Accuracy isn’t the metric, recoverability is

By Prajwal · May 30, 2026 · 2 min read

Key takeaways

• Every model hallucinates; even grounded systems never reach zero, so accuracy always has a tail.
• Two systems at 99 percent accuracy can behave very differently on the 1 percent that goes wrong.
• Recoverability is what the system does at the tail: detect, escalate, close the gap, and update before the customer churns.

Accuracy is the metric every AI-for-support vendor leads with, and it is the one most likely to mislead a buyer into thinking a system is safer than it is. The reason is simple: the model is going to be wrong. Every model hallucinates. Published research on ungrounded LLMs in customer service puts hallucination rates between 15 and 30 percent of responses. Even grounded systems on their best day sit at 0.7 to 1.5 percent, which still means tens of thousands of wrong answers per million tickets at enterprise scale. Buying for accuracy alone is buying a metric that will always have a tail. The real question is not how often the AI is right. It is what happens the moment it is wrong. Recoverability, not accuracy, separates a tool from a system.

The accuracy ceiling is structural

Hallucination is not a bug in the formal sense. It is a consequence of how LLMs generate text: token by token, returning the most-likely sequence rather than the most-true one. When those diverge, the output is fluent and wrong at once. A peer-reviewed 2025 study of LLMs in customer service found hallucinations in 31.4 percent of real-world cases, rising to 60 percent in complex domains, and 63 percent of production systems showed dangerous hallucinations within their first 90 days. Guardrails help. Grounded retrieval, system prompts, verification pipelines, and real-time monitoring can cut hallucination risk by 71 to 89 percent, and published guardrail benchmarks reach a 97 percent detection rate. None of it reaches zero. The structural ceiling is always above zero. A vendor promising 100 percent accuracy is either misstating the rate or misstating what it measures.

Why accuracy is the wrong selection metric

Two systems with identical 99 percent accuracy can behave very differently on the 1 percent tail. System A confidently delivers the wrong answer, the customer accepts it, and the failure surfaces in a chargeback three weeks later. System B detects the low-confidence case, escalates to a human, the customer recovers, and the knowledge gap that caused the failure closes within a week. Both read “99 percent accurate” on the demo slide. Only one is fit for production. Buyers anchor on accuracy because it translates straight from machine-learning research, where benchmarks are the currency of model comparison. The translation breaks in customer service, because the cost of a wrong answer is not the cost of a wrong token. It is the cost of a customer who never came back. That gap is what recoverability fills.

What recoverability looks like as a system

Recoverability is the property of a system that, when wrong, detects the error, surfaces it to the right owner, closes the underlying gap, and updates its own behavior before the customer churns. In Auralis Audit it lives as four instrumented signals. Confidence is calibrated against historical correctness, so “90 percent confident” actually means wrong one time in ten. Anything below the threshold is routed human-in-the-loop, not auto-resolved. Every closed conversation is scored after the fact against accuracy and against signals of dissatisfaction: reopens, escalations, second-channel contact, and churn within 30, 60, and 90 days. Detected errors trigger a knowledge-base gap that Auralis drafts, the customer reviews, and goes live within the week, so the next ticket in that category routes against corrected content. Accuracy is the number we report. Recoverability is the property we tune for.