The Third State Between Empty and Set
TRIGGER
When AI suggestions for labels, assignees, and other properties look identical to human-set metadata, users lose track of what's authoritative versus speculative. This creates confusion about the source of truth and erodes trust when AI suggestions turn out to be wrong.
APPROACH
Linear's team built Triage Intelligence with a dedicated module in the triage view that uses Linear's existing visual language for suggestions (assignee, labels, projects, related issues) while maintaining clear provenance distinction. Input: AI-generated property suggestions with confidence scores and reasoning. Output: visually distinct suggestion chips that transform into standard metadata upon acceptance. The UI shows a "thinking state" with timer during multi-step reasoning, hover reveals plain-language explanations and alternative suggestions, and a full thinking panel displays the model's research trace (context pulled, decisions made, guidance applied). This transparency design supports the opt-in automation path where trusted suggestion types can be auto-applied while maintaining visibility.
PATTERN
“Treat AI output like human input and you'll ship 'bugs' that were never bugs—just unreviewed suggestions users assumed were facts. AI suggestions occupy a third state between 'empty' and 'set': proposals. The UI must make provenance visible so users know what to trust and what to verify.”
✓ WORKS WHEN
- AI suggestions coexist with human-set values on the same entities
- Users need to quickly distinguish authoritative data from AI proposals during review
- Suggestions can be wrong and accepting incorrect ones has meaningful consequences
- The product has an existing visual language that AI elements must integrate with
- Workflows involve both reviewing AI suggestions and setting values manually
✗ FAILS WHEN
- AI is the only source of the data—no human-set alternative exists to distinguish from
- Suggestions are high-confidence and verification isn't expected (fully automated pipelines)
- Adding visual distinction creates clutter that outweighs the transparency benefit
- Users don't care about provenance, only correctness (consumer-facing products)
- All values must go through explicit approval anyway (distinction is redundant with workflow)