporate clients believe AI is being used and how it is actually used, giving rise to issues involving confidentiality, intellectual property, efficiency and employee behavior.

The 3M case is a concrete example. On the eve of 3M's trial over a 2020 gas explosion at a Houston plant, attorney Will Moye discovered signs of AI in the report by the opposing independent expert witness Josh Autenrieth: abnormal formatting and fonts, and timestamps showing that 14,500 pages were assembled within hours. Chat records showed that Autenrieth had uploaded a photo of a gas detector to ChatGPT and asked "What am I looking at?", and he billed 3M $90,000 for this report asserting that 3M bore no liability.

The jury ultimately found that 3M failed to properly inspect its gas detection system, and awarded $61.5 million to 24 homeowners. Moye said that after the verdict, the jury "expressed deep disbelief" that the company would rely on such a report. It should be noted that the source cannot determine the actual degree of influence the AI-generated report had on the verdict, but the case at least shows that when AI-generated content enters high-risk scenarios such as law and compliance, its "plausible" appearance can bring serious consequences.

Putting the above facts together, a clear logical chain can be seen: companies rapidly introduced AI tools, expecting to gain productivity improvements directly; but AI produces "plausible" content whose errors are not easy to spot immediately, and instead spread through the process and consume employee time; when AI is used in human-related scenarios such as scheduling and performance monitoring, errors are converted into unreasonable pressure; at the same time, data security and compliance risks rise, leading companies to restrict use, recall experienced employees and increase training.

Whether this shift is a short-term adjustment or a long-term trend, the source does not give a judgment. The scale, timetable or quantified impact of companies' contraction in AI use is also not disclosed. But what can be certain is that Microsoft's self-summary is representative: "Access and use do not equal transformation." The value of AI does not depend on how many tools have been deployed, but on whether the processes it is embedded in are sound, whether employees receive sufficient training, and whether governance can keep up.

For companies that are introducing or planning to introduce AI, these facts point to a more cautious path: fix the process first, then accelerate; train people first, then deploy agents; clarify data and compliance boundaries first, then expand the scope of use. Otherwise, what AI brings may not be productivity, but more errors that need to be cleaned up.