Workers say they want AI as a true partner, not an autopilot — but is anyone actually building it that way?

This explores which control and safety features workplace users actually ask for in AI agents, and whether products are delivering them. The most direct evidence comes from a multi-method study of business users. They ranked human control, reliability, context-awareness and safety as practical necessities, not nice-to-haves. In their view, an agent that lacks these features won't be adopted, however capable it is What UX principles do workplace users want in AI agents?. What workers want is a working partnership, not hands-off automation. A survey of 1,500 workers across 844 tasks found that equal human-AI partnership was the most-wanted arrangement in 45% of occupations. Yet 41% of startup investment targets levels of automation that don't match what workers say they want What collaboration level do workers actually want with AI?.

One way to see past the usual "more automation means less control" trade-off comes from Shneiderman's framework. It treats automation and human control as two separate dials that can both be turned up at once. Patient-controlled pain medication is the example: the pump runs automatically, and the patient still decides when to take a dose Can automation and human control increase together?. Read this way, when users ask for control they aren't asking the agent to do less. They want to keep the right to steer it. A related idea from organizational research splits shared work by two questions: who sets the rules, and who picks each action How do humans and technology share control over organizational routines?. That gives a sharper vocabulary for what "control" might mean in a given workflow.

Salesforce's design system shows what these demands look like once they become product features. It sorts safeguards into five groups: deliberate friction before risky steps, labels that mark AI-generated content, bias checks, explanations of what the AI did, and hallucination reduction. The safeguards are concentrated at irreversible actions, because that's where users tend to stop trusting the system How can product design build human trust in AI systems?. Shneiderman makes a further point: reliability, safety and trust come from different sources. Reliability comes from engineering practice, safety from management culture, and trust from independent oversight. A good confirmation dialog therefore can't make up for a weak organizational safety culture Can we separate reliability, safety, and trust into distinct mechanisms?. Simkute and colleagues bring in older lessons from automation design: continuous feedback, a clear split of who does which task, and interfaces that keep the work stable. These proposals have not been tested on generative AI yet Can Human Factors design fixes restore GenAI productivity losses?.

The good news is that workplace tools may already be ahead. A review of nine deployed agents checked them against 14 interaction-design principles. General-purpose productivity assistants like ChatGPT and Claude scored highest. Companion and advisory agents scored lowest, and the principles they broke most often were dependency prevention and making the agent's intent visible Do productivity agents actually protect users better than emotional ones?.

There is also a control need that doesn't show up on any UX checklist. Across four experiments with more than 4,400 people, AI users expected colleagues to rate them as less competent and less diligent, and they were less willing to tell managers they had used AI Do people fear judgment when they use AI at work?. That points to a demand the feature lists miss: control over how one's own use of AI is seen at work. This is an inference from the corpus, not something the studies measured. The corpus is also thin overall. Most of these sources are frameworks or survey rankings, and few test which specific safety features change how people behave.