Special Offer

Get 3 months free of PEO*

articleIcon-icon

Article

8 min read

Companies are measuring AI adoption. They still don’t know if it’s working.

Kaila Caldwell

Author

Kaila Caldwell

Published

November 13, 2026

Photo of woman at computer

Companies are measuring AI adoption. They still don’t know if it’s working.

Employer-sponsored AI tool subscriptions rose from 16% to 33% in SHRM's 2026 Employee Benefits Survey, the largest increase of any benefit it tracked. Over the same period, formal training to develop new skills and training to maintain existing skills both fell seven percentage points.

A 2026 National Bureau of Economic Research study of nearly 6,000 senior executives across the US, UK, Germany, and Australia found that 69% of firms actively use AI, while 89% of executives reported no impact on labor productivity over the previous three years.

Across 88,753 assessments, only 11% of employees accurately judged their own AI proficiency, according to Workera's 2026 AI Skills Enterprise Benchmark. Yet 85% of L&D leaders said they trust self-reported skills data.

Developers using AI were 19% slower in a randomized METR study of 16 experienced open-source developers completing 246 real tasks in codebases they knew well. They had expected AI to make them 24% faster and, even afterward, still believed it had made them about 20% faster.

Companies can measure training completion, AI use, and self-reported proficiency. Those measures don't establish whether comparable work gets done faster, whether output increases, or whether it requires less rework.

The tools and the training were bought separately

37.8% of US knowledge workers have received no AI training at all, according to Section's June 2026 AI Proficiency Report, a survey of 5,026 workers. Among those who have, 15% were trained only on safety and compliance.

Martin*, a veteran IT project manager at a global chemical manufacturer, has access to AI but has largely been left to figure out how to use it in his own work. His employer's guidance is "use it if it helps, just check the output and don't leak anything confidential or secret," which he calls "technically advice and practically a shrug."

"No one has ever walked me through a real task from my actual job and shown me where the tool tends to go wrong," Martin says. "It's been general awareness material, theoretical and tidy, which is a bit like teaching someone to drive using only a pamphlet about traffic laws."

"There is definitely a sense that AI is now something you're supposed to understand," Martin says. "So admitting that you're not particularly good at using it can feel a bit like admitting you're behind the times. I'll admit that includes me."

Use it first, learn later

21% of employees are required to use AI for core job responsibilities, and another 45% are encouraged or expected to use it, according to Study.com's 2026 State of AI Jobs and Skills Report, a survey of 1,000 workers. Yet among workers who received training, only 18% say it prepared them to work independently.

Alex*, who worked at Meta until the company's May 2026 layoffs, says AI use went from encouraged to tracked and mandatory within months. By late 2025, employees were expected to use AI at least once a day, four days a week. "There were tools that showed us how AI native we were," he says. "There were reports that compared performance across individual employees and teams." Employees were later told AI use would become part of the 2026 performance review cycle. No formal training came with the new metric.

At first, Alex says, "we were just using AI to use it, to check the box. I remember someone telling us to use it first thing in the morning to get it out of the way."

"Not all of us knew how to use AI in a meaningful way and not every job lends itself for AI automation," he says. "We knew that eventually we had to use it in a way that we could defend and hopefully show results, come performance review time."

At Founders, Inc., Adrianna Lakatos solved a problem Meta employees faced: finding where AI was actually useful in the job. She used AI to build an internal application her team still uses, but had no formal training in how to verify the work. At first, she says, checking AI output meant asking, "does this seem to work?"

Her AI use became part of a business process before there was a clear standard for how that work should be checked, maintained, or handled when something went wrong. When the app malfunctioned, Lakatos had to develop those standards through experience. "Once people rely on what you make, you're also the person dealing with it when it breaks."

What AI productivity numbers miss

47% of CHROs haven’t established clear productivity measures for AI, according to CHRO Association's 2026 CHRO Survey of roughly 150 CHROs at major corporations.

Joy Papini, an organizational psychologist and president and founder of Cammino Insights who has spent four years helping large organizations implement AI, cautions against treating skill gains as productivity gains. "Skills are an input to productivity, not productivity itself," Papini says. "You should expect a bigger jump in skills than in productivity, because productivity also depends on the opportunity and motivation to use those skills on the job."

Martin used an AI agent to pull technical network data from several dozen PDFs in about two hours, work he estimates would have taken 20 hours by hand. But the AI merged two different network switches into a single entry, and the error looked plausible. "It wasn't obviously broken, it was convincingly broken," Martin says.

Reviewing the remaining output took about 15 hours. "Checking an AI result is not the same thing as proofreading it," Martin says. "You have to know enough about the subject to challenge the answer." Once the review time was included, the job took roughly 17 hours instead of 20. The AI-assisted step suggested a 90% time saving; the full workflow reduced the time by about 15%.

