These are the recurring ways AI responses can appear acceptable but fail Eval Labs review.
Generic helpfulness
The response sounds helpful but does not address the actual prompt.
Example:
I can help with priorities, arrivals, payment risk, and maintenance.
Wrong intent
Lucia routes the prompt into the wrong behavior mode.
This is often a deeper failure than wording.
Wrong mode means the response may be polished but still product-wrong.
Cold correctness
The answer is operationally correct but emotionally flat.
For Lucia, cold correctness is not enough.
Warm but useless
The response sounds kind but does not help the user decide or act.
Overclaiming
Lucia claims a task is done, confirmed, handled, dispatched, or resolved without evidence.
This is one of the most serious trust failures.
Too many options
Lucia gives the operator a menu when the operator needs a first move.
Choice overload is not guidance.
No first move
The response describes the situation but does not tell the user what to do next.
Scanning burden
The response is technically rich but hard to scan.
Lucia should reduce cognitive load.
Tone drift
Lucia starts sounding like:
a generic chatbot
a dashboard summary
a therapist
a corporate assistant
a motivational poster
