# Outbound AI Voice Benchmarks: Why 90% Task Rate and 300ms Latency Aren't Comparable

Dr. Samuel Ortiz · November 13, 2026

> A blended 90% task rate can mask failing high-risk intents. Report completion across 5 intent strata and audit 300ms latency with reproducible logs.

Report task completion separately for at least five intent strata: confirm, reschedule, cancel, billing question, and transfer-to-human. A blended 90% can hide a failing high-risk intent when high-volume, low-risk intents dominate the numerator. A high-risk intent with low volume can be invisible in a blended rate.

A reproducible audit should expose the call set, attempt log, disposition timestamps, and intent labels so a reviewer can recompute both the clock and the denominator. Publish the stratified counts and timestamps, not only the blended percentage. Without those artifacts, a 90% task rate and a 300ms latency figure describe different measurements and cannot be compared.

![Report task completion separately for at least five — Outbound AI Voice Benchmarks](https://static.mm-ais.com/article-images-ai/outbound-ai-voice-benchmarks-why-90-task-ai-f160bbe5.jpg)

## Three-source audit kills the 90/300 myth

The sources reviewed here do not report an outbound voice-agent result that pairs a task-completion rate with a sub-second latency figure on the same call set. Treat any such pairing as unproven until a reproducible audit discloses the measurement conditions behind both numbers. A three-source check shows why: each source contributes a disclosure practice or a cost anchor, but none supplies outbound voice performance evidence.

According to Artificial Analysis, evaluating Gemini 4 Hydrogen (xhigh) costs $3.16 per task. Use that figure only as an audit-budget anchor when planning how many scored runs you can afford. It is not evidence of voice latency, not evidence of task completion, and not a proxy for production call performance. Before accepting a vendor claim, require a line-item accounting of how many tasks were evaluated, what each task represented, and whether the reported evaluation spend covers the full outbound attempt set.

Turing reports that ServiceNow’s EnterpriseOps-Gym used 1,000+ enterprise prompts to benchmark long-horizon planning, tool use, and policy adherence. Copy the disclosure requirement, not the domain result: a credible outbound audit should state the prompt or scenario count, the intent categories covered, and the policy checks applied to each interaction. Without that level of prompt-count and scope disclosure, a headline completion rate cannot be compared across agents or across test runs.

Apply a simple audit rule: if a source cannot supply outbound task evidence, borrow only its disclosure discipline or its cost anchor. Budget the evaluation using per-task cost data, document scenario coverage with explicit prompt counts, and demand instrumented latency records for the same calls being scored. Any claim that cannot survive this three-source check should be treated as marketing, not benchmark evidence.

![Three-source audit kills the 90/300 myth — Outbound AI Voice Benchmarks](https://static.mm-ais.com/article-images-ai/outbound-ai-voice-benchmarks-why-90-task-ai-861979dd.jpg)

## Compare p95 SLO, not mean latency

The production comparison is not a latency number alone: the winner is a p95 end-to-end SLO paired with an intent-stratified task rate, while mean latency and first-token latency lose because each can look good while hiding tail failures or non-model delays.

| Option | Proves | Hides | Call |
| --- | --- | --- | --- |
| Mean latency | dashboard smoothness | tail failures | loses |
| First-token latency | model responsiveness | trunk, ASR, TTS, writeback | loses |
| p95 end-to-end + intent task rate | production SLO with denominator | nothing if denominator is published | winner |

Mean latency can be dominated by fast calls, and first-token latency can be fast while trunk setup, ASR, TTS, and disposition writeback add delay. p95 end-to-end captures the full path, including the slow tail, so it is the latency measure that can be paired with a task rate.

A 300ms figure must be p95 end-to-end on the same outbound call set as the 90% task rate. If a vendor reports only mean latency instead of p95, or does not disclose the p95, the 90% claim is not production-equivalent because the task rate may exclude slower or failed calls.

Require the failed-task denominator and intent strata to be published with the p95 figure. The task rate should be calculated over all dialed attempts, with completion broken out by intent, so a single aggregate percentage cannot hide which tasks fail or how many attempts never reach a disposition.

![Compare p95 SLO, not mean latency — Outbound AI Voice Benchmarks](https://static.mm-ais.com/article-images-pixabay/outbound-ai-voice-benchmarks-why-90-task-bc38a20c.jpg)

## Cost the audit before the agent

Before comparing outbound agents, price the audit that will prove or disprove a vendor claim. A reproducible audit carries its own cost structure, and skipping that math makes any headline metric difficult to defend. According to Artificial Analysis, Gemini 4 Hydrogen (xhigh) is priced at $1.50 per 1 million input tokens and $7.50 per 1 million output tokens. Those rates establish a token-cost line, but they do not describe the full cost of running an outbound audit.

Convert token prices into per-call cost using transcript tokens and completion tokens per call, not per minute. Multiply the transcript input tokens by $1.50 and divide by 1,000,000, multiply the completion output tokens by $7.50 and divide by 1,000,000, then add the two results: per-call token cost = (transcript input tokens ÷ 1,000,000 × $1.50) + (completion output tokens ÷ 1,000,000 × $7.50). This keeps the unit aligned to the call record under review rather than to wall-clock duration, which can vary independently of how much text the model processes or generates.

