Map
I map the task, agent actions, risk points, user decisions and current failure modes.
AI AGENT UX / HUMAN–AI INTERACTION
Design AI agents users can understand, supervise and recover from.
I design one critical AI workflow around trust, autonomy, progress, approval and recovery—so users know what the agent is doing, what it needs from them and what happens when it gets something wrong.
5–10 daysOne workflowAll critical statesFixed quote
THE INTERFACE IS THE TRUST LAYER
An agent can be capable and still feel unsafe.
Traditional software waits for users to act. AI agents can plan, decide and execute on a user's behalf. That changes the UX problem from “which button comes next?” to “how much should the system do, what should it reveal and when should the user intervene?”
- Invisible action. The system does work, but users cannot see what is happening or what has changed.
- Wrong autonomy. The product either asks for approval constantly or acts too far without permission.
- No recovery model. A wrong result becomes a dead end because undo, correction or escalation was never designed.
- Chat-first by default. A conversational shell is used where a structured workflow would be faster, safer and easier to inspect.
WHAT I DESIGN
The full lifecycle of one agent workflow.
- Before action. Intent capture, scope, permissions, plan preview, constraints and the right level of user confirmation.
- During action. Progress, tool use, intermediate states, rationale, uncertainty, pauses and clear signals when the agent needs input.
- After action. Result review, source/context visibility, audit trail, correction, undo, retry and escalation.
- Failure states. Partial completion, timeout, missing data, low confidence, conflicting instructions and safe fallback behavior.
- Human handoff. When automation should stop and how context survives the transition back to a person.
HUMAN–AI UX SPRINT
One workflow. Every state that affects trust.
Design
I redesign the flow around the right autonomy, visibility, approval and recovery patterns.
Specify
You get the interaction states, reusable AI-specific patterns and implementation notes your team can build from.
COMMON QUESTIONS
AI UX is not a new coat of paint on chat.
Do we need a separate AI design system?
Usually not. I extend the existing product system with the AI-specific components and states the workflow needs—such as progress, confidence, review, approval, intervention and recovery.
Do users need to see the model's reasoning?
Not necessarily. Users need enough evidence and system state to make a decision. The interface should explain outcomes, status and risk without dumping internal model mechanics into the product.
How much autonomy should an agent have?
It depends on consequence and reversibility. Low-risk reversible actions can often run with less friction; high-impact or novel actions need stronger review, permission or approval boundaries.
What is the deliverable?
A redesigned critical flow with the relevant AI states, interaction rules and implementation-ready specifications. Scope is fixed before the sprint begins.
START WITH ONE REAL FLOW
Send me the agent workflow users hesitate around.
I will tell you where trust breaks, what should be redesigned first and what a bounded Human–AI UX sprint would include.
Send the AI flow