Voice agents can't get the details right — and nobody is measuring it

The detail problem

Sierra built τ-Entity to ask a simple question: can a voice agent capture a detail correctly, recognize when it might be wrong, and repair the mistake? The benchmark runs 200 tasks across ten entity types. The results are sobering: GPT Live 1 captures values correctly in 60.3% of regular calls, Gemini 3.8 Live reaches 43.8%, and Voice Think Fast 2.0 reaches 44.2%. When all three test conditions must pass, GPT Live 1 succeeds on 34.0% of tasks. As Sierra's Soham Ray put it: "Even in this simple setting, today's voice agents struggle to get the details right." m:…0973

Worse than the miss rate is the recovery rate. Only 24-37% of initially incorrect captures are ultimately corrected by the agents — asking a follow-up question, Ray notes, is only useful if the agent can incorporate the answer correctly. Multi-field calls compound the problem: three-field tasks have a 53% success rate, though validating each field and allowing retries lifts multi-field success from 64% to 82%. m:…0973

The measurement problem

The deeper problem is that most firms would not know. Parloa's 2026 CX Observability Study, conducted with Metric Sherpa, found that 78.5% of operators whose organizations use AI have no metric for hallucinations or factual accuracy. Traditional CX metrics — speed, containment, transfer rates — do not measure whether the agent's answer is correct. Yet 94.7% of respondents trust that their current metrics reflect the real customer experience, and blind spots persist at roughly the same rate across all stages of AI adoption. As Parloa's Alexa Krzyzanowski put it: "Accuracy is the least-measured AI risk." m:…2961

The gap runs all the way up. Only 27.9% of executives report that hallucination or factual-accuracy data reaches them, and 76.6% of operators who say they are very confident in their AI still lack accuracy metrics. In regulated industries, where rules like the EU AI Act are raising expectations that companies prove their AI is accurate, 79.3% of operators report no accuracy metric at all. m:…2961

The fix layer arrives

The tooling to close the gap is starting to ship. Decagon launched Experiment Hub, an experimentation platform for AI agents that lets teams test one change while the rest of the agent keeps shipping updates. It supports up to 20 variants against a shared control group, measures impact with over 20 conversation-specific metrics, and provides automated analysis with verdicts, explanations, confidence intervals, and significance testing — automatically stopping tests when a variant performs significantly worse than the control. m:…5376

On the detection side, n8n's guidance is to stop looking for one metric and layer several: faithfulness checks on whether claims are supported by context, self-consistency across multiple responses, semantic entropy on whether responses mean the same thing, and source verification against authoritative databases or APIs. Hallucinations, n8n notes, are hard to catch because they resemble normal, coherent answers — so confidence thresholds should route high-risk outputs to human-in-the-loop review, with evaluation checks built directly into the workflow canvas. m:…8326

What executives should do

The two findings together define the decision. Sierra shows verification works but costs time: prescribed verification strategies raise success 14-31 percentage points while adding 21-28 seconds per call. Parloa shows most organizations are not measuring what verification would protect. The implication is that the tradeoff is now explicit and priced — accuracy measurement, verification seconds, and human-review routing are line items to budget, not afterthoughts. Scale the agents, but instrument the answers first. m:…0973 m:…2961