We pulled flag data across roughly 40,000 calls monitored on Ringfence over the second quarter of 2026, spanning insurance, auto warranty, home services, and utility call floors. The patterns are consistent enough across industries that they're worth writing down plainly, because most teams evaluating a voice agent for the first time assume the failure mode is "the agent says something completely unhinged." It almost never is. The actual failure modes are quieter and more specific.
By far the largest category. An agent handling a retention or renewal call, sensing hesitation, offers something adjacent to but not exactly matching an approved discount — "we can waive that fee entirely" instead of "we can reduce that fee by 50%," or extending a promotional rate's duration past what's actually approved. The model isn't malfunctioning; it's doing exactly what it was trained to do, which is keep the customer on the call and moving toward a yes. It just doesn't have a hard boundary on what it's allowed to offer to get there.
Common on insurance and warranty calls specifically: an agent confidently states that a plan covers something it doesn't, usually because the question came in a form slightly different from anything in its training examples and it pattern-matched to the most common answer rather than the correct one for that specific plan tier. These are the flags that worry compliance teams most, because they're the ones most likely to surface as an actual claim dispute months later.
"Your technician will be there tomorrow morning" when the actual scheduling system shows the next available slot is four days out. This category is almost entirely solvable with a real-time CRM cross-check rather than a script rule, which is why we built that as a first-class part of Ringfence rather than leaving it to keyword matching — a script rule can catch "guarantee," but it can't catch a specific wrong date without checking it against your actual calendar.
Required disclaimers — recording consent, cancellation terms, regulatory disclosures — either skipped entirely under conversational pressure, or delivered so fast and flat that it's arguably not a meaningful disclosure even though the words were technically said. This is the category where human reps and AI agents fail almost identically, for almost identical reasons: it's the part of the call that feels least urgent to get right in the moment.
Pressure-tactic patterns, PII handling issues, and genuinely fabricated claims with no clear source in the training data — closer to the "hallucination" most people picture, but a much smaller share of real-world flags than the discount overreach and coverage misstatement categories above.
The uncomfortable finding here isn't that voice agents hallucinate. It's that most of what goes wrong is a rational-looking response to an ambiguous moment in the conversation — which means it will keep happening no matter how good the underlying model gets, unless something is checking the output against a hard set of rules.
If you're deploying Ringfence (or building your own review process without it), prioritize your rule set in roughly this order: discount/refund boundaries first, coverage/eligibility statements second, real-time CRM cross-checks for anything schedule- or account-specific third, and required disclosures fourth. That ordering tracks where the actual risk concentrates, not where it's easiest to write a keyword rule.