How should AI interfaces handle the shift from doing to supervising?

Nielsen argues that AI interfaces need a new design model because intent-based computing changes what users do: "Users are changing from doing the work (operating the UI) to supervising the work." He proposes that "mature intent-based systems will settle into a triple-layered design model": an Intent Surface where "the user states an outcome" and, as it matures, will "increasingly rely on implicit intent inference" drawn from ambient context; an Orchestration Surface, "the critical negotiation layer" where the agent must "reveal its proposed plan, expose the provenance of its data, and seek consent," and afterward issue "post-action receipts" that "summarize what it changed, which systems it touched, what assumptions it used, and what can still be undone"; and a Direct-Manipulation Surface, the familiar GUI kept as a fallback "reserved for edge-case editing, granular corrections, and emergency overrides."

His reasoning is that each layer answers a different failure mode, not a single one. The articulation barrier — "It's often hard to put something into words" — is handled at the Intent Surface by multimodal input and context that drafts the prompt for the user rather than waiting for better phrasing. The loss of "implicit knowledge" of what happened, since the user no longer executes each step, is handled at the Orchestration Surface, which "must manufacture legibility after the fact" through transparency and consent rather than relying on the user having watched the work happen. And the residual need for precise, low-level correction is handled by demoting the GUI rather than removing it. He ties this to a rewrite of his own usability heuristics: "Visibility of system status" becomes showing "what the system believes the user intends," and "User Satisfaction" is replaced by "Trust Calibration," which he calls "the primary functional metric of an intent-based system," achieved partly through counterfactual explanation — "I chose Plan A over Plan B because cost mattered more than speed" — rather than a confidence score.

This formalizes, from a UX-practice angle, the same gap Why can't users articulate what they want from AI? describes as a cognitive asymmetry: the Intent Surface is a design answer to it, betting on ambient-context inference rather than better prompting, and the "articulation barrier" names the same phenomenon that How do users actually form intent when prompting AI systems? frames as unresolved intent maturation. Nielsen's shift "From Error Prevention to Clarification Quality" describes, without measuring, the failure that Why do AI agents miss most of what users actually want? quantifies directly. And the Orchestration/Direct-Manipulation split echoes, at the UI-design level, the planning-versus-grounding separation Can structured interfaces help language models control GUIs better? proposes at the agent-architecture level — both conclude that negotiating what to do and executing it precisely need different interfaces.

The excerpt is a practitioner essay, not an empirical study: the triple-layered model, the heuristic rewrites, and the "chauffeur" analogy are design proposals offered without usability testing, adoption data, or named products that implement them. It does not establish that users actually calibrate trust the way the model assumes, or that post-action receipts reduce errors in practice — only that a veteran usability researcher believes this is the structure such systems will need. Read as a design hypothesis rather than a finding, it is worth testing against the measured intent-and-trust gaps the neighboring notes document, not a substitute for that evidence.