A fatal incident is a breach serious enough that the rest of the scorecard stops mattering. The agent may have opened correctly, listened well and closed properly — and still handled the call in a way no business would defend.
Most QA programmes add this category eventually, usually after a call reaches someone senior and the scorecard says 89%.
What usually counts
The exact list is yours to set, but these appear in almost every programme:
- Abusive or hostile language toward the customer.
- Dismissive or condescending delivery — sarcasm, mockery, audible impatience. Often more damaging than an outright insult because it is deniable.
- Refusing to assist without a legitimate reason.
- Ending the call improperly — hanging up on a customer mid-conversation.
- Prolonged unresponsiveness — leaving a caller talking to silence.
- Compliance breaches — a required disclosure omitted, or information given that should not have been.
Why they cannot be a deduction
The instinct is to make a fatal incident cost a lot of points. That does not work, because averaging absorbs it.
| Agent | Calls | Average score | Fatal incidents |
|---|---|---|---|
| A | 180 | 91% | 0 |
| B | 180 | 91% | 3 |
| C | 180 | 84% | 0 |
Agents A and B are indistinguishable on score. B is the problem. C looks worse than both and is simply less polished. Any system that reports only the average will route coaching to C and miss B entirely.
Fatal incidents are a count, not a percentage. The moment they become part of an average, they stop being visible.
The detail that generates false accusations
"The agent ended the call" and "the customer ended the call" produce nearly identical audio: a conversation, then a disconnect. Systems that cannot tell them apart flag the agent.
That is a serious failure mode. Being wrongly accused of hanging up on a customer is the fastest way to destroy an agent's trust in a QA programme, and once a floor decides the system is unfair they stop engaging with all of it.
Xperia takes hangup attribution as an input and uses it to suppress false flags, and it excludes legitimate transfers to another agent from fatal detection. It also accounts for call direction, so an agent is not penalised for behaviours irrelevant to an outbound call. More on how fatal incident detection works.
Tone is the hard part
Explicit abuse is comparatively easy to detect. The common case is harder: an agent who says nothing objectionable while making it entirely clear the caller is wasting their time.
Detecting that requires evaluating delivery, not words — pitch, pacing, clipped responses, the particular flatness of forced politeness. A transcript-based system cannot see any of it, which is why transcription-then-scoring pipelines under-report the most common form of poor service.
What to do with the count
A fatal incident is not a coaching note. Treat it as an event:
- Review the recording. Every time, with no exceptions — automated detection is a filter, not a verdict.
- Confirm or clear it. A supervisor should be able to override the flag, and the original detection should remain on record either way.
- Handle it as a conduct matter, not a QA score. It belongs in a different conversation from parameter coaching.
- Track the rate, not just the total. Three incidents across 2,000 calls and three across 200 are different situations.
Dead calls are not fatal incidents
Worth separating: a recording containing only IVR audio, hold music, or a call that never reached an agent is not a quality failure. It is not a call. Scoring it produces a meaningless low number that drags down an average.
Xperia marks these as dead calls and excludes them from scoring, which is usually the first thing that makes a team's dashboard match their intuition.
Fatal incidents are rare, expensive, and invisible to averages. They need their own category, their own count, human review before anyone acts, and enough coverage to be detected at all — a 2% sample will miss most of them.
Get the attribution right too. A system that accuses agents of hanging up on customers they did not hang up on will lose the floor's trust faster than it earns anything.