A generated flow can look complete because every screenshot shows a stable moment: form filled, result returned, dashboard populated. Production is mostly the transitions between those moments.
A product state is the information the interface needs to represent when the expected path is not simply “ready → action → success.” In AI workflows that space is larger because the system can be slow, uncertain, partially successful or capable of acting beyond the current screen.
Eight states worth checking before launch
- Initial. What can the user do before any data exists?
- Loading / working. What is happening, and can the user safely leave?
- Empty. What does zero data mean, and what is the next useful action?
- Permission. What is blocked, why, and who can resolve it?
- Uncertain. What does the system know poorly enough that the user should verify?
- Partial completion. Which parts succeeded, which failed and what remains actionable?
- Failure. What exactly failed and what can the user do next?
- Recovery. Can the user retry, undo, correct, restore, escalate or hand off?
AI adds two state dimensions traditional CRUD products can avoid
The first is epistemic state: confidence, source quality or ambiguity may affect whether a result should be acted on. The second is agency state: an AI feature may be suggesting, preparing, waiting for approval, executing, paused or already committed.
When those dimensions remain invisible, users have to infer too much. A single “Generate” button can hide a chain of decisions about scope, tool use, external actions and irreversibility.
| Bad generic state | Better product state |
|---|---|
| “Loading...” | “Checking 14 invoices · you can leave this page” |
| “Done” | “12 updated · 2 need review” |
| “Error” | “Could not access Drive · reconnect or upload manually” |
| Disabled action | State why the action is unavailable and how to unlock it |
State debt grows faster than screen debt
Every new action multiplies possible states. If teams keep generating new screens without stabilizing the shared state model, inconsistency becomes structural: one flow uses inline errors, another uses toasts, another silently retries, and a fourth simply fails.
The fix is not to design every theoretical edge case. Start with high-frequency and high-consequence failures. Then turn their interaction logic into reusable product rules.
A state audit can be smaller than a redesign
Take one revenue or activation-critical flow. Walk it with bad network, missing data, denied permissions, repeated actions and an intentionally wrong AI result. Record each point where the interface stops explaining what happened. Those gaps form a concrete backlog.
This is why product readiness is not synonymous with visual polish. A coherent state model often improves perceived quality more than another pass on typography or shadows.