BLUEBERRY
Making AI Work Visible, Reviewable, and Safe
COMPANY
Self-Initiated Concept
ROLE
Product Designer
EXPERTISE
Enterprise Product Design
YEAR
2026
Blueberry is a concept for supervised AI work in customer operations.
During a service outage, a support operations manager needs to move quickly without losing control of customer communication. Blueberry makes the agent’s plan, access, actions, uncertainty, and approval points visible across one connected workflow.
The challenge
AI can accelerate work, but opaque intent, broad permissions, and irreversible actions make delegation feel risky. The design question was how to help a manager stay accountable without turning every task into a control-panel exercise.
My role
I defined the scenario, information architecture, interaction model, visual system, and clickable prototype as a self-initiated portfolio concept.
Design the trust contract before the screens.
I began by mapping the decisions that carry risk: what the agent will do, what it can access, when it should pause, and how a person can understand or reverse a consequential action.
01 — Make intent legible
Preview the plan with evidence, sequence, and stopping points before work begins.
02 — Scope power to the task
Use least-privilege permissions in plain language, rather than blanket access.
03 — Pause at judgment
Surface uncertainty and reserve consequential choices for a human reviewer.
One clear workflow from task to accountable outcome.
The experience moves through a command center, task creation, plan review, permissions, live activity, uncertainty handling, approval, completion, and a reversible audit trail.
The interaction model
Blueberry avoids a generic AI dashboard. Each moment answers a specific operator question: What will happen? What can it touch? Is it still on track? Do I agree? What happened?
Key states
Plan previews expose intent before execution. Task-scoped permissions make access visible. A live timeline shows evidence and uncertainty. Grouped approvals keep human judgment at consequential steps. Audit and undo make the resulting action accountable.
A design direction, not a claimed product outcome.
Blueberry has not been shipped or user-tested. It was created to explore how AI operations could remain understandable and controllable in a high-stakes support scenario.
What the next phase would test
A scenario-based study would evaluate whether operators can predict the agent’s next action, notice uncertainty, intervene effectively, and identify who approved a send or how to reverse it.
Reflection
The central tradeoff is deliberate friction. Plan, permission, and approval reviews should scale with consequence — not appear identically for every task.