We Automated the Answer, Not the Responsibility

Based on a Sept. 2, 2026 Kolmogorov Law survey release. Reporting note: the survey measured respondents’ beliefs about legal responsibility; it did not determine actual liability in individual cases.

AI’s most consequential workplace mistake may not be producing a wrong answer. It may be producing a wrong answer that looks finished.

A national survey released by California business litigation firm Kolmogorov Law on September 2 found that 65% of 500 employed U.S. adults who use AI for work said they do not always independently verify an AI answer before acting on it or passing it along. More strikingly, 42% said they had accepted an answer they suspected was wrong because using it was faster or because the answer sounded confident. Thirty percent said a wrong or inaccurate AI answer had already caused a problem in their work, ranging from a mistake in a deliverable or a bad decision to lost time, money, or a client or legal issue.

Those numbers describe something more complicated than ordinary carelessness. AI has entered many workplaces as a productivity tool, but the surrounding work system has not always been redesigned around it. Employees can be encouraged to use AI, praised for moving faster, and quietly expected to exercise professional judgment about when the output is trustworthy. The organization gets the benefit of acceleration. The employee inherits the checking. When the result is wrong, everyone suddenly remembers that a human was still the one who clicked “send.”

That is the real gap in the survey: not machine accuracy, but the distance between tool use and responsibility.

A belief about liability is not a legal ruling

The survey also produced a headline-friendly liability result, but it deserves careful wording. Asked who they thought would be legally responsible if they acted on a wrong AI answer at work and it caused financial harm, 51% chose “me personally,” while 18% chose their employer and 12% chose the company that made the AI. That measures workers’ beliefs about liability; it does not establish that every employee would, in fact, be personally liable for every AI-assisted mistake. Actual responsibility can depend on the facts, the type of claim, the worker’s role, contractual obligations, professional duties, and state law. Under the long-standing doctrine of respondeat superior, for example, employers can be legally responsible for wrongful acts committed by employees within the scope of employment, and claims may involve both employer and employee.

The distinction matters because the survey is revealing even without turning a perception question into a legal conclusion. A majority of respondents already believe accountability does not disappear just because software helped generate the answer. They are using AI with some awareness that a human may still have to answer for the result. Yet that awareness is not reliably producing verification behavior.

In fact, the survey contains an even stranger contradiction. Among the 173 respondents who said they always verify AI output, 40% also said they had accepted an answer they suspected was wrong, and that group reported work problems from bad AI answers at roughly the same rate as the full sample. Self-reported behavior is messy, and a 500-person online survey should not be mistaken for a census of the American workplace. But the inconsistency is useful because it suggests that “I check” may not describe one standardized activity. For one worker, checking may mean opening a source. For another, it may mean asking the chatbot a second time. For someone else, it may mean reading the answer once and deciding that it feels plausible.

That is not a minor procedural detail. Verification is a task, and tasks need definitions.

Verification is work

Only 22% of respondents said their employer had a written policy requiring AI output to be verified before it entered work product. Fifty-nine percent said no such policy existed, and 19% did not know. The absence of a written rule does not prove the absence of good management, but it does show how easily organizations can introduce a new production tool without specifying the quality-control step that comes after it.

The U.S. Department of Labor’s AI Literacy Framework, released in February, treats that step as a core part of competent AI use. Its guidance says workers should be able to cross-check factual accuracy against trusted sources, identify gaps and logical errors, apply human judgment, and avoid treating AI responses as final or authoritative without review. That is a much more useful description of AI literacy than simply knowing how to write a good prompt. The prompt starts the work. The judgment determines whether the work is usable.

And judgment takes time.

A separate BambooHR study released one day before the Kolmogorov Law announcement found that 1,608 full-time salaried U.S. desk workers reported spending an average of 87 minutes per day using AI. They said 42% of that AI time went to troubleshooting errors and iterating on prompts, compared with 35% spent on work that directly advanced their workload. BambooHR annualized the troubleshooting share to roughly 20 eight-hour workdays. That calculation should not be read as proof that AI causes 20 days of net productivity loss; troubleshooting may still be faster than doing the underlying task without AI. But it does make the hidden labor visible. AI does not merely create output. It creates output that someone must evaluate, correct, contextualize, source, approve, and sometimes reject.

The invisible final mile

Put the two surveys next to each other and a recognizable workplace pattern appears. Organizations are automating the first draft faster than they are designing the final mile. The technology can compress drafting, research, summarization, or analysis, but it does not automatically create a verification standard, an escalation path, a risk threshold, or extra time in the estimate for checking. Those responsibilities often land on the individual worker as informal labor: important enough to be blamed for when skipped, but not always formal enough to be named as part of the job.

That arrangement becomes especially fragile when speed itself is the reason people are using the tool. If the value proposition is “this gets you there faster,” verification can begin to feel like giving the time back. The worker now has two competing signals: use AI to move quickly, but slow down enough to establish that the answer is correct. Without clear organizational rules, the employee has to decide which expectation matters more on every task.

The Kolmogorov Law survey found the consequences were more common among the 113 respondents who said they use AI for legal, financial, or compliance questions: 43% said a wrong answer had already caused a work problem, while 64% said they still did not always verify. The subgroup is small and carries a wider margin of error, but it illustrates why a single universal instruction such as “double-check the AI” is not enough. A lunch recommendation, a meeting-summary draft, a contract clause, and a regulatory statement should not all pass through the same review process.

Design the review, not just the prompt

A mature workplace AI policy would treat verification as risk-based work. Low-consequence drafting may need a quick human review. Factual claims may require source confirmation. Legal, financial, medical, regulatory, or client-facing material may need approved sources, documented checks, a second reviewer, or a clear escalation path. Just as important, employers need to decide where that checking time lives. If AI shortens the first 70% of a task but adds a new quality-control step at the end, that step is still part of the work. It belongs in the process, the training, and the estimate.

This is why the most useful question for employers is not simply, “Are our people using AI?” It is, “What happens between the AI answer and the moment we rely on it?” If the answer is mostly individual instinct, the organization has not automated responsibility. It has distributed responsibility without designing the system around it.

AI can draft the paragraph, summarize the file, suggest the clause, assemble the analysis, and present all of it with remarkable confidence. It can make the first move look suspiciously like the final one. But the obligation to decide whether the work is actually fit to use has not vanished. We automated the answer. We did not automate the responsibility.

Methodology note: The Kolmogorov Law survey covered 500 employed U.S. adults ages 18–64 who reported using AI for work at least occasionally. It was fielded Sept. 1, 2026 through Pollfish; results were unweighted, with a reported margin of error of approximately ±4.4 percentage points for the full sample. Subgroup results carry wider margins. The BambooHR research was a separate survey of 1,608 full-time salaried U.S. desk workers, including 520 HR professionals.

SOURCE NOTES

• Kolmogorov Law / Pollfish survey and analysis, Sept. 1, 2026
• Kolmogorov Law press release, PR Newswire, Sept. 2, 2026
• U.S. Department of Labor, Artificial Intelligence Literacy Framework, Feb. 13, 2026
• BambooHR, Redesigning Work: AI’s Performance Review, Sept. 1, 2026
• Cornell Legal Information Institute, Wex: respondeat superior