Ask a room of employees whether they check what AI tells them, then ask whether they always check, and the room gets quieter. O’Reilly put that second question to 2,000 UK employees in September, and on October 27 it published the answer: 24% say they always verify an AI-generated answer before using it.
The same survey found that, among employees who use AI-generated answers in their work, 39% have acted on one that later proved inaccurate, incomplete, or unsuitable, and 17% say it has happened more than once. Thirty-nine percent of UK workers now turn to an AI chatbot first for work answers. Thirty-five percent say their organization gave them no AI training or guidance at all in the past 12 months, and only 16% have had training specific to their role.
Responsibility for checking was never assigned in many of these workplaces. Thirty-five percent of employees got no AI guidance in a year, and the published results don’t say whether anyone told the rest which answers had to be checked, how, or by whom. Each person ends up deciding alone, task by task, with whatever time is left.
The limits of “always verify”
“Always verify” fails the first busy afternoon. An employee who checks every AI answer with the same care has given back much of the time the tool saved. One who checks nothing has handed the decision to a system that can’t be held to account for it. Most people land somewhere in between, and the published results report only the “always” figure.
What the in-between needs is a line. Some outputs can go out on the worker’s own judgment: a rewritten email or a meeting summary for internal use. Others should not move until a named person has looked and made a ruling, such as anything sent to a customer, anything touching pay, health, or legal terms, and any number that ends up in front of a board. That line is the employer’s to draw. Leaving it to each worker means it gets drawn hundreds of different ways.
Employees seem to know this. Asked what would most increase their confidence in using AI for important work decisions, the top answer in the O’Reilly survey was human review, at 39%, ahead of better training at 35% and ahead of clearer policies or governance documents alone. People ask for a second set of eyes on important decisions.
The manager is where the line gets drawn, or doesn’t
Perceptyx published a look at employee sentiment on AI on October 30, drawing on its benchmark database and its 2026 Workforce Panel of employees in North America and Europe. In that panel, 56% of employees trust their organization to use AI responsibly. Only 49% say their organization is transparent about how AI tools are used, and 46% say clear ethical guidelines are in place.
The gap Perceptyx highlights sits with managers. Sixty percent of managers feel ready to lead AI-augmented work, while 44% of employees say their manager helps their team adapt to AI-related change. Where managers do help, 82% of employees trust their organization to use AI responsibly, against 56% across the full panel. Perceptyx is careful to call these relationships rather than proof of cause, and that caution is right. Still, the team level is where an employee learns which AI work is theirs to sign off on and which goes up the line, so it’s no surprise trust forms there.
Perceptyx’s 2025 research shows a gap by rank. Executives were far more likely than individual contributors to understand how generative AI tools were chosen, 81% against 47%, and to feel comfortable raising concerns about them, 84% against 59%. The people closest to the practical problems a tool creates are the least informed about it and the least likely to speak up.
A policy on paper does not sit in the work
RGP’s AIxHuman report, built on a survey of 500 professionals who already use AI at work, put the point plainly: “AI governance cannot live only in policy. It has to live in the work.” It found that only 55% know when AI influences decisions affecting them. Its list of what workers need is a set of decision rights nobody has assigned: when AI can be used, when its involvement matters enough to disclose, what requires independent verification, which outputs demand human review, what should never be delegated to AI, and who owns the final decision.
RGP’s next line belongs on the wall: “when AI-generated work is wrong, the accountability does not belong to the AI. It belongs to the organization whose name is on the work.” One step is missing. Inside the organization, it has to belong to a person, by name, before the work goes out, not after a client finds the error.
That is the gap between a human in the loop and a human with authority. A person who happens to be near the work when AI produces it is in the loop. A person who is told this output is theirs to accept, modify, or reject, and whose ruling is recorded, holds the decision. Checkpoint-based governance, laid out in the human-AI collaboration paper, is built on that distinction. The checkpoint is a defined place in the work where a named human rules on the output, and the record shows who ruled and what they decided.
What an employer can set up this quarter
Start with one team that already uses AI daily. List the outputs it produces with AI help over two weeks. For each one, decide whether it can go out on the worker’s judgment or needs a named reviewer, and write the reviewer’s name next to it. Keep the second list short enough that people can actually honor it.
For the outputs on that second list, log each ruling, accept, modify, or reject, next to the reviewer’s name and the date. That record is how a team learns what its AI gets wrong, because every modify and reject is a correction somebody can study. Track two numbers: the share of listed outputs that went out with a recorded ruling, and the number of AI errors found after release. If the first number rises and the second falls, the line is probably in the right place. If errors keep appearing in work that was never on the list, move the line.
Then train to the list. Role-specific training reached 16% of the O’Reilly sample. Training built around the outputs a person is actually responsible for checking is role-specific by design, and it gives a worker something more useful than “be careful”: a short, named set of decisions that are theirs. Our work on measuring AI literacy as a governance competence treats the ability to question and own AI-assisted work as a skill a person can build and an employer can measure.
Alexia Pedersen of O’Reilly called the findings “a training and trust problem, not an AI problem.” The assignment sits underneath both. Training lands when a person knows what they are being trained to rule on, and trust follows when the people who rule are named. Twenty-four percent say they always verify. The number worth tracking inside an organization is the share of consequential AI outputs that someone was told to check, and did. Then watch whether errors keep surfacing in work that was never on the list.
Sources
- O’Reilly. (2026, October 27). Employee AI use is moving faster than training and governance, new O’Reilly research reveals [Press release]. https://www.oreilly.com/pub/pr/3490
- RGP. (2026). Humans still required: The new rules of work in the age of AI [AIxHuman research report]. Resources Connection, Inc. https://rgp.com/wp-content/uploads/2026/10/AIxHuman_Research_Report_FY27_RGP.pdf
- Warman, Z. (2026, October 30). Trust in transition: A global look at employee sentiment on AI and change. Perceptyx. https://blog.perceptyx.com/trust-in-transition-a-global-look-at-employee-sentiment-on-ai-and-change
FAQ
What did the O’Reilly survey find about verifying AI answers at work?
Censuswide surveyed 2,000 UK employees for O’Reilly between September 18 and 22, 2026. Only 24% said they always verify an AI-generated answer before using it. Among employees who use AI-generated answers in their work, 39% said they had acted on one that later proved inaccurate, incomplete, or unsuitable.
Should employees be told to verify every AI output?
A blanket rule is hard to follow under real workloads. A more workable approach is a short list of outputs that need a named reviewer before release, such as customer-facing, financial, or legal work, with the rest left to the worker’s judgment.
Why do managers matter so much for AI trust at work?
Perceptyx found that 82% of employees whose managers help them adapt to AI-related change trust their organization to use AI responsibly, compared with 56% across its full panel. Perceptyx describes this as a relationship, not proof of cause. The team level is where employees learn what AI work is theirs to approve.
What is a checkpoint in AI-assisted work?
A checkpoint is a defined point in a workflow where a named person reviews an AI-assisted output and accepts, modifies, or rejects it, and the ruling is recorded. It turns being near the work into holding authority over it, which is the core of checkpoint-based governance.
How can an employer tell whether its AI review list is working?
Track the share of listed outputs that went out with a recorded ruling and the number of AI errors found after release. Rising coverage with falling errors suggests the list is probably in the right place. Errors appearing in work that was never on the list mean the list needs to grow.
AI disclosure: This column was researched and drafted with AI assistance by HAIA Agents (RAI by GrokBot), using primary sources verified at time of publication. #AIgenerated
Leave a Reply