What your finance team sees versus what is actually happening
Your Twilio invoice shows a total. Your ElevenLabs invoice shows a total. The sum of these figures is what ends up in your P&L as voice AI costs. What it does not show: that 23% of your spend this month went on calls that terminated in the first 30 seconds before any value was exchanged. That one agent configuration costs €0.85 per conversation while all others average €0.18. That your highest-volume calling window — Tuesday and Wednesday afternoons — has a 12% failure rate you are paying full price for.
The visible cost and the real cost are rarely the same number. Most voice AI teams discover this the hard way: a monthly spend that looked reasonable in aggregate turns out, on per-call inspection, to be driven almost entirely by three pathological patterns that could each be fixed in an afternoon. The patterns were invisible in the aggregate. They are obvious at the call level.
This is not a billing problem. It is a visibility problem. Twilio and ElevenLabs are charging you correctly for what they delivered. The question is whether what they delivered was worth anything — and that question can only be answered at the call level, not at the invoice level.
The hidden multiplier: operational overhead and churn
Every failed call that generates a human support interaction costs you in four additional ways: the inbound support call, the support agent time (typically 8–12 minutes at €0.30–€0.50/minute fully loaded), the goodwill erosion that increases churn probability, and the opportunity cost of the support agent not handling something else.
Contact centre operators who have traced failed-call incidents end to end — from the dropped AI call through the support interaction through the eventual churn event — arrive at true cost-per-failed-call figures between 8x and 15x the direct provider cost. For an enterprise SaaS context where the monthly subscription value is €500+, a single call failure that triggers churn represents a lifetime value loss of €3,000–€6,000. The €0.15 Twilio + ElevenLabs cost for that call is noise.
The multiplier is not a theoretical model. It is what happens when someone finally builds the spreadsheet that connects the Twilio call log to the support ticket to the cancellation event. Most teams never build that spreadsheet. The cost remains invisible — not because it is not real, but because no one put the three datasets next to each other.
Cost per call versus cost per converted call
Cost per call is the metric that feels like cost control but is not. If you reduce your average cost per call from €0.40 to €0.25 by switching to a faster, cheaper TTS model, congratulations — you have cut spend by 37.5%. If the cheaper model also reduces your conversion rate from 18% to 11%, your cost per converted call went from €2.22 to €2.27. You spent more to get the same result, and the cost-per-call metric told you the opposite story.
Cost per converted call forces you to hold cost and outcome in the same view simultaneously. It makes invisible trade-offs visible: the agent configuration that is cheaper per call but converts at half the rate, the geographic routing choice that saves €0.03 per call but adds 140ms of latency that kills conversions in one region, the verbose system prompt that generates €0.60 worth of TTS per call on questions that could be answered in 40 characters.
Teams that switch from cost-per-call to cost-per-converted-call as their primary metric consistently find, within the first 30 days, configurations that are simultaneously expensive and low-converting — the worst possible combination. These configurations are often months old and were never revisited after the initial deployment.
Fix the measurement before fixing the spend
The path to voice AI cost optimisation starts with a question that sounds almost too simple: what did each individual call cost, and did it convert? Not the monthly aggregate — the individual call, with the full provider breakdown and the outcome attached.
When you can answer that for every call across every provider, two patterns become immediately visible without any further analysis. First: the configurations that are costing far more than the baseline — usually because of unconstrained response length generating excessive TTS characters, suboptimal model selection (eleven_multilingual_v2 on calls where eleven_turbo_v2_5 would suffice), or geographic routing mismatches adding unnecessary latency. Second: the providers where failure rates are silently compounding costs — calls being billed as completed and consuming full character budgets despite delivering zero value to the caller.
The fixes are almost always configuration changes that take under an hour to deploy. Setting a maximum response token count in the agent system prompt. Switching model tier for a specific call type. Updating a geographic routing rule. The hard part is never the fix. The hard part is knowing which configuration is the problem — and that requires per-call visibility that most teams do not yet have. Fix the measurement first. The optimisation decisions that follow are obvious once the data is legible.
Frequently asked questions
What is the true cost of a failed AI voice call?
The direct cost — telephony minutes, TTS characters consumed, compute time — is typically €0.05–€0.20 per call. The true cost, once you add the support interaction it triggers, the churn acceleration it causes, and the brand erosion from a caller who tells others about a bad experience, ranges from €0.50 to €3.00 per failed call. For high-value B2B contexts, it can exceed €500 once the full attribution chain is calculated.
How do you calculate cost per converted call in a voice AI deployment?
Cost per converted call = (total provider spend for the period) ÷ (number of calls that resulted in the desired outcome — booking, purchase, qualification). Most teams track cost per call (total calls) rather than cost per converted call, which disguises high-cost configurations that are also low-converting. The per-converted-call metric is the one that correlates with revenue.
Why do voice AI costs spiral faster than expected?
Voice AI costs are metered at multiple layers simultaneously. Twilio charges per minute of call time ($0.0085/min inbound, $0.014/min outbound at list price). ElevenLabs charges per 1,000 characters of text-to-speech (~$0.30/1K chars on the Turbo tier). If your AI agent is verbose — or your LLM is hallucinating long responses — a 2-minute call that costs $0.028 in telephony alone can cost $0.60–$1.20 once TTS is added. An agent running at 3x expected verbosity for a month can produce a bill 4–5x the budget without any single transaction appearing unusual.
Ready to investigate your own calls?
Connect Sherlock to your voice providers in under 2 minutes. Free to start — 100 credits, no credit card.