Your AI feature needs a UX pattern, not just a chat box

Somewhere in your product right now there's a sparkle icon. Click it and a chat panel slides out, empty, with a cheerful placeholder: "Ask me anything." Your team shipped it because every competitor shipped one, usage spiked for a week, and now the analytics show what the industry already knows: the panel gets opened once, produces one mediocre answer, and gets ignored. The problem isn't the model. It's that a chat box is one UX pattern, and it got treated as the whole discipline.

What is agentic UX?

Agentic UX is the design work for software that acts instead of just responding. Jakob Nielsen calls AI the first genuinely new UI paradigm in roughly 60 years: interaction shifts from commands (click this, fill that) to intent (here's the outcome I want), and the software plans and executes the steps. Gartner puts task-specific agents inside four in ten enterprise applications by the end of 2026. That's a lot of software making decisions on behalf of users, designed mostly by teams whose last new pattern was the modal dialog.

The stakes are concrete. Most workplace-AI trust surveys land near the same number: roughly seven in ten employees say they won't use an AI tool they don't trust. Trust isn't a model property. Models are probabilistic and occasionally wrong no matter what you buy. Trust is a design property, built from how the product exposes what the AI is doing, and that's exactly the part a bare chat box hides.

Why the chat box fails as a default

Three structural reasons:

• The blank-page problem. An empty text field puts the entire burden of imagining what the AI can do on the user. Most users type one generic question, get one generic answer, and file the feature under "toy."
• No capability boundaries. Chat implies the system can do anything, so users ask for things it can't do, watch it fail, and downgrade their trust in the things it does well.
• Hidden state. When an agent works through a multi-step task behind a spinner, the user can't tell progress from a hang, or a considered answer from a hallucination. Opacity reads as risk.

Chat still has a place: exploratory questions, low-stakes drafting, support. As the single front door to your AI capability, it's a dead end.

What UX patterns does a trustworthy AI agent need?

Six patterns cover most of what we ship for B2B products:

1. Scoped entry points. Put the AI where the work is: "Draft the renewal email" on the account page, "Explain this anomaly" on the chart. Suggested actions in context beat an omniscient empty box, because they teach capability while being used.
2. Plan before action. For anything multi-step, the agent shows its plan first: the steps it intends to take, the data it will touch. The user approves the plan, not each keystroke.
3. Visible progress. Stream the steps as they happen, with a log the user can expand. "Reading 14 invoices, found 2 discrepancies, drafting summary" is a trust engine. A spinner is a trust drain.
4. Approval gates matched to risk. Reversible and internal actions run free. Irreversible or outward-facing ones (send, delete, charge, publish) stop and ask. The gate placement is the product decision, and it deserves design review, not a default.
5. Undo as a first-class citizen. Every agent action a user can see needs a path back. Recoverability is what lets people delegate without anxiety, and it's also what makes gate number four tolerable to loosen over time.
6. Sources and confidence. When the agent asserts something, show where it came from. When it's unsure, say so in the interface, not in a disclaimer nobody reads.

Should every agent action require approval?

No, and over-asking is its own failure. Approval fatigue turns confirmation into a reflex within a day, and a reflex approves mistakes as readily as it approves good work. The design question isn't "how do we stay safe" but "which actions have a blast radius that earns an interruption." Match the gate to reversibility and visibility, then loosen deliberately as the agent earns a track record the user can inspect.

One more thing this changes: your design system. Agent plans, progress logs, approval sheets, and undo affordances are components, and they belong in the system with tokens and states, not rebuilt ad hoc per feature. We covered how that discipline holds up in design systems developers actually use, and the strategy behind the agents themselves in AI agents for B2B.

This seam between product design and applied AI is where our Studio practice and AI practice work as one team, because the seam is where trust gets designed. The DGTL Readiness Index scores that maturity as one of eight dimensions, if you want the wider picture.

Related: Design systems developers use → · AI agents for B2B → · Brand design that converts →