## Frequently Asked Questions

**Which intent categories should be reported separately when measuring outbound AI voice task completion?**

Task completion should be reported separately for at least five intent strata: confirm, reschedule, cancel, billing question, and transfer-to-human.

**How can a failing high-risk intent be hidden in outbound voice benchmark results?**

A high-risk intent with low volume can be invisible in a blended rate, especially when high-volume, low-risk intents dominate the numerator.

**What artifacts must a reproducible audit expose for outbound voice benchmarks?**

A reproducible audit should expose the call set, attempt log, disposition timestamps, and intent labels so a reviewer can recompute both the clock and the denominator.

**What specific data should be published alongside a blended task completion percentage?**

Publishers should provide stratified counts and timestamps, not only the blended percentage.

**What did the three-source audit find regarding paired task-completion and latency results for outbound voice agents?**

The sources reviewed do not report an outbound voice-agent result that pairs a task-completion rate with a sub-second latency figure on the same call set.

**When should a pairing of task-completion rate and sub-second latency be considered proven?**

Any such pairing should be treated as unproven until a reproducible audit discloses the measurement conditions behind both numbers.

## Quick answers

| Why can a blended task rate be misleading? | A blended task rate can hide a failing high-risk intent when high-volume, low-risk intents dominate the numerator, and a high-risk intent with low volume can be invisible in a blended rate. |
| --- | --- |
| Which intent strata should task completion be reported for separately? | Task completion should be reported separately for at least five intent strata: confirm, reschedule, cancel, billing question, and transfer-to-human. |
| What should a reproducible audit expose? | A reproducible audit should expose the call set, attempt log, disposition timestamps, and intent labels so a reviewer can recompute both the clock and the denominator. |
| Why can't the prominent task rate and latency figure be compared without certain artifacts? | Without artifacts such as stratified counts and timestamps, the prominent task rate and latency figure describe different measurements and cannot be compared. |
| How should a pairing of a task-completion rate with a sub-second latency figure be treated? | Any such pairing should be treated as unproven until a reproducible audit discloses the measurement conditions behind both numbers, as none of the reviewed sources supplies outbound voice performance evidence. |

Also worth reading: **Enterprise AI’s new frontier: Autonomous agents reshape workflows**: [Enterprise AI’s new frontier: Autonomous](https://enterpriseailabs.io/blog/enterprise_ais_new_frontier_autonomous_agents_reshape_workflows.php) · **Fine-Tune LLMs on Internal APIs: 40% Fewer Hallucinated Imports**: [Fine-Tune LLMs on Internal APIs:](https://enterpriseailabs.io/blog/fine-tune-llms-on-internal-apis-40-fewer-hallucinated-imports.php) · **Driving superior enterprise AI performance with optimization algorithms**: [Driving superior enterprise AI performance](https://enterpriseailabs.io/blog/driving-superior-enterprise-ai-performance-with-optimization-algorithms.php)

### Related reading

- [OpenBioLLM Llama3-Based Models Set New Benchmarks in Biomedical AI Performance](https://enterpriseailabs.io/blog/openbiollm_llama3_based_models_set_new_benchmarks_in_biomedi.php)
- [Analyzing Online C Compilers Performance Benchmarks and Feature Comparisons in 2024](https://enterpriseailabs.io/blog/analyzing_online_c_compilers_performance_benchmarks_and_feat.php)
- [Performance Benchmarks Running a 400B Parameter LLM on Consumer Hardware in 2024](https://enterpriseailabs.io/blog/performance_benchmarks_running_a_400b_parameter_llm_on_consu.php)
- [Stanford BMIR: Retrieval Beats Generation, 62% Latency Gain](https://enterpriseailabs.io/blog/stanford-bmir-retrieval-beats-generation-62-latency-gain.php)
- [2026 LLM Routing: Latency Is Architectural, Not a Serving Artifact](https://enterpriseailabs.io/blog/2026-llm-routing-latency-is-architectural-not-a-serving-artifact.php)
- [Benchmarking Redis Cluster Performance Docker-Compose Implementation Shows 47% Latency Improvement in AI Workloads](https://enterpriseailabs.io/blog/benchmarking_redis_cluster_performance_docker_compose_implem.php)

### Latest

- [How to Verify Phi-4 vs Llama 3.3 70B for Enterprise RAG](https://enterpriseailabs.io/blog/how-to-verify-phi-4-vs-llama-33-70b-for-enterprise-rag.php)
- [Sand.ai's reported $200M round: MAGREF claims stay unverified](https://enterpriseailabs.io/blog/sandais-reported-200m-round-magref-claims-stay-unverified.php)
- [AI test scores deceive: Eval debt inflates pass rates by 65% vs. fresh golden...](https://enterpriseailabs.io/blog/ai-test-scores-deceive-eval-debt-inflates-pass-rates-by-65-vs-fresh-golden-sets.php)

Canonical: https://enterpriseailabs.io/blog/outbound-ai-voice-benchmarks-why-90-task-rate-and-300ms-latency-arent-comparable.php
Markdown: https://enterpriseailabs.io/blog/outbound-ai-voice-benchmarks-why-90-task-rate-and-300ms-latency-arent-comparable.php/index.md
