Designing trust into autonomous agents

PointFive was adding agent capabilities — AI that doesn't just flag a cloud-cost issue, but fixes it autonomously. The moment agents could act on their own, something shifted: people started to fear them, not trust them, and quietly avoided turning them on.

This wasn't a technology problem — it was an experience problem. A user deciding whether to let an agent touch their cloud bill is making a trust decision, and trust is pure design: control, transparency, and a clear mental model of what the agent will do. If users don't trust what they can't see, the capability goes unused — and the AI-native bet fails on adoption, not on the model.

• Pending decisions were buried inside each agent's page, behind an unlabeled icon — users running several agents never knew something was waiting.
• There was no sense of an agent's blast radius — how much, and what, it was about to affect.
• Approving actions one agent at a time didn't scale — five decisions meant five separate screens.
• For an agent someone else set up, it was unclear what it was doing or even whether it was running.

I designed against PointFive's five brand values as a lens — Efficiency, Depth, Creativity, Bias to action, and Trust. The challenge was holding them at once: dense, live agent data had to stay calm and scannable, and powerful controls had to stay safe.

I grounded the work in two kinds of research. I studied 10 agentic products for oversight and governance patterns, looked closely at direct FinOps competitors (Vantage, Finout, ProsperOps, nOps, CloudZero), read 10 studies on AI trust and human-in-the-loop, and pulled signal from McKinsey, Gartner, SurveyMonkey and NN/g. One gap stood out: even the most advanced competitor runs oversight through Slack or routes it to ticketing tools — none offered a single in-product pane to see every agent, its scope and its pending decisions. From both streams I distilled five principles that move trust: Intent Preview, Action Audit + Undo, Variable Autonomy, Confidence Signals, and a Single Pane of Glass.

I explored the overview screen as interactive prototypes built with Claude as live HTML — fast to feel in a browser, not just look at. The core question was how to surface the decisions waiting for review: keep them quiet in the dashboard, or pull them up as a dedicated call to action. Each concept made a different bet on how hard the screen should push the user toward pending approvals.

I'm honest about where this stands: it's a starting point, not a finish line. I focused on the place where the tension between human and AI is most visible — the moment a person decides whether to let an agent act. The whole design works to make the user feel they are the one in control, the one in the loop, even as the agents do the work.

What's next: expanding and focusing the single agent page itself; a deeper end-to-end flow (edit / delegate / snooze branches); validation with the FinOps users who live in this screen; and the edge and empty states — no decisions, all snoozed, a failed run with recovery.

• Used AI as a research partner — clustering the product scan into themes and generating 20 live prototype directions — with my judgment as the filter.
• Distilled 10 products and 10 studies into 5 evidence-backed trust principles that drove every design decision.
• At FinOps X 2026, customers and partners repeatedly called PointFive's NG agent direction a "game changer" and asked how soon they could use it.
• In an AI-native world, a designer's value isn't producing screens fastest — it's being the layer of judgment: asking the questions no one else asks, understanding the user, finding the holes in a flow before they reach development.