At HostZealot, founder and CEO Daniel Yeromka tracked comparable infrastructure and support tickets across 40 to 50 employees for six months. Median resolution time fell from 45 minutes to 38, a 15.6% reduction. In surveys, employees estimated AI was saving them closer to 30%.

Early gains were larger because engineers used AI mostly for routine requests. As they moved to more complex infrastructure tickets, "the generation time remained small and the verification time increased," Yeromka says. "An AI assistant provides an answer in two minutes, and the engineer spends an additional 18 minutes verifying it. That counts as a 20-minute workflow, not a two-minute solution. I'd advise them not to measure what the AI does, but what happens to the work." HostZealot later standardized verification and created checklists for AI-generated code and configurations.

When AI saves one worker time and costs another

AI can save the person producing the work time while adding work for the person reviewing it. About 40% of 1,150 full-time US desk workers received low-quality AI-generated work from a colleague in the previous month, according to research from BetterUp Labs and Stanford Social Media Lab. Recipients spent nearly two hours dealing with each incident. About half viewed the colleague who sent the work as less capable or reliable, and 42% as less trustworthy.

At Whitepage, founder and CEO Tanya Slyvkin received a 12-page competitor analysis from a direct report who had generated it with ChatGPT without supplying source documents. It listed a discontinued $2,500-a-month price as current and included a market statistic Slyvkin says "didn't exist." "Honestly, I lost confidence in all of it considering she left everything to ChatGPT," Slyvkin says. A properly researched analysis would normally take her 30 to 45 minutes to review. Instead, she spent about three hours "going back to square one, conducting the primary research myself, and rewriting the text."

At SOLO MEDIA GROUP, founder Johnathan Solorzano asked an employee to explain the integration section of a client proposal. He couldn't. The employee had used AI for a first draft without reviewing it, leaving Solorzano with two to three hours of corrections.

"[The employee] thought that was what a first draft was for. That's a fair thing to think. We'd never said otherwise," Solorzano says. He now requires employees sending client-facing work to explain every claim and identify its source. "The person knew how to use the tool. What they didn't have was a standard for what 'finished' means when generating a draft costs nothing."

The work is the test

Employee surveys can capture whether an AI tool feels useful or frustrating, but not whether it improves productivity, Papini says. "Self-reporting isn't unreliable; it's answering a different question," she says. "Asking someone to rate their productivity is really measuring self-efficacy, not productivity itself." Instead, she recommends identifying "a specific point in the workflow that's observable and quantifiable, and directly valuable to the business," establishing a baseline before AI is introduced and measuring it again after employees have had time to adapt.

Daniel W. Rasmus, founder and principal analyst at Serious Insights, says companies need to compare the work before and after AI, including work the technology adds. "What got eliminated? What got transformed? What new work was generated by the new process?" he says. "AI use is a signal. AI value is an argument that must be proved in the work."

At HasData, CEO and CTO Roman Milyushkevich compared similar engineering tasks before and after AI adoption, tracking active work time, review rounds, reopened tasks, and corrective work. "The harder part was defining what counted as a comparable task," he says.

Active coding time fell from eight hours to five and a half, but once review and reopened work were included, the overall improvement was 12.5%. "We did not treat faster code production as the final productivity result," Milyushkevich says. "The real unit of measurement should be completed work that survives review."

Evgenii Arsentev, PhD, an AI transformation executive at ARSENTEV.AI, says volume metrics need a "failure twin." "Tickets closed, code commits and lines changed, documents produced, and marketing content published. All four rise almost automatically with AI," he says. Pair them with reopened tickets, defects found after release, documents rewritten by someone else, and content that gets no response. "AI makes production cheap, so the scarce work moves to verification and judgment." Arsentev recommends tracking rework and its cost, downstream errors, time spent verifying AI output, and the outcome the work was meant to produce.

Michael Podgortsev, an independent researcher and former CTO focused on measurement bias and human-AI fit, applies three questions to claims of time saved: How was the saving measured? What is it being compared with? And where did the saved hours go? If an employee says AI saves five hours a week but output hasn’t changed, Podgortsev says, "the time may have felt saved without being saved."

Martin's network-data project carried another risk. Had he not checked the source documents, the merged switch entry could have entered internal reference material used by other employees. "AI can take some of the work away," he says. "It doesn't take the responsibility away."

*Names have been changed at the sources' request.

Table of Contents
Kaila Caldwell

About the author

Kaila Caldwell is a freelance journalist contributing to Deel Works, where she reports on workforce trends, management, and the future of talent. Her work combines original reporting, expert interviews, and primary data to produce long-form features for business leaders and decision-makers worldwide. Before Deel Works, Kaila spent several years as an editor and journalist covering the future of work, AI, workforce transformation, economics, and sustainable finance. She has lived and worked in the US, France, and Tunisia, and is currently based in Washington, D.C. Connect with her on [LinkedIn](https://www.linkedin.com/in/kaila-caldwell/).