AI products are shifting from click-driven UI to attention-driven interfaces. Inspired by Latent Space’s “The Attention Interface”, here’s how to turn that idea into shippable UX patterns for your app.
Why attention is the new UI
Traditional UI asks users to navigate. Attention interfaces let users declare intent, then the system assembles context, tools, and next steps.
Done right, this reduces friction, increases trust, and makes AI feel like a focused teammate—not a chat box.
7 practical patterns to ship now
- Pin & scope context: Let users “pin” entities (docs, rows, tickets) so the model’s attention is grounded. Example: chips for “Q3 plan” and “ACME account” visible above the assistant.
- Preview before action: Always show a dry‑run. For edits, display a side‑by‑side diff; for automations, list planned API calls. Make “Apply” a deliberate step.
- Cite sources & uncertainty: Inline citations to files, rows, or URLs. When confidence is low, show a short rationale or alternatives instead of a single answer.
- Correct & teach back: Allow quick fixes (“That’s the wrong client”) and reflect the new understanding. Keep an editable system note or “assumptions” panel.
- Guardrails on commit: Add bounded actions: limits, approval requirements, and undo. For risky ops, require a second check or policy‑based blockers.
- Memory with consent: Separate session memory from durable memory. Let users save a “profile” or “project brief” explicitly; show and delete memory items easily.
- Traceability & logs: Provide an inspectable trace: prompt, context, tools used, and outputs. Useful for debugging, audits, and trust.
Concrete examples
- CRM copilot: Pins = account + opportunity; Preview = email draft + CRM field diffs; Commit = update stage + send email with undo.
- Spreadsheet AI: Pins = selected range; Preview = formula explain + result preview; Guardrails = scoped to sheet with change history.
- Support agent assist: Pins = ticket + policy; Citations = knowledge base refs; Trace = tools invoked (summarize, translate, retrieve).
What to measure
- Time‑to‑first‑value (TTFV): Seconds from prompt to useful preview.
- Correction rate: % of sessions with successful teach‑backs instead of abandon.
- Approval/commit conversion: Previews that become applied changes.
- Rollback/undo rate: High rates signal risky defaults or unclear previews.
Trust and UX guidance
Design for legibility and control. Small touches—visible pins, clear diffs, and inline citations—go further than bigger models.
For broader AI UX principles, see Nielsen Norman Group’s overview on Designing with AI.
Key takeaway
Treat attention as a first‑class interface: let users set scope, preview everything, cite sources, and commit with guardrails. That’s how AI earns trust—and usage.
Like nuggets like this? Subscribe to our newsletter for weekly, practical AI insights: theainuggets.com/newsletter.
