What do office workers actually need from an AI agent before they'll trust it enough to use it at all?

This explores what workplace users want from AI agents when it comes to safety and control, and how designers are turning that into real interface decisions. The most direct evidence comes from a multi-method study of business users. They named human control, reliability, context-awareness and safety as practical requirements. To them these decide whether a tool gets adopted at all, not whether it's merely nice to use What UX principles do workplace users want in AI agents?. The interesting part is what 'control' turns out to mean once you look at how designers put it into practice.

It mostly doesn't mean watching the agent constantly. Gunasekaran argues that what makes an agent feel trustworthy is *perceived* control: the user can see a clear way to step in, even if they rarely use it Does perceived control matter more than actual control for AI trust?. Control also moves earlier in the process. Instead of a single goal box, agent interfaces should collect intent up front: limits, risk tolerance, forbidden actions, and the conditions under which the agent needs permission. A vague instruction gets carried through every step the agent takes without anyone reviewing it Should agentic AI interfaces require users to specify risk tolerance upfront?. Designers are settling on a shared vocabulary for this. One example is the 'autonomy slider', a per-task setting that runs from suggest-only to draft to execute. Nobody has yet tested it against other ways of shaping control Should agent autonomy be a continuous slider or fixed levels?. Salesforce's design system goes further and builds 'mindful friction' into the moments that can't be undone, because that is where users tend to lose trust How can product design build human trust in AI systems?.

The less obvious finding is that safety at work can't be fully designed into the agent itself. One framework applied to more than 2,000 real job tasks found risks that come from how agents, goals, work environments and human relationships interact. These risks appear even when every single part works as intended Can workplace AI risks emerge from interactions alone?. Shneiderman's framework reaches a similar conclusion from the organizational side. Reliability comes from engineering practice, safety from management culture, and trust from independent oversight. Improving one doesn't automatically improve the others Can we separate reliability, safety, and trust into distinct mechanisms?. So when a workplace user asks for 'safety', part of what they need lives outside the product.

Two more findings complicate the picture. First, control may start flowing both ways: agentic systems can hand tasks and responsibilities back to humans, so the user is no longer the only one delegating Can AI systems delegate tasks back to humans?. Second, these demands for control rest on today's limited capability. Leading agents complete only about 30% of tasks in a simulated workplace and struggle most with social interaction and unfamiliar professional software Why do AI agents fail at workplace social interaction?. Part of why users want visible ways to intervene is simply that agents still fail often.

The surprise is that workplace tools may be the good case. A review of nine deployed agents found general productivity assistants like ChatGPT and Claude scored highest on interaction principles. Companion and advisory agents scored lowest, especially on preventing dependency and making the agent's intent visible Do productivity agents actually protect users better than emotional ones?. Business settings seem to push design toward control in a way that emotionally charged settings haven't yet.