Production and case studies
AI interaction models
Chat, inline assist, copilot, autonomous agent or background batch. Pick the shape that matches stakes and how structured the task is.
Wrong shape is the most expensive UX bug: a full agent UI for a single-field form, or a bare chat box for irreversible account changes. Pick the interaction that matches frequency, stakes and how much autonomy users want.
Concept
Before you start: product strategy decides whether AI belongs at all. This page is the interaction once you have decided to ship.
The big idea
| Model | Feels like | Best when | Failure mode |
|---|---|---|---|
| Chat | Conversation thread | Open-ended tasks, exploration | Endless chat as a junk drawer for every feature |
| Inline assist | Ghost text or sidebar in the artifact | Writing, code, sheets — user stays in flow | Over-aggressive suggestions interrupting typing |
| Copilot | Sidekick with tools on a canvas | Multi-step work with a durable document | Opaque actions the user cannot undo |
| Autonomous agent | Goal in, result out | Clear goals plus verifiable outcomes | Silent wrong side effects |
| Background automation | No UI until exception | High volume, low ambiguity pipelines | No path to notice or correct errors |
Default to copilot / assist when stakes are high or taste matters (code merge, customer email). Use agents when the task is ambiguous but outcomes are checkable (tests pass, form validates). Use chat when the job really is dialogue — support, tutoring — not as a lazy wrapper around a form.
Many products mix models: chat for discovery, inline edit for the artifact, human approval before irreversible tools.
Walking through a refund
"Design the UX for an agent that can refund customers." Chat alone is the wrong shape. The irreversible action needs a confirmation, an undo window and a visible tool trace. Assist-plus-approval is the product: the model drafts refund $40 on order 1842, the human clicks, the system records who approved.
A single-field "reason for return" form does not need an agent canvas. A million-document extraction pipeline does not need a chat thread. It needs a batch job plus an exception queue.
Common mistakes
- Chat as the default for every feature. Forms stay forms.
- An agent UI with no review when the outcome is not verifiable.
- Background automation with no exception path. Errors ship silently.
Check yourself
- A user is editing a customer email. Which model? Inline assist or copilot, so they stay in the draft.
- Outcomes cannot be checked automatically. What do you require? A human in the loop before side effects.
Key takeaways
- Match chat, inline, copilot, agent or batch to task structure and stakes.
- Assist when taste or risk is high. Agents when outcomes are checkable.
- Mix models in one product. Chat finds; inline edits; a gate sits on writes.
- Background jobs need an exception queue, not a chat box.
Further learning
Once the shape is right, calibrate what users believe in calibrated trust.
Related
Designing AI product experiences
Interaction design for probabilistic systems: trust, control, failure states and human-in-the-loop.
Agents and agent architectures
Workflows vs agents, the core patterns, ReAct loops, planning and knowing when to stop.
AI product strategy
What to build with AI, how to defend it, how to price it, and how evals become product requirements.