A manager announces that the company wants everyone to become more productive with artificial intelligence. The staff are expected to start using unfamiliar tools, yet nobody has explained which information is safe to enter, how to check an answer or when a human must make the final call. That gap is what makes AI training at work more than an optional perk.
Research published by the Federal Reserve Bank of New York in April 2026 found that 38% of surveyed employed Americans considered training in generative AI important, while only 15.9% said their employer offered any. The New York Fed analysis used responses collected in November 2025. It describes a mismatch between interest and formal provision, rather than proving that every employer should buy the same course or that all workers want AI in their jobs.
People are already experimenting while the rules catch up
The survey found that 39% of employed respondents had used AI tools in their current job within the preceding 12 months. Some reported quicker completion of tasks or greater personal productivity. Others were still learning how to use the software and said the learning process made work slower. That mix will sound familiar to anyone who has watched a colleague produce an excellent first draft and then spend an hour checking what the software invented.
A tool that writes confidently can be useful and risky at the same time. It may produce a clear summary yet misunderstand a customer's contract, misquote a number or reveal details that should never have left a company's system. The lesson is not simply how to type a better prompt. Workers need to understand what the tool is allowed to do, how it arrives at an answer and which mistakes would have serious consequences.
Some employers appear to have avoided the issue by restricting access. In the Fed survey, 37% of employed respondents said their workplace did not offer AI tools, while a further 11% reported that their employer actively prohibited their use. Those responses reflect workplace access, not necessarily a personal desire for training. A company with sensitive records may have legitimate reasons to limit a public chatbot until approved alternatives are available.
AI training at work still follows existing advantage
The Fed's findings reveal differences in who uses AI at work. Among respondents, the proportion reporting workplace use was substantially higher among college graduates than among people without a degree. It was also higher among better-paid and full-time workers. Such patterns suggest that access and confidence are not being distributed evenly across the labour market.
This matters because workplace technology can amplify a person's existing advantages. Someone with a supportive manager, time to experiment and reliable access to approved tools may learn quickly. Another worker may be judged against higher expectations without receiving the same opportunity. A productivity target that ignores training can become a disguised test of who already had the resources to practise.
The subject connects with earlier LiveAIWire reporting on older workers and the AI skills gap. Age alone cannot explain someone's ability to learn a system, but past access to training and confidence with new software can affect who volunteers to try it. Employers should be wary of assuming that enthusiastic early adopters represent everyone on the payroll.
A practical programme should begin with actual duties. A warehouse administrator, a customer-service assistant and a financial analyst encounter different information and different risks. One uniform demonstration of a chatbot's most impressive features will not prepare them equally. The right skills are those that help each person perform their real work with appropriate safeguards.
The value of training looks different depending on who gets it
The Fed also asked respondents about hypothetical pay trade-offs. Workers without employer-provided AI training were asked what salary they would surrender to move to an otherwise identical role that offered extensive training. The average answer was 11.4% of salary, but the median was zero. Around 61% of respondents without training were unwilling to accept any pay cut for it.
It would be misleading to report only the average and conclude that employees happily want to sacrifice a large part of their wages. Some people gave high values, while a substantial share gave none. The question was hypothetical and measured stated preferences, not actual job changes or paid training purchases. A person may value learning while reasonably believing the employer should pay for skills the job now requires.
The reverse question produced a different pattern. Among workers who already had access to training, the Fed reported a higher average salary increase required to switch to an otherwise identical job with no such provision. The comparison may reflect loss aversion, selection into different jobs or other factors. It does not prove that offering training alone causes an equivalent increase in pay or productivity.
What emerges is more human than a simple technology statistic. People often value a benefit differently once they have experienced it, and they differ sharply in whether they expect to use the skill. An employer trying to improve capability should listen to that variety rather than turn a survey average into a compulsory training budget.
A course cannot replace workplace judgement
Useful AI training should include examples of failure. If a model invents a source, confuses a client with a similarly named person or produces a plausible but false spreadsheet explanation, employees need a routine for spotting and correcting the problem. The exercise should make clear who remains accountable for a decision and when the work must be escalated.
This is particularly important for people learning a profession. Research discussed by LiveAIWire about junior lawyers and AI assistance raised questions about what happens when a tool supplies an answer before a newcomer has developed the expertise to evaluate it. A training programme that teaches only speed may leave staff less able to recognise a serious mistake. Practice in questioning and verification is part of the skill, not an obstacle to it.
Confidentiality is another area where policy and training must meet. Workers should know which applications are approved, what data may be uploaded and where records of AI-assisted work belong. The right instruction may sometimes be to avoid AI altogether for a particular task. Clear boundaries can be more empowering than a vague warning to 'be careful'.
Managers also need to examine whether the technology changes workload fairly. If an employee saves time on drafting but is then assigned more work without support, the apparent efficiency gain may feel very different to the person doing the job. Software adoption is a management decision as well as a training exercise, and it should be evaluated in the light of what actually happens to staff.
A global survey points to a similar concern, not the same statistic
A separate PwC global workforce survey published in September 2026 describes uneven access to AI and learning opportunities across a much broader international sample. Its findings reinforce the idea that some workers are advancing faster than others, but its definitions, population and dates differ from the New York Fed research. Their percentages should not be placed next to one another as though they came from an identical experiment.
The shared issue is an organisational one. Companies may invest in licences and demonstrations while neglecting the skills, trust and process changes needed to use the tools well. A business can announce an AI strategy and still leave the person answering customer enquiries uncertain about what is permitted. Training closes only part of that gap; managers must also redesign procedures and accept responsibility for outcomes.
This is related to LiveAIWire's reporting on the wider workplace AI divide. Employees are not one homogeneous group waiting for the same software. Their duties, confidence, incentives and support differ. Measuring success by the number of accounts activated would miss whether the people most affected can use the tools safely and productively.
Start with the task that goes wrong most often
The clearest first step for an employer is not a lengthy presentation on the future of AI. It is to identify a routine task where errors and delays already matter, decide whether an approved tool can genuinely help and teach employees how to verify the output. Record what improves and what does not. Include people who are sceptical, because they may understand the risks better than the enthusiasts.
Training should be repeated when tools or company policies change, and staff need somewhere to report unexpected results without embarrassment. The goal is neither to persuade everyone to use AI for everything nor to force employees to discover the safe method alone. It is to give workers enough knowledge to decide when the technology is useful and when it is not.
The New York Fed findings expose an uncomfortable possibility: employers may be moving faster in their expectations than in their support. If AI becomes a normal part of work, learning to challenge it will be just as important as learning to operate it. A company that understands that distinction is more likely to benefit than one that simply hands out logins.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.