AI & Emerging Interfaces

AI features fail socially before they fail technically: people stop using what they cannot predict or challenge. So the core surfaces are the explanatory ones: what the system concluded, how confident it is, what evidence it used, how a person overrides it. A visible, defensible score turns output from a verdict into an argument a professional can take responsibility for.

Uncertainty gets designed with the same care as success: calibrated language for each confidence band, empty and error states that say what the system did not know, and escalation to a human that does not read as failure. Every failure state is specified with the same rigor as the happy path: what the system says, what it exposes about why, and what the person can do next.

Chat is the right surface when intent is genuinely open-ended, and the wrong one when a form, a table, or a button would answer faster. Across text, voice, and video, the design work is conversational structure: turns, memory, interruption, repair, and the handoff between modalities and to humans.

The rest is choreography with the surrounding product: where an assistant may act on someone's behalf, where it must show its work first, and how its actions land in the same audit trail as everything else.

AI has changed how the argument gets made, not just what gets designed. Working prototypes, increasingly AI-assisted, settle in days what static specs debate for weeks; stakeholders react to behavior, not promises. The same loop evaluates AI features themselves: prototype the behavior, test it against real cases, and let the observed failure modes design the guardrails.

None of this suspends the evidence rules. AI features get instrumented, tested with users, and labeled by how their outcomes were established, exactly like everything else.