The agent did not fail.
The microphone did.
A quality system that cannot say "I could not tell" will say "no" instead, and the difference between those two answers is somebody's performance review.
9 September 20264 min read
Where this shows up
A rubric asks whether the agent disclosed the price. The system returns a value per call, and the dashboard aggregates it into a compliance rate per agent.
Now take a call where thirty seconds are unintelligible — a bad line, background noise, the customer's phone cutting out. The price may well have been disclosed inside those thirty seconds. The system does not know.
If the data model has two states, the answer is "no". The agent's compliance rate drops for a call in which they may have done everything right, and nothing in the report distinguishes that case from a genuine failure.
Why it is not a rare edge case
Contact centre audio is mono, compressed, and recorded over a telephone channel. Some fraction of every batch is partly unintelligible, and that fraction is not evenly spread: it clusters by line quality, by shift, by whichever queue takes calls from the worst connections.
So the bias is not random. It falls hardest on the agents who happen to take the noisiest calls, which is the opposite of what a quality programme is for.
A client caught this in a demo. It is the kind of defect that survives testing because the sample used to build the system is usually clean audio, and it only appears at the scale where bad recordings are guaranteed.
The fix is in the schema, not in the model
It is tempting to treat this as an accuracy problem — better transcription, a better model. It is not. No amount of accuracy recovers information that is not in the audio.
The fix is that the value has three states rather than two: yes, no, and not determinable. Then the dashboard can do the only correct thing, which is to compute compliance over the calls where the answer was determinable, and report separately how many were not.
That second number is useful on its own. A queue with 12% non-determinable calls has a recording problem, and that is worth knowing regardless of what anyone's compliance rate says.
What to ask a vendor
Ask what the system returns when it cannot tell, and ask to see it in the export rather than in the demo. A model can be prompted to answer carefully; a column either has three values or it does not.
Then ask what the dashboard does with that third value. Counting it as a failure and dropping it silently are both wrong, in opposite directions — and the second one is harder to notice, because the number that appears looks perfectly reasonable.