Companies are rethinking AI rollouts as costs and errors mount

The rollout is creating work of its own

Meta is rebuilding parts of its management ranks after cuts in 2023, particularly among middle managers, according to employees’ LinkedIn posts. Earlier this year, the company also paused tracking employees’ computer activity for AI training. It has told workers that using AI tools will no longer count toward performance reviews, The Information reported.

At Microsoft, executive vice president Ryan Roslansky recently described a familiar failure mode: AI-generated material that adds little. After receiving a document that appeared to have been prepared by AI, he asked his system to summarize it and learned nothing new. He warned of a “closed loop” in which AI reads and writes, leaving an “ocean of sameness.”

Walmart workers have reported a more operational version of the problem. Business Insider said AI tools used to assign tasks, make schedules and track performance produced unrealistic demands, errors and false alerts.

The broader cost is not just bad output. Errors consume employees’ time, increase privacy and security risks, and add expense. Microsoft and McKinsey have restricted employee use of AI amid the volume of new tools and their resource demands. Reuters reported that Palantir, Booz Allen Hamilton and Nvidia have also imposed restrictions over concerns about corporate data security.

The productivity case is still unproven

The scale of investment has not yet translated into clear gains. In a working paper from the Federal Reserve Bank of Atlanta, 90% of executives said AI adoption efforts had not yet increased productivity. In an Ernst & Young survey published last month, 36% of AI leaders at companies with annual revenue of at least $1 billion reported AI-related incidents. The consequences included data loss, financial and reputational damage, and operational disruption.

Training is another weak point. More than half of workers surveyed by Indeed said they were not getting the AI training they needed. Most executives, meanwhile, said their employees understood the technology better than they did; only 45% of managers felt ready to lead employees who grew up in the AI era.

Some companies are replacing one-off webinars with longer, role-specific training, said Johnny Taylor Jr., president and chief executive of the Society for Human Resource Management. Microsoft says it has also learned that adding agents to a poorly designed process will not fix it: speeding up one stage can simply create a longer queue at the next. In a recent blog post, Kathleen Hogan, Microsoft’s chief strategy and transformation officer, wrote that access and usage alone do not produce transformation. The company said it improved productivity by focusing on training, including pairing junior engineers with more experienced colleagues.

The reliability problem is harder to train away. AI can produce plausible-looking answers with flaws that are not immediately obvious, said Jessica Reif, an associate professor of management at the University of Pennsylvania’s Wharton School. That makes errors more likely to travel through a workflow and accumulate before anyone catches them.

When AI becomes evidence

The risks can extend beyond internal inefficiency. Lawyer Neil Elan of Stubbs Alderton & Markiles LLP sees a growing gap between how corporate clients imagine AI being used and how employees actually use it, with consequences for privacy, intellectual property, efficiency and workplace conduct.

A case involving 3M shows how visible that gap can become. Before trial over a 2020 gas explosion at a Houston manufacturing facility that killed three people and damaged hundreds of homes and businesses, lawyer Will Moye reviewed an independent expert report commissioned by 3M. He noticed unusual formatting and timestamps suggesting that 14,500 pages had been prepared within hours. Moye suspected the author, Josh Autenrith of Knighthawk Engineering, lacked the necessary qualifications and had relied on ChatGPT, despite billing 3M $90,000 for a report concluding that the company was not responsible.

Court filings included Autenrith’s ChatGPT logs. After uploading a photograph of the gas detector at the center of the case, he asked the chatbot, “What am I looking at?”

Jurors found that 3M had failed to properly test the gas detection system and ordered it to pay $61.5 million to 24 homeowners. A 3M representative said Autenrith independently examined the facts and reached his own conclusions, using AI as a tool while preparing the report. His lawyer, Janina Wojtkowski, said ChatGPT did not generate his conclusions or replace his professional judgment; it helped with formatting, structure and drafting under tight deadlines.

I think the more important signal is not that AI appeared in one disputed expert report. It is that the tools can make weak work look finished, while the responsibility for checking it stays with people. The announcement from companies pulling back says little about how they will measure that cost—or who will be accountable when an AI-assisted mistake survives review.

The promised gains have not arrived

A May study by University of Pittsburgh researchers found no productivity increase after companies invested in AI and cut staff. Its authors argued that management optimism had not translated into better results. Challenger, Gray & Christmas said U.S. employers had announced more than 112,000 job cuts citing AI since the start of 2026.

Pittsburgh professor and AI-layoff researcher Mark Ma said executives focused on short-term results and moved too quickly to invest in AI and lay off workers. He also said company shares often fell after AI-related layoff announcements, suggesting investors were not convinced those decisions would succeed.

“Many mistakes could have been avoided,” Ma said.

Companies now face a tension their early adoption plans did not resolve: AI can accelerate parts of a job, but unreliable output, weak training and poorly designed processes can shift the work rather than remove it. Until the gains show up beyond the announcement, pulling back may be less a reversal than a test of whether the original case was ever there.