Use this when
An AI step can end up uncertain or in an unexpected state, and you need a defined fallback that does the least harm.
The pattern
UNCERTAIN or UNEXPECTED?
↓
TAKE THE SAFE DEFAULT
the option that does the least harm if it is wrong
↓
(usually: do nothing, hold, or escalate to a person)Rather than letting the AI improvise when it is unsure, you define in advance what happens: the choice that is safest to be wrong about.
Why it works
When an AI is uncertain, the riskiest thing it can do is act confidently anyway. A safe default replaces that with a predefined, low-harm response. The safe default is usually to do nothing consequential: hold the case, take no irreversible action, and hand off to a person. That way, an uncertain moment cannot turn into a costly action.
It is the calm answer to "what should happen when we do not know what should happen." It works hand in hand with confidence thresholds and escalation design.
How to apply it
- For each step, ask what the least-harmful option is if the AI is wrong or unsure.
- Make that the defined fallback, not an improvised choice.
- Prefer defaults that avoid irreversible actions: hold, decline, or escalate.
- Make sure the safe default routes to a human fallback where a person is needed.
Good fit
- Any step that acts, where being wrong has real consequences.
- Uncertain classifications, extractions and decisions.
Poor fit
- Purely informational, reversible steps where any reasonable guess is harmless and stopping adds friction.
What this does not solve
A safe default limits harm when the AI is unsure; it does not reduce how often it is unsure, and an overly cautious default can send too much to people. Combine it with improving the step and tuning the threshold, so the safe default is the exception, not the norm.
