Cost per Call: Benchmarking Voice Agents Against Human Agents for Insurance Intake

Insurance carriers evaluating voice agents for intake lines routinely compare per-minute API pricing against human agent hourly rates and stop there. That comparison misses most of the cost. A voice-agent-augmented intake line carries per-call charges for automatic speech recognition (ASR), large language model (LLM) inference, and text-to-speech (TTS); it absorbs failed-call retries; it pays for human handoff on calls the agent cannot complete; and it carries ongoing compliance and integration maintenance overhead. The dominant variable in total cost of ownership (TCO) is the human-handoff rate — the share of calls the agent escalates to a live agent. This piece builds an illustrative TCO model comparing a human-only intake line against a voice-agent-augmented one across three hypothetical call-volume scenarios, with sensitivity to handoff rate as the primary lever. The scenarios, volumes, and rates below are illustrative examples constructed by the author, not figures drawn from industry research or vendor pricing — they are meant to show the shape of the model, not to prescribe specific numbers. ## What a voice agent actually costs per call Every voice agent call chains three inference steps: ASR transcribes the caller's speech, an LLM generates a response, and TTS synthesizes the spoken reply. The i-LAVA architecture — described in a paper titled i-LAVA: Insights on Low Latency Voice-2-Voice Architecture for Agents, published September 25, 2025, and classified under Computer Science > Sound — provides a useful reference for where latency and cost accumulate in such pipelines. The architecture tested in that work uses CSM1b, a model that ingests both audio and text from prior exchanges to generate contextually appropriate speech. Within that pipeline, the TTS component has the highest impact on Real Time Factor (RTF) — the ratio of processing time to audio duration. That matters for cost because RTF drives how much compute each call consumes, and TTS is the segment where compute is hardest to compress without quality loss. The i-LAVA experiments explored reducing Residual Vector Quantization (RVQ) iterations in the TTS decoder and reducing the number of codebooks used in Mimi, finding these to be the most important optimizations for V-2-V implementations based on CSM. But the same work notes that optimizing RVQ iterations comes at the cost of decreased voice quality. The Real World Voice EQ Bench (RW-Voice-EQ Bench) — a multidimensional benchmark for evaluating voice AI across TTS, speech-to-speech (STS), speech understanding (SU), and ASR — underscores why that quality trade-off is not trivial. For TTS, the benchmark treats naturalness, expressiveness, identity stability, and reliability as largely independent evaluation dimensions. A degradation in one does not necessarily mean the others hold. For STS, the benchmark notes that access to audio does not guarantee use of vocal affect — some agents remain largely transcript-driven even when the input signal carries emotional information. The implication for insurance intake is that a cheaper TTS component may produce speech that is intelligible but flat or inconsistent, which could affect caller experience and potentially increase the likelihood that a caller asks for a human. This causal chain — TTS quality degradation leading to higher handoff rates — is the author's hypothesis, not a finding established by the cited research. It is a reasonable extrapolation from the quality dimensions the benchmark identifies, but carriers should validate it against their own call data before relying on it for procurement decisions. ## The handoff rate problem The handoff rate is the share of calls the voice agent cannot complete autonomously and must escalate to a human agent. In a human-only intake line, every call is handled by a human at fully loaded human cost. In a voice-agent-augmented line, the agent handles the calls it can at per-call API cost, and a human handles the remainder at human cost. The economics turn on where that split falls. At a low handoff rate, the agent absorbs most calls cheaply and the human line is a thin overflow resource. At a high handoff rate, the human line is nearly as large as it would be without the agent, and the carrier pays both the agent's per-call cost and the human cost on top. The sensitivity is non-linear because the human line must be staffed to peak capacity, not average — a 30% handoff rate on 100,000 calls means 30,000 human-handled calls, but the staffing must cover the distribution of those calls across hours, not just the total. The voice agent does not eliminate the human line; it changes its shape. This is the core TCO insight: the agent's value is not the cost per call it handles, but the cost per call it deflects from the human line. ## An illustrative TCO model The model below compares two configurations across three hypothetical call-volume scenarios. All figures are illustrative examples constructed by the author to demonstrate the model's structure — they are not derived from vendor pricing, industry benchmarks, or the cited research. Carriers should replace them with their own volume data, negotiated API rates, and loaded human costs. ### Assumptions (illustrative) - Per-call voice agent cost (ASR + LLM + TTS): Not specified — carriers should input their negotiated rate. - Human agent fully loaded cost per call: Not specified — carriers should input their own figure. - Failed-call retry rate: A percentage of agent-handled calls that fail (due to ASR errors, timeouts, or caller abandonment) and must be retried or handed off. Not specified — carriers should measure their own. - Compliance overhead: Fixed periodic cost for call recording, disclosure scripting, audit logging, and regulatory review. Not specified. - Integration maintenance: Fixed periodic cost for connector upkeep, prompt tuning, model updates, and monitoring. Not specified. ### Scenario matrix (illustrative) | Scenario | Calls/month | Handoff 10% | Handoff 30% | Handoff 50% | |---|---|---|---|---| | A | 5,000 | Agent handles 4,500; human handles 500 | Agent handles 3,500; human handles 1,500 | Agent handles 2,500; human handles 2,500 | | B | 25,000 | Agent handles 22,500; human handles 2,500 | Agent handles 17,500; human handles 7,500 | Agent handles 12,500; human handles 12,500 | | C | 100,000 | Agent handles 90,000; human handles 10,000 | Agent handles 70,000; human handles 30,000 | Agent handles 50,000; human handles 50,000 | The call volumes (5,000, 25,000, 100,000) and handoff thresholds (10%, 30%, 50%) are hypothetical examples chosen to span a range a mid-size carrier might see — they are not drawn from industry data. At 10% handoff, the human line is a thin overflow tier and the agent carries most of the volume. At 50% handoff, the human line is half the total volume, and the carrier is paying for both the agent infrastructure and a human line nearly as large as the baseline. The crossover point — where the augmented line costs more than the human-only line — depends on the specific per-call rates, but the model shows it exists and that handoff rate is the variable that moves it most. ## What the research tells us about the components The cited research does not provide insurance-specific cost data. It does provide technical findings that bear on the cost structure: ASR accuracy. At NADI 2025, a submission using a fine-tuned Whisper-large-v3 encoder with data augmentation ranked first for the Arabic Dialect Identification (ADI) subtask, achieving 79.83% accuracy on the official test set. The same group ranked second for multi-dialectal Arabic ASR, using SeamlessM4T-v2 Large (Egyptian variant) fine-tuned separately for each of eight dialects, obtaining an average word error rate (WER) of 38.54% and character error rate (CER) of 14.53% on the test set. These are top-ranking results in a shared-task setting, but the WER of 38.54% also illustrates that ASR in challenging conditions — here, multi-dialectal Arabic — can carry substantial error rates. For insurance intake, ASR errors on policy numbers, names, and addresses are not just quality issues; they are retry and handoff triggers. The NADI 2025 results are specific to Arabic dialects and should not be generalized to English-language intake lines without separate validation. TTS latency and quality trade-off. The i-LAVA findings identify TTS as the highest-impact component on RTF and show that the main optimization levers — reducing RVQ iterations and Mimi codebooks — trade quality for speed. For a carrier, this means the TTS component is where cost reduction is most available but also where quality reduction is most likely to affect the caller experience. Whether that quality reduction translates to higher handoff rates is, as noted above, an untested hypothesis. Evaluation dimensions. The RW-Voice-EQ Bench's separation of TTS into naturalness, expressiveness, identity stability, and reliability means a carrier cannot rely on a single