From UX to AX – Why Agentic AI Is Shifting the Design Problem

Maeda draws on Don Norman's classic HCI model and reframes it for the age of autonomous agents. In the UX era, designers bridged the Gulf of Execution – they helped people carry out tasks. They designed flows, screens, menus, and paths. In the AX era, the problem shifts to the Gulf of Evaluation: how do I help someone know whether a task was done well?

The difference is fundamental. UX assumes a human acts and the interface guides them. AX assumes an agent acts and the human judges the result. Instead of designing clicks, we design delegation. Instead of optimizing flows, we shape feedback loops: action, outcome, evaluation, correction.

This goes well beyond adding a text field with AI behind it. Chatbot design and conversational UI remain subsets of UX – the user types, the system responds. AX begins where an agent independently decides which steps to take, in what order, which tools to use, and when to bring the human in. The designer's task shifts from the screen to the system, from the interface to orchestration.

When agents act autonomously, users need new control mechanisms. Victor Yocco – UX researcher at ServiceNow and author of the forthcoming book Designing Agentic AI Experiences – published the most concrete pattern framework to date in Smashing Magazine. Six patterns, organized along three phases: pre-action, in-action, post-action. Three of them deserve particular attention.

Intent Preview – the agent shows its plan before acting. Not a technical log, but a human-readable preview: "I will cancel flight AA123 to San Francisco." The user has three options: approve, adjust the plan, or take over the task entirely. It sounds like a small detail. In practice, it is the decisive moment where trust is built or broken. Yocco describes the Intent Preview as deliberate deceleration – a design decision that prioritizes control over efficiency. Yocco's benchmark: if the plan acceptance rate drops below 85 percent, something is wrong with the agent – the model needs an update, not the interface.

Autonomy Dial – a control that defines what the agent is allowed to do on its own. Not globally, but per task type. Four levels: observe and suggest, plan and propose, prepare and request confirmation, act autonomously. Granularity is everything. A user who allows the agent to reschedule calendar entries does not thereby authorize it to cancel contracts. Yocco's warning: a single failure at the wrong autonomy level leads users to abandon the agent entirely – rather than simply dialing back its permissions.

Action Audit & Undo – a chronological log of all agent-initiated actions with time-limited undo windows. The single most effective trust-building element. The ability to reverse an agent's action lowers the barrier to delegation more than any explanation ever could. Yocco points to the Service Recovery Paradox: a user who experiences a failure followed by a smooth correction can become more loyal than one who never encountered a problem. Threshold: if the reversion rate for a given task type exceeds five percent, automation should be disabled for that type.

Alongside these patterns, Microsoft Design published a principles framework that takes a remarkable shift in perspective. Rather than masking uncertainty, Microsoft treats it as an inherent design element: a certain degree of uncertainty is to be expected with agents – and that is precisely why reasoning and confidence levels must be made visible.

The framework rests on three principles. Transparency – users know that AI is involved, how it works, and how to provide feedback. Control – users can set preferences and adjust the system, including the ability to have data deleted. Consistency – the experience works across devices and modalities with familiar UI elements.

In practice, this is already visible in Copilot Cowork, Microsoft's agentic work mode within Microsoft 365. Complex requests are broken into multi-step tasks, progress is visible, and users can intervene at any point. Employees interact with the agent the way they would with a colleague – in Teams chats, not in a separate tool. The boundary between delegating to a human and delegating to an agent is blurred by design.

The role shift is already underway. Maeda puts it this way: the designer of the future is less of an executor and more of a critic, editor, and judge. That doesn't mean every designer needs to learn to code. But every designer needs to understand how systems, loops, and feedback mechanisms work.

Figma's data makes the gap tangible that AX exists to close: 78 percent of designers say AI makes them more efficient – but only 32 percent trust AI output. That discrepancy is not a communications issue. It is the design problem that AX exists to solve.

A practical planning tool comes from the autonomy framework by Feng, McDonald, and Zhang (University of Washington), published by the Knight First Amendment Institute. It defines five levels – from Operator (user retains full control at all times) to Observer (agent operates independently, user only monitors). The key insight: autonomy is a design decision, not a capability question. A highly capable agent can be deliberately designed for low autonomy – because the context demands it. If you are working on a customer portal today, you should already be asking: at which level does the agent operate in this workflow? The answer determines which patterns need to be implemented.

The window of opportunity is reminiscent of 2008, when the first teams took mobile UX seriously, or 2012, when responsive design went from experiment to standard. Those who establish AX patterns now will shape how the next generation of digital products handles delegation, trust, and autonomy.

But the enthusiasm has limits. Gartner predicts that more than 40 percent of agentic AI projects will fail by the end of 2027 – due to escalating costs, unclear value, or insufficient governance. KPMG's Q1 2026 AI Pulse shows that 63 percent of companies now require human validation of agent outputs – up from just 22 percent a year earlier. The trend is not toward blind trust. It is toward structured skepticism.

And that is exactly where the design mandate lies. AX without trust design will not scale. Agents that make opaque decisions, offer no undo options, and lack a graduated autonomy model will not be adopted – no matter how capable they are technically.

The design problem has shifted – from execution to evaluation, from interface to system, from click to delegation. This shift is irreversible. The question is not whether it affects your products, but how quickly your team responds. Three steps that make sense right now: First – review your existing products for tasks that lend themselves to delegation. Second – add trust patterns like Intent Preview, Autonomy Dial, and Action Audit to your design system backlog. Third – define autonomy levels as a deliberate design decision, not a technical byproduct. Those who embed these patterns now will build products that users actually entrust with tasks – unlocking efficiency and automation potential that was never achievable with traditional interfaces.

Facing the question of how your digital products should handle autonomous agents? 21TORR combines UX expertise with AI strategy – from autonomy analysis to trust pattern design. Let's talk about what AX means for your next project.