How to Design Agentic AI Defaults That Preserve User Agency

The last fortnight of April delivered five vendor releases pointing the same direction. Microsoft 365 Copilot’s agentic mode went generally available on April 22. OpenAI’s Workspace Agents launched the same day. Anthropic shipped persistent memory for Claude Managed Agents on April 23. Google announced Gemini Enterprise’s agentic taskforce. Salesforce launched Agentforce Operations on April 29. If you are a Builder shipping an AI feature this quarter, the design question on your desk is no longer whether to add an agentic default. It is where the agent gets to act, and how cleanly users can see what it did.

What the research shows

A team from Northwestern, the University of Chicago, and Microsoft Research interviewed twenty freelance science journalists at CHI 2026 in Barcelona last month. The paper is titled “Helping Me Versus Doing It for Me.” They built four prototype AI writing tools that varied along three dimensions: who starts the task, how much of the work the AI does, and how much the user can configure. Then they walked each journalist through every tool and asked what they would actually use.

The journalists drew a sharp line. AI that gathered information or critiqued their writing was widely accepted because their decision-making stayed intact. AI that wrote first drafts or generated core ideas was treated as a threat to autonomy, skill development, fulfillment, and professional relationships. The pattern held even for tasks that look automatable, like adjusting writing voice with a slider.

The authors named two kinds of agency at play. The moment-by-moment kind, where users keep control over what gets decided and executed. And the slower kind, where doing the work yourself is what builds the craft and the values you will need to evaluate AI output later. Both kinds erode together when the AI does too much. This is the same conference that produced the National University of Singapore work on AI chatbots and self-image research a few weeks back. Different paper, different lab, related thesis: when an AI shapes a user’s work and self-description in the same loop, both shift.

What it doesn’t tell us yet

Twenty freelance science journalists writing pitches is not the whole knowledge economy. It is one professional context, one task type, and hypothetical Figma prototypes rather than tools they actually used over months. The methodology is qualitative, the numbers are small, and one of the authors is at Microsoft Research, the same quarter Microsoft shipped Copilot’s agentic mode. There is also a parallel finding from the International Journal of Information Management this March. A randomized experiment grounded in psychological reactance theory showed that when people perceive their autonomy is reduced, they distrust the AI and push back, especially in low-stake tasks. Different methods, same direction. The pattern is not proven across all users. The question every product team is now asking is whether it holds where they ship.

One thing to notice in your work today

If you are shipping an agentic default this quarter, here is the question worth carrying into your next design review. For each AI behavior in the workflow, ask three things. Who initiates it. How much of the user’s decision-making does it replace. Can the user see exactly what was changed and undo it cleanly. Information-gathering and feedback usually pass that test. Auto-drafting and idea generation usually do not. The same scope question came up yesterday in the workforce piece on how to audit your team’s AI recovery time, and earlier this season in the harness piece on where coding-agent ROI shows up first. Same lever, different surface. The research is small and qualitative. The line it draws is the one users will draw, too.