"The agent gave a wrong answer" is not a complaint but a request for diagnosis. And starting the diagnosis in the wrong place costs hours.
This is the final article of the series and a checklist: the four possible sources of a wrong answer, and the order to check them in.
First, pin down the case
The only thing to do before diagnosis is making the case reproducible:
- Which call? You need the recording and transcript.
- What exactly did the customer ask? The wording matters, not a paraphrase.
- What exactly did the agent say?
- What should the answer have been?
Any investigation without those four written down is guesswork.
Source 1: a rule in the instructions
Check here first, because it is the most common cause and the fastest to fix.
A significant share of wrong answers comes not from the agent failing to find information but from a rule blocking it. A rule like "when asked about price, ask about budget first" turns a simple pricing question into a three-turn interrogation, and the customer never gets an answer.
Symptom: the agent clearly knows the topic but will not say it; it circles the answer.
Source 2: contradiction or gap in the knowledge base
Information comes second. Two distinct situations live here, with different symptoms:
- Missing information. The agent says "I do not have that" or speaks generally. Fix: add the document.
- Contradictory information. The agent answers differently each time. Fix: delete the old document.
The second is sneakier, because a single test can return the right answer and the issue looks closed.
Source 3: the update never arrived
If the information is right and the document current but the agent still gives the old answer, the problem is content processing. A changed document has to be reprocessed; until it is, the agent keeps seeing the old version.
Symptom: open the document and the correct information is right there, while the agent insists on the old one.
Source 4: the model
It sits last on the list because it is the rarest cause. Model-related errors usually appear as: misreading numbers, confusion in long lists, losing context in very long calls.
The fix is usually not changing the model but shrinking the job asked of it: moving long lists into a tool, taking calculation away from the agent, shortening the call.
Two traps when testing
Testing inside an existing conversation. The conversation context carries the old answer; every test needs a fresh conversation.
Asking the question the way you know it. Because you know the answer, you phrase the question correctly. Customers do not. Test with real sentences taken from real transcripts.
Verifying the fix
Before declaring a fix done, ask the same question three times and confirm the answers are consistent. A single correct answer does not show the problem is solved — only that the answer was right once.
End of the series
Across this series we went from the agent's voice to campaign operations, from sector applications to data structure and reporting. The common theme: the value of a voice agent lies not in the model but in the operation built around it.
You can explore how the voice agent works on the AI call center page, or hear it directly by having it call you from the demo page.