Workers Feel AI-Fluent—but Half Say It Has Wasted Their Time

WalkMe published its third annual AI at Work Pulse Survey on August 25, 2026, reporting a wide gap between confidence and results among U.S. employees who use AI at work. Of 2,037 respondents, 90% said they felt confident using AI, but only 24.6% said it worked on the first try; 50.2% said they had spent longer getting AI to complete a task than manual completion would have taken, findings also covered by Fortune on August 25.

The August 25 announcement describes self-reported experiences, not the results of a skills test or controlled productivity trial. Propeller Insights fielded the WalkMe survey from June 10 through June 20 among working U.S. adults who already used AI on the job; 51% also said AI had led their manager or team to expect more output in the same amount of time.

What the survey measured—and what it did not

The 90% figure measures perceived confidence, not verified proficiency. A worker can be comfortable choosing a tool, writing prompts and interpreting an answer while still needing several attempts before the output is usable.

The 24.6% first-try figure addresses a different question: whether respondents believed AI worked on its initial attempt. The published methodology does not define a common acceptance standard or separate results by model, application, occupation, task complexity or consequence. A rough draft, a calculation and a high-impact recommendation therefore sit inside the same broad measure despite requiring different levels of reliability.

The time-loss result is narrower than a finding that AI reduces productivity overall. It establishes that half of these respondents had experienced a task on which using AI took longer than doing the work manually; it does not reveal how frequently that happened, how much time was lost or whether savings on other tasks outweighed the delay.

The sample also excludes workers who do not use AI on the job. All three headline figures—confidence, first-pass success and time lost—should be read as self-reported evidence from current workplace AI users, not as observed performance across the entire U.S. workforce.

Five signals separate adoption from performance

The findings become more useful when the measures are kept separate. An adoption rate can show that a tool is present in a workflow, but it cannot establish whether the first output was usable, whether corrections erased the time saved or whether employees trusted the result when the stakes were high.

• Confidence: whether employees believe they can use AI comfortably. In this survey, 90% said they did.
• First-pass reliability: whether the initial output works without another attempt. Only 24.6% said it did.
• Rework and net time: the prompting, correction and verification required before completion. The 50.2% result records an experienced delay, not total time saved or lost across all work.
• Contextual trust: whether an output reflects the relevant business rules, permissions and consequences rather than merely sounding plausible.
• Output pressure: whether the availability of AI changes expectations independently of its reliability. Here, 51% reported pressure to produce more in the same time.

Rework and net time are not interchangeable. Several prompts may still produce a faster result than manual work, while a polished first response can create hidden labor if every factual claim requires checking. A defensible productivity measure would combine elapsed time, correction time, downstream errors and validated final quality.

A separate global study illustrates why workflow friction belongs in that calculation. WalkMe’s State of Digital Adoption 2026, based on 3,750 enterprise workers and executives, estimates that software and AI friction consumes 51 workdays per employee annually. That estimate comes from a different international sample and cannot be merged with the August U.S. pulse survey.

Rising expectations complicate the confidence result

The reported confidence gap exists inside workplaces where AI can raise the expected volume of output before its net value has been established. Employees may keep using a tool because it is available or expected even when prompting, correction and verification consume the anticipated savings. High use and high confidence can therefore coexist with weak first-pass reliability.

The same survey found that 33% of respondents said they had pretended to be more skilled with AI than they were, while 32.5% said they had passed off AI-generated work as their own. On questions tracked from 2025, however, the share saying they had pretended to understand AI in a meeting fell from 45.2% to 28.3%, and the share passing off AI work as their own declined from 48.7% to 32.5%.

Those changes do not identify a cause. They could reflect greater skill, more normalized use, different interpretations of what counts as AI-assisted work or changes in willingness to disclose the behavior. The survey did not independently test which explanation is correct.

Technical and organizational context also affects performance. SAP and Oxford Economics research identifies data readiness, governance, workforce transformation and responsible scaling as continuing challenges as businesses embed AI in core operations. That supports treating poor results as a possible workflow or system problem, rather than assuming every gap can be fixed with more employee training.

What employers still cannot infer

The survey does not show which tools or tasks generate the most rework, whether first-pass failures are minor drafting issues or consequential errors, or whether AI produces a net productivity gain over a longer period. It also cannot establish that confidence causes poor performance; the two measures were reported together, but no causal test was conducted.

Before interpreting the gap as a training deficit, managers would need to distinguish among operational skill, missing business context and workflow friction. Useful questions include whether the first output met a predefined acceptance standard, how much correction and verification followed, whether the system had access to the required context and whether the validated result was faster than the manual alternative.

For now, the confirmed finding is narrower than a national productivity verdict: current U.S. workplace AI users report widespread confidence, but far fewer report dependable first attempts, and half have encountered at least one task on which AI cost them time. Observed task results, consistent quality standards and measured rework would be needed to determine whether that confidence translates into reliable workplace performance.

Also read:

• Data Science Is Set to Grow 34%—but the Job Is Broader Than Coding
• AI-Generated CVs Flood Hiring Queues, but the Real Loss Is Trust
• LinkedIn’s Creator Marketplace Is Live—but Access Still Has Hard Limits