FG—01AI Design Field Guide

Search the field guide

Search chapters, principles, patterns, worksheets, and glossary terms

PTN—41

Graceful degradation

When the AI can’t perform — offline, rate-limited, low confidence, out of scope — fall back to something still useful.

AI features fail more ways than traditional software: the model is overloaded, the context is too long, the question is out of distribution, the confidence is garbage. Graceful degradation is deciding, at design time, what each failure hands back to the user. The floor should never be a spinner that gives up or a cheerful non-answer; it should be a smaller, honest capability — search instead of synthesis, a template instead of a draft, the raw documents instead of the summary.

The design move is a capability ladder. Full capability at the top; below it, degraded modes that still move the task forward; at the bottom, the manual path, which must always exist and always be visible. Products that bet everything on the model — no search fallback, no manual editor, no cached results — convert every model hiccup into a total outage of the user’s task, not just of your feature.

Degradation must be legible, not cosmetic. If the fast model silently substitutes for the good one under load, users get inexplicably worse answers and blame themselves — or stop trusting the good days too. A one-line banner (“High demand — responses may be slower / using a faster model”) costs a moment of polish and buys back the whole relationship. Failing quietly is the expensive option.

When to use

Design considerations

Pitfalls

In the wild

Perplexity
Grounded in search: even when synthesis is weak, the cited sources remain useful — the degraded mode is a good search engine, not an empty state.
Claude
Declines out-of-scope or beyond-cutoff questions by stating the limit and suggesting how to get the answer, rather than improvising one.
GitHub Copilot
Fails soft by design — when no completion arrives, the editor is just an editor. The manual path is the default surface, so degradation is nearly invisible.
Google Search (AI Overviews)
Simply omits the AI overview on queries where it isn’t confident, degrading to classic results instead of forcing a synthesis.

Principles served

Discussed in the guide