This post is for operations leaders and CTOs deciding whether a voice agent will pay for itself. It walks through the per-call cost stack behind a production deployment handling 4,000+ calls a day, the latency budget that sets the cost ceiling, and the three cases where voice agent ROI does not work.
Voice agents are the AI category where per-call unit economics either work cleanly or not at all. A correctly architected voice agent runs at 10-30 cents per call on telephony, model, and infrastructure combined. A mis-architected one runs at $1.20 a call before you count the engineering time spent on tail-latency incidents. The difference is not the LLM. It is the latency budget, the model choices at each pipeline stage, and the telephony provider.
The numbers come from the cost model behind a production voice agent we ship for a B2C services brand at 4,000+ calls a day, and the back-office voice tooling that has released 1,400 hours a month for an EU insurance customer. Both engagements are profitable at the per-call rates below. Neither would be if the pipeline were architected the way most voice-agent demos are.
The five components in the per-call cost stack
A production voice agent has five cost components, each billed differently and each with different optimization levers. Most vendors quote the LLM piece and ignore the rest, which is why the projected ROI in the pitch deck rarely matches the bill at month three.
- Telephony (inbound or outbound carrier minutes, DID/SIP termination): $0.008-0.030 per minute depending on geography and provider.
- Speech-to-text (streaming STT, audio in): $0.004-0.012 per minute depending on provider and accuracy tier.
- LLM inference (input and output tokens, plus prompt-cache hits): $0.001-0.020 per turn depending on model and prompt size.
- Text-to-speech (premium voice, audio out): $0.012-0.030 per 1,000 characters of output speech.
- Infrastructure (orchestration, observability, eval platform, session state): $0.005-0.015 per call amortized.
On a typical 90-second consumer support call: 1.5 minutes of telephony at $0.015 (about $0.023), 1.5 minutes of STT at $0.008 (about $0.012), four LLM turns at $0.004 each (about $0.016), 600 characters of TTS at $0.018 per 1,000 (about $0.011), and $0.008 of infrastructure overhead. Total: roughly $0.07 per call for a tight architecture. That is the floor for English consumer voice in 2026.
Where the cost blows out
The vendors quoting $1.20 a call are not lying. They are running an architecture that is two to three years out of date for production voice. The failure modes are consistent:
- Sending the full conversation history to the LLM on every turn instead of using prompt caching: 3-5x the input-token bill.
- Using a flagship model for routing decisions a small classifier could make in 50ms: 10-20x the LLM bill on the turns that route.
- Using non-streaming STT that waits for the user to finish speaking before transcribing: adds 800ms-1.5s of latency and forces users to repeat themselves when they get impatient.
- Premium-voice TTS on every turn, including yes/no acknowledgements: 3-4x the TTS bill for no perceived quality gain.
- No silence detection or VAD tuning, so the agent waits 2-3 seconds at every utterance boundary: calls drag on and telephony minutes accumulate.
On our 4,000-calls-a-day deployment, prompt caching reduced the per-turn LLM cost by 64% on a single architecture change. A 7B router model in front of the flagship LLM handles 80% of turns at 12% of the flagship's cost. The ROI works because the architecture works, not because the carrier rates are special.
The latency budget sets the cost ceiling
Conversation feels natural under about 2 seconds of p95 turn latency. Above 3 seconds users start interrupting and the call structure breaks down. Above 5 seconds abandonment climbs into double digits. The latency budget determines the cost ceiling because every model choice that reduces latency also reduces cost. The streaming and speculative architecture behind this is the core of our voice agent development work.
The canonical sub-2-second budget for a customer-service voice agent:
- STT first partial: 150-300ms (streaming STT with VAD tuning)
- Routing decision: 50-150ms (small classifier model, prompt-cached)
- LLM first token: 300-600ms (mid-size model, aggressive prompt caching, streaming response)
- TTS first audio: 150-400ms (streaming TTS, low-latency voice tier)
- Network and orchestration overhead: 100-200ms
Total p95: 750-1,650ms. The teams that hit sub-2-second consistently instrument each component with p50/p95/p99 and shave 50ms here and 100ms there. The teams running at 4-5 seconds typically have one or two components taking 1.5 seconds or more on their own, and no observability to find which.
The 4,000-calls-a-day reference deployment
The reference engagement: a multi-location consumer services brand handling outbound appointment reminders, inbound rescheduling, and post-service follow-up. Before deployment, the call volume was consuming roughly 28 hours per location per week of front-desk time. After deployment, the voice agent handles 4,000+ calls a day across the locations, with human handoff on roughly 9% of calls (complex rescheduling, billing escalations, complaints).
- Per-call cost: $0.09 fully loaded (telephony, STT, LLM, TTS, infrastructure).
