Voice AI’s Per-Minute Pricing Leaves Out The Decisions That Shape Cost And Call Quality

With the right setup, a minute of voice AI can cost five or ten cents, though the model chosen for a single step can cost ten or twenty times more than another. Much of that spread comes from decisions contact center teams make themselves, like how a call's work is divided and what the system knows before the caller speaks. Those decisions shape how long a customer waits in silence, making voice AI optimization as much a customer experience discipline as a purchasing one.

Alexander App is Director of Customer Service at wirkaufdeinauto.de and its sister brands across Europe, and has spent around 20 years in customer service, building and leading teams at fast-scaling tech companies across Europe and Asia. Since joining the AUTO1 Group three years ago, he has led the scaling setup of its European customer service operations, including teams that manage appointments for more than 765 branches, and introduced AI agents for inbound calls in 9 European countries supporting 7 languages. That work has given him a practitioner's view of which voice AI design choices drive cost and which shape the caller's experience.

"The trap with voice AI is focusing on the cost per minute rather than the all-in expense," says App. Vendors lead with that figure because it's easy to compare, but it says nothing about the design work that drives both cost and call quality. That work starts the moment a caller stops talking.

Voice AI has less than a second to answer

A three or four-second delay barely registers in chat or email. "On a phone call, that feels like an eternity," App says. "Silence is a huge issue." Every response runs a relay. Speech-to-text converts the caller's words, an LLM checks the knowledge base and other systems, and text-to-speech reads the answer back. "Ideally, all of that happens in less than a second," App says. Responses start to feel unnatural around 800 to 900 milliseconds, and every lookup eats into that budget.

Accuracy pulls the other way, since more advanced models can be more complex than a task needs and keep the caller waiting. When a pause can't be avoided, a short "Let me think about that," or even the sound of typing, tells the customer something is happening.

Specialized agents give faster, more accurate answers

A long master prompt can slow replies, because the model has more to process before responding. App's answer is to split one voice agent into what he calls "a group of smaller, specialized sub AI agents."

On an inbound call, one sub-agent identifies the caller and a second works out why they're calling, a step where App recommends a more advanced model because intent is easy to misread. Specialists such as an appointment agent handle the rest. Because every agent shares the same voice, a caller might pass through ten of them in two minutes without noticing.

The knowledge base gets the same treatment. App divides hundreds of FAQs by topic, with separate agents for branch visit knowledge, appointment sheduler, and specialized agents with knowledge for business customer insead of private. A bot choosing among 500 or 1,000 questions has far more room for error than one working from the 50 that apply. "The context is already established, so you improve response quality," App says.

Routine steps can run on cheaper models

Splitting the call also lets each step run on the cheapest model that can handle it. Booking a slot, for instance, simply follows a rule. "You don't need one large language model for every activity in your voice bot," App says. "Depending on the model, the cost difference can be tenfold or even twentyfold," he adds.

The greeting itself can do work. While the bot welcomes the caller, it can query other systems, assembling context before the customer asks for it so the model doesn't stall later. "Rather than waiting for the customer to ask a question, you provide context in advance," App says.

Language adds another variable for a European operator. Speech models perform well in English and French but can drop sharply in less common languages, where App says teams may need a pricier LLM with extra context to compensate. The cheapest setup in one market can be the wrong one in the next.

The costs that never show up in the per-minute price

Before comparing vendors, App asks leaders to establish what a call costs today, including preparation, wrap-up, hiring and overhead, and what they expect voice AI to deliver, more cost efficiency or additional effectiveness. The vendor's price rarely stops at the per-minute rate.

Many platforms charge for access and visibility into AI calls. In Europe, servers in the US or China raise data residency questions. Quality assurance may be built in or require a second LLM run over calls afterward. Uptime counts too, since 99.9% and 98% look close but separate under an hour of monthly downtime from more than half a day. "If the servers go down, calls go down too," App says.

First deployments often need hands-on support as well. "A company might tell you it charges three, four, or five cents per minute, but then there's another $10,000 for a forward-deployed engineer, $20,000 for the platform, and $30,000 for quality assurance," he says.

Pricing structure decides how much of this a team controls. Outcome-based contracts charge only for calls the AI resolves without escalating to a human agent, which keeps the risk low for teams without much experience. "But it can be expensive, potentially costing $1, $2, or $3 per resolved call," App says. "With per-minute pricing, you might get costs down to $0.05 or $0.10 per minute, giving you more levers to control your expenses."

Small, closely watched pilots keep voice AI improving

App's operating advice starts with scope. Companies often try to do everything at once, sometimes under a senior mandate to use AI everywhere, and stall on legacy systems and organizational silos. "Start with a very clear use case, keep the scope small, and learn quickly," he says. A narrow pilot also lets a team show value to management before anyone asks why the money is being spent.

Learning quickly means watching real conversations closely and adjusting often, treating the system as something to keep coaching. The company went through 27 versions of its inbound voice AI in 12 months, changing answers, knowledge base structure, adding more functionality and model choices along the way.

Some steps shouldn't be left to a probabilistic system at all. Certain processes in insurance, banking, and pharmaceuticals can't tolerate even a 0.5% error rate. "Don't use a probabilistic agentic system for something that needs to happen the same way every time," App says. For those, he favors rule-based workflows and conventional automation.

The newest model is often more than a caller needs

The same restraint applies to model upgrades. App recommends trying each new release in a controlled quality assurance environment before it goes anywhere near production. For voice, he finds that models released six months or a year ago already handle many of the needs callers bring. "You're usually not having a deep research discussion with a PhD," he says. "You're talking to someone who has a question, and a simpler model can often answer it."