Cost per call: how to model voice AI economics for a 50-doctor practice

Voice AI vendors quote cost per call. Some quote $0.15. Some quote $0.50. Some quote $1.20. The numbers look comparable. They are not. The difference between a $0.20 quote and a $0.80 quote often comes down to what is bundled, what is excluded, and how first-call resolution rate is measured. This article walks through how to actually model voice AI economics for a multispecialty clinic practice or an RCM operation, with the assumptions that matter.

Why cost per call is the wrong starting question

The right starting question is not "what does it cost per call, but "what does it cost per resolved unit of work." A scheduling call resolved by AI saves you the labor of one scheduling call. A scheduling call partially resolved by AI and then handed to staff for cleanup may save you nothing.

Three numbers determine the actual economics:

• Cost per call — the AI agent unit cost
• First-call resolution rate (FCR) — the percentage of calls completed without human intervention
• Cost of human cleanup — the labor cost of finishing the calls AI did not

Effective cost per resolved call is roughly: (AI cost) + ((1 - FCR) × human cleanup cost). At 95 percent FCR, AI is dramatically cheaper. At 70 percent FCR, AI may not save money — and may actually slow operations down by adding a handoff step.

The cost components vendors don't always explain

A voice AI call has multiple cost layers. Some vendors bundle them into one rate. Others charge separately. Knowing what is in the rate is the difference between a clean comparison and a misleading one.

Telephony

The cost of placing or receiving the actual phone call. Typically $0.0085 to $0.014 per inbound minute, slightly higher for outbound. This is a pass-through cost from the carrier and is hard to compress.

Speech-to-text (STT)

Real-time transcription of the conversation. Industry pricing ranges from $0.0043 per minute to $0.025 per minute depending on provider and accuracy tier. STT is often the second-largest cost component.

Reasoning model

The large language model that drives the agent's decision-making. Cost depends on the model and the verbosity of the conversation. For a typical 4-minute prior auth call, reasoning costs land between $0.01 and $0.10 depending on provider and prompt design.

Text-to-speech (TTS)

Synthesizing the agent's spoken responses. Pricing from $0.07 per minute to $0.18 per minute depending on voice quality. Higher-quality voices cost more but improve perceived professionalism, which matters in healthcare.

Platform overhead

Compute, audio routing, observability, audit logging, integration infrastructure. Often the smallest line item but the one that scales with concurrency.

For a 4-minute call at a competent vendor, the all-in component cost typically lands between $0.50 and $1.00. Vendors quoting below $0.30 are usually subsidizing during the pilot or quoting a narrow definition. Vendors quoting above $2.00 are typically charging for managed-services overhead, not the call itself.

Modeling a 50-doctor multispecialty practice

For a 50-provider multispecialty group with conservative assumptions:

• 400 inbound calls per business day, 250 business days per year → 100,000 calls per year
• 70 percent of those calls are routine (scheduling, reminders, refills, intake)
• Average handle time for routine calls: 3 minutes for AI, 5 minutes for staff
• Loaded labor cost for front-desk staff: $25 per hour

At the human-only baseline, those 70,000 routine calls consume roughly 5,800 staff hours per year, or $146,000 in labor cost.

At an AI-handled-routine model with 90 percent FCR:

• 63,000 calls handled by AI at $0.75 per call → $47,250
• 7,000 calls handled by staff at $2.08 per call (5 minutes at $25/hr) → $14,560
• Total: $61,810 — a $84,000 annual saving against the human-only baseline

This calculation does not include downstream gains: reduced no-shows from better reminders, recovered revenue from no-show rebooking, lower staff turnover from removing repetitive call load, and capacity unlocked for staff to focus on higher-value patient interactions. In real deployments those gains often equal or exceed the direct labor savings.

The same model for an RCM operation

RCM unit economics are sharper because the calls are more repetitive and the dollar value of each resolved call is higher. For a mid-size RCM team running prior authorization at scale:

• 20,000 payer calls per month, primarily prior auth and eligibility
• Human cost per resolved call: $7.00 (14 minutes at $30/hr loaded plus overhead)
• AI cost per resolved call: $0.85 at 4.5 minutes AHT
• FCR: 91 percent

Annual labor cost at human-only: $1,680,000. AI-handled with human cleanup on the 9 percent: $228,600. Annual saving: roughly $1.45 million. Turnaround time on prior auth status drops from days to hours, which compresses cash cycle and reduces denials downstream.

The questions to ask before you sign

Six questions to take into any voice AI pricing conversation:

1. What is bundled in your per-call rate, and what is passed through separately? Telephony, STT, reasoning, TTS, observability — all of these are real costs. A vendor that does not break them down is hiding something.
2. What is your first-call resolution rate, measured how, on what call mix? An FCR measured on a curated demo set is meaningless. Insist on FCR measured on a representative sample of your actual calls.
3. What is your handoff cost — does the human start at zero or with full context? The cost of cleanup depends entirely on this answer.
4. How does pricing scale with concurrency? Some platforms charge per concurrent call, which can dominate the unit economics at peak hours.
5. What happens to pricing if our call mix changes? If you move from scheduling to prior auth, does the rate stay flat or shift?
6. Will you commit to an SLA on accuracy and a benchmark on a sample of our calls? Vendors who refuse this are not confident in their numbers.

How we price

At iBridge, we run a 100-call benchmark on a representative sample of your real calls before quoting a unit rate. The benchmark is at no charge, the methodology is transparent, and the result is the floor of what we will commit to in a contract. We do this because the only honest way to quote voice AI for healthcare is on data from your actual operation.

For a clinic considering voice AI at any scale, the cheapest move is to spend 30 minutes modeling the math against your own numbers before you take a single sales call.