- Equivalent human-handled cost per call: roughly $4.50 (loaded staff cost plus retry overhead).
- Net saving per call: $4.41, applied across the 91% of calls that complete without human handoff.
- Monthly saving at 4,000 calls a day: roughly $480,000 against the human-only baseline.
- Build cost recovered: roughly 6 weeks after launch.
The build was a 12-week engagement with eval-first delivery; the eval set was 280 representative call scripts covering happy path, edge cases, refusal-required, and escalation categories. The architecture and the metrics that gated the rollout are in our case studies.
The 1,400-hours-a-month EU insurance build
The second reference deployment is an EU insurance customer whose back-office team was spending roughly 1,400 hours a month on inbound policy-question calls, many of them repetitive: coverage lookup, claim status, document requests. The voice agent fronts the call, handles 73% of inquiries without escalation, and routes the remaining 27% to a human with full call context pre-populated.
The cost math is different here. The customer is not measuring per-call cost reduction; they are measuring back-office hours released for high-judgment work. 1,400 hours a month at the loaded cost per back-office FTE is roughly €52,000 a month of reclaimed capacity, against a per-call cost of €0.14 (premium voice tier, EU-resident infrastructure, GDPR-compliant audit trail). The ROI is real, but it works because the alternative cost, loaded back-office labor, is high. The same architecture in a low-wage market might not.
When voice agent ROI does not work
Voice agent ROI fails in three recurring cases, and an honest vendor will say so on the first call:
- Low call volume (under about 200 calls a day). The fixed infrastructure cost (orchestration, observability, eval platform) does not amortize. A well-built chatbot or async messaging is usually the better channel.
- Highly variable conversation structure with no repeatable patterns. If 60% or more of calls need human judgment, the voice agent is a routing layer with extra cost; a smarter IVR is cheaper.
- Very low labor cost. If the per-call alternative is $0.40 of human time, saving $0.09 a call does not justify the build cost.
What buyers should ask in the procurement call
Six questions surface architecture quality fast:
- Show me your per-call cost breakdown on a recent deployment, by component (telephony, STT, LLM, TTS, infrastructure).
- What is your p95 turn latency on production traffic? Show me the latency dashboard.
- Which model handles routing versus the flagship turns, and why?
- How is prompt caching configured, and what is the cache hit rate in production?
- What does the eval set look like for a voice deployment: happy path, edge cases, refusal categories, escalation triggers?
- Show me the audit log for a single call: the full trace from STT input to TTS output with timing and cost per step.
A vendor running a real production voice stack can answer all six in narrative on the call. A vendor reading from a deck will hedge on at least three. The hedge points are the architectural decisions that determine whether the per-call cost lands at $0.09 or $1.20.
What the regional buyer should weigh
Voice agent economics vary by region, mostly because labor cost (the alternative) and telephony cost (the bill) vary by region.
- United States: ROI works above about 300 calls a day in most service industries; loaded labor cost is high enough that per-call savings compound fast.
- United Kingdom: similar economics to the US; UK GDPR audit-trail requirements add infrastructure cost, but the math still works above about 400 calls a day.
- Canada: PIPEDA and Quebec's Law 25 add a small infrastructure layer; Quebec deployments may need a French-language voice tier with separate STT and TTS providers.
- Australia: Privacy Act and APP-compliant data residency add infrastructure cost, and Australian-resident telephony is more expensive, but the math works above about 500 calls a day in financial services and healthcare.
Wrapping up: the per-call math is the engagement
Voice agents are not a category where the ROI conversation can wait for a phase-2 measurement. The unit economics either work in week one or they do not work at all. The decisions that set the per-call cost (routing model, STT provider, prompt-caching strategy, TTS tier per turn) all get made in the first 14 days of the engagement. A vendor without an opinion on each of them will deliver a working demo and a $0.90-per-call production bill.
If you are scoping a voice agent build and want a written cost model (per-call breakdown by component, latency budget per stage, and a projected monthly run rate against your call volume), book a discovery call. One conversation, a fixed-price scope inside 72 hours, and an honest read on whether the per-call math works for your traffic.
Frequently asked questions
How much does a voice agent cost per call?
A tight architecture runs at roughly $0.07-0.30 per call across telephony, STT, LLM, TTS, and infrastructure. Architectures without prompt caching, streaming STT, or a router model run at $0.90-1.20 a call.
What call volume does a voice agent need to pay off?
In most high-wage markets the math works above about 300-500 calls a day with a repeatable conversation structure. Under about 200 calls a day the fixed infrastructure cost does not amortize.
Why does latency affect the cost of a voice agent?
The model choices that cut latency (streaming STT, a small router model, prompt caching, a low-latency TTS tier) are the same choices that cut cost, so the latency budget effectively sets the cost ceiling.