AI at work is often sold with the same promise: give employees the right tools and they will finish more tasks in less time. New workplace research described by Stanford suggests that message can miss what makes people want to use the technology in the first place. Staff were more engaged when managers presented AI as a way to make jobs more interesting and backed that promise with time to learn.
The Stanford workplace study account, published on 7 October 2026, describes research led by Arvind Karunakaran with colleagues Roshni Raveendhran and Tami Kim. Their work examined AI adoption in professional settings rather than measuring performance on an isolated benchmark. The contrast they found was striking, but it is important to understand what was observed before assuming the result applies to every company.
Why AI at work gets a different reaction
The researchers examined generative AI deployments at a law firm, an advertising agency and an IT services business. One part of the project drew on interviews with 184 employees across different roles. A closer study followed comparable groups at a law firm, where paralegals had access to the same AI drafting tool but encountered different messages from their managers.
In one division, the technology was introduced mainly as a way to draft documents more quickly and shorten turnaround times. Workers described feeling less involved in the substance of their work. They tended to use the system for established assignments rather than explore new possibilities. The result is understandable: if the most visible reward for adopting a tool is that management expects more output, experimentation may look like an unnecessary risk.
In another division, managers asked what staff found repetitive and which more challenging activities they wanted opportunities to pursue. Paralegals were encouraged to explore legal research and case analysis rather than treat the system purely as an output accelerator. Researchers reported greater use and experimentation in this group.
The second group used the tool 58% more and experimented 70% more than the first, according to Stanford’s account. These are results from the studied divisions, not universal expected gains for every employer. The groups were similar on important features, but management style was intertwined with training and the wider work environment.
The promise had to be backed by real opportunities
A manager saying that software will make work meaningful is not enough. In the second law-firm division, staff received formal training, opportunities to ask the vendor questions, regular knowledge-sharing sessions and time to learn independently. They also exchanged advice about useful applications and sought access to more complex legal discussions.
Those details are essential to interpreting the finding. The useful intervention was not a motivational slogan delivered at a meeting. It was a different set of opportunities for staff. An employee who is expected to maintain full output while learning a complicated tool has less reason to experiment than someone who is given protected time and recognition for doing it.
Organisations introducing AI at work can test this distinction without making grand predictions. They might invite staff to identify tasks that cause frustration, involve them in choosing trials, set safe boundaries for sensitive information and assess whether the tool actually improves the quality of the work. That process is slower than declaring a productivity target, but it can reveal whether the technology fits the job.
It is equally important to distinguish employees who are reluctant because they lack training from those whose concerns are about job security, accuracy or professional identity. A workshop may help with the first issue without resolving the others.
A faster report is not always a better report
The study also challenges how organisations measure value. If managers count documents produced or minutes saved, they may overlook work that becomes more demanding and more useful. Stanford describes paralegals taking on deeper research and case analysis, activities that may require more time than routine drafting even when AI provides assistance.
An employer can end up rewarding the wrong behaviour if it treats every use of an AI system as progress. Some businesses track how much AI-generated text staff produce or process. That may be easy to measure, but volume is not the same as customer value, sound judgement or fewer mistakes. An unnecessary query should not count as productive merely because it uses an expensive model.
A sensible evaluation begins with the task and its consequences. Was the document correct? Could the professional explain the result? Did a customer receive a better service? Were colleagues asked to correct hidden errors? The answers cannot always be reduced to a counter on a dashboard.
Related reporting on a workplace performance gap raises another question: who benefits from AI assistance, and under what conditions? The useful measure may change as workers learn new skills, not remain fixed at the metric chosen before the software arrived.
Why employee resistance can be rational
Resistance is often described as fear of change. Yet workers may be responding to the incentives placed in front of them. If a programme promises that repetitive work will disappear, but the freed time simply fills with more of the same assignments, employees notice the mismatch. Some will limit their use of the tool because they do not believe it benefits them.
This also helps explain AI workflow fatigue, where tools are added to processes without removing old steps or clarifying responsibilities. A professional may have to draft with AI, check every sentence, transfer the result into another system and record compliance evidence. The headline generation time shrinks, but the total work may not.
A healthier approach makes the entire workflow visible. Teams should record which steps have improved, which steps require additional checking and where discretion must remain with a named person. Employees should be able to report when a tool makes a task worse without being accused of resisting innovation.
The Stanford case studies cannot prove that every business following this approach will achieve similar gains. They do, however, identify a plausible mechanism. People are more willing to adopt tools when they can see a credible path from learning the tool to doing work that matters.
The kind of job makes a difference
The research also points to differences between occupations. Some jobs combine closely connected activities that cannot easily be separated. A salesperson may research a prospect, build trust, negotiate terms and handle the final agreement. AI can help with information gathering, but those elements still influence one another. Other jobs contain more discrete components that can be reorganised or delegated more readily.
That means an identical deployment programme can have very different consequences even within one organisation. A legal assistant preparing documents, an account manager handling clients and a technician investigating faults do not necessarily need the same automation boundaries or measures of success. Treating the technology as a universal productivity layer ignores how responsibility flows through each role.
Research into human and AI teamwork provides relevant context because teams have to decide when to accept a suggestion, when to challenge it and who owns the final decision. Training should therefore include the limits of the assistant, not merely the shortcuts that make it impressive.
Executives have genuine reasons to seek efficiencies. Employees have equally genuine reasons to ask what happens to their jobs when efficiencies are realised. A credible AI plan explains both sides. If the benefit is simply more output, say so. If the ambition is better work, provide the time, authority and opportunities needed to make that true.
What a thoughtful rollout would look like
For an organisation considering a new system, the best starting point may be a small, well-defined trial. Choose a process whose quality can be assessed. Ask the people doing the work which parts are frustrating or error-prone. Let them compare the old and new methods and record total time, correction effort and user satisfaction rather than only the speed of the first draft.
Managers should also make a decision about rewards. If someone learns to use AI to prepare stronger research or offer more useful advice, will that contribution be recognised? If the work becomes more complex, are there prospects for progression? Those questions determine whether employees experience the technology as support or as an instrument for squeezing more output from an unchanged role.
The Stanford AI Index workplace evidence places such experiments within a rapidly changing adoption landscape. It does not, on its own, prove how any particular employer should redesign its jobs. The practical conclusion from the new workplace research is narrower and more useful: people respond to the purpose of a tool and the conditions in which they are expected to use it. A company that wants genuine adoption must be willing to change management habits as well as software.
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.