Summary

• An O'Reilly survey of 2,000 UK employees found that 39% turn first to AI chatbots for work-related answers, rising to 50% among those aged 25–34.
• Among employees using AI-generated answers at work, 39% reported acting on information that subsequently proved inaccurate, incomplete, or unsuitable.
• Only 16% received role-specific AI training during the preceding year, while 35% reported receiving no AI training or guidance.

Almost two in five UK employees now turn first to AI chatbots when seeking answers to work-related questions, according to research from O’Reilly, although many are using the technology without formal training or consistently checking the information it provides.

The survey of 2,000 employees found that 39% preferred asking an AI chatbot before consulting a colleague, manager, or search engine. Among respondents aged 25–34, the figure reached 50%, indicating that the technology has become a particularly established source of workplace information within that age group.

Yet 35% of employees said their organisation had provided no AI training or guidance during the preceding 12 months. Only 21% had received a formal AI training course, while 16% reported instruction specifically relevant to their job.

O’Reilly commissioned Censuswide to conduct the research between 18 and 22 September 2026, publishing the findings on 27 October. The results also raise questions about verification: among employees using AI-generated answers at work, 39% said they had acted on a response that later proved inaccurate, incomplete, or unsuitable.

Chatbots enter everyday information gathering

Employees regularly need information that sits outside their immediate knowledge, whether they are interpreting technical documentation, examining a software problem, preparing a report, or learning an unfamiliar business process. Traditionally, that work might involve consulting colleagues, searching for reference material, or reviewing internal guidance.

AI chatbots offer another route by allowing users to describe a question in ordinary language and receive a response assembled from information available to the system. They can also help users reformulate questions, compare possible approaches, and explore terminology before deciding how to proceed.

The convenience can change working habits, particularly when a response is available more quickly than locating an appropriate colleague or examining several documents. However, the apparent completeness of an answer does not establish that the information is accurate.

AI-generated responses can contain factual mistakes, omit qualifications, or combine information drawn from different contexts. A technically plausible answer may also be unsuitable because it describes an outdated version of a system or assumes conditions that do not apply to the organisation using it.

These limitations have different consequences depending on the task. An error in an informal explanation may be corrected before affecting any work, while an incorrect technical recommendation could influence a production system, financial calculation, contractual decision, or customer-facing service.

The survey found that 17% of employees using AI-generated answers had acted on unsuitable or incorrect information more than once. The figure concerns respondents’ reported experiences and does not establish the severity or financial cost of those incidents.

Only 24% of employees using AI-generated answers said they always verified responses before relying on them. This does not mean every unchecked answer was inaccurate, but it indicates that consistent checking was far from universal among respondents.

Training has not matched employee behaviour

While employees are introducing AI into their own work, many organisations have not established a corresponding programme of instruction. The 35% reporting no training or guidance included employees who had not received formal courses, practical workshops, or other assistance with using the technology.

General familiarity with a chatbot does not necessarily equip someone to evaluate its output. Employees may learn how to obtain fluent answers without understanding how to determine whether those answers are supported by appropriate evidence.

Role-specific training can address the different risks associated with particular functions. Software engineers evaluating generated code, for example, may need to consider testing, security, dependency management, and maintainability.

A finance employee using AI to investigate a variance or prepare calculations needs reliable source figures, appropriate assumptions, and a record of how the result was established. Staff handling sensitive customer information must also understand what data may be entered into an external system.

The relevant controls therefore extend beyond teaching employees how to write prompts. They include identifying suitable use cases, protecting information, selecting authoritative references, and recognising when a decision requires specialist review.

O’Reilly’s respondents indicated that human involvement remains an important source of confidence. When asked what would most increase their trust in using AI for consequential workplace decisions, 39% selected human review and 35% identified better AI training.

Those preferences suggest that employees value practical support when assessing outputs. They do not, however, prove that a particular training programme would prevent every reported error or produce a quantified productivity improvement.

Verification needs to fit the task

Businesses adopting AI-assisted workflows must decide when an employee can use generated information directly and when additional checks are required. Applying the same review procedure to every request may consume much of the time saved through automation, while allowing consequential decisions to proceed unchecked can introduce avoidable risk.

One approach is to distinguish exploratory tasks from actions affecting live systems, financial commitments, regulated information, or customers. Employees can use AI to investigate possible explanations or identify relevant documentation, with established approval and testing procedures retained before changes are implemented.

Technical recommendations can be checked against current product documentation and tested in an appropriate environment. Financial information may require reconciliation with verified records, while legal or regulatory interpretations need examination against the applicable rules and organisational circumstances.

These processes work best when reliable reference material is readily available. An employee cannot easily verify an AI-generated answer against internal procedures if the relevant documentation is outdated, incomplete, or difficult to locate.

Maintaining current information is therefore part of the wider challenge of using AI consistently. Organisations may also need to review whether the systems they permit employees to use provide adequate controls over confidential material and records of consequential decisions.

Alexia Pedersen, senior vice-president international at O’Reilly, has emphasised the need for structured, role-specific learning and verification practices rather than simply discouraging staff from using AI.

That approach requires management to understand how employees are already using the technology. Policies developed without reference to actual working practices may leave important activities outside established controls, particularly where staff use publicly available tools alongside approved corporate applications.

Expert information becomes part of AI systems

O’Reilly has separately developed Expert Intelligence, a service that connects enterprise AI tools with technical and professional information from its publishing and training library.

Introduced to selected customers in July 2026 and opened to wider founding-member applications in October, the offering includes a Model Context Protocol-compatible connector and predefined agent skills for specialist tasks.

The connector is intended to make identifiable reference material available inside tools employees already use, allowing responses to draw on information associated with named authors and publication dates. Its agent skills provide guidance for work such as preparing architectural decisions, building threat models, and planning technical projects.

Access to traceable sources can make an AI response easier to investigate, particularly where an employee needs to understand the origin of a recommendation. However, citation alone does not establish that the interpretation is correct or appropriate for a particular business environment.

Technical information may be accurate when published but subsequently superseded by software updates, changes in regulation, or new operating requirements. Users still need to establish whether the source applies to the question being considered.

The company’s product announcement predates the workplace survey and is a separate commercial development, rather than an intervention whose effectiveness has been demonstrated by the research findings.

Likewise, the Censuswide survey measures employees’ reported behaviour and experiences. It does not establish the frequency of incorrect answers across all workplace AI use, identify the financial consequences of reported mistakes, or provide an independent assessment of employees’ verification practices.

The findings nevertheless show how quickly an information source can become part of everyday work before organisations establish consistent procedures for using it. The next practical test for employers will be whether training, reliable reference information, and appropriate review processes can keep pace with employees’ existing reliance on AI-generated answers.