Automatic issue detection
Automatic issue detection is the machine-driven flagging of problems across a stream of phone calls. Instead of waiting for a human to notice a bad conversation, the system evaluates calls as they complete — looking for patterns such as dead air, repeated loops, questions the agent could not answer, failed tool calls, abrupt hang-ups, or callers asking for a human — and groups recurring findings into named issues an operator can triage.
The output is a worklist rather than a score: each issue describes the pattern, points to example calls as evidence, and — in more capable systems — proposes a fix, such as a prompt adjustment or a missing knowledge-base answer. Issues have a lifecycle (open, acknowledged, resolved) so the same pattern is not re-reported forever.
Detection differs from its neighbors. Call quality assurance is a review process, sampling calls against a rubric. Alerting fires on threshold breaches in aggregate metrics. Issue detection sits between: it reads individual conversations at scale and promotes recurring problems to first-class objects.
Why it matters for AI phone calls
No team reviews every AI call; at even modest volume, unsampled human review is impossible. That leaves a gap where a failure mode can run for days — an agent mispronouncing the business's name, a tool silently erroring for one class of request — hurting a small share of calls that never shows up in a topline success rate.
Automatic detection closes that gap by putting every call under some level of inspection. Its limits deserve respect: detectors judged by language models inherit their blind spots, clustering can merge distinct problems or split one problem into several, and a proposed fix is a hypothesis, not a diagnosis. Treating detected issues as leads for a human to confirm — with the example calls as evidence — keeps the system honest.
In practice on ThunderPhone
ThunderPhone's Issues page (under Quality & Testing) collects problems detected across calls and groups repeat occurrences into patterns — fifty instances of the same mistake show as one row with a trend line, not fifty reports. Issues are raised automatically by the platform's call analysis, with example calls attached as evidence and proposed fixes for the operator to review.