Voice AI for a business phone line: build vs buy in 2026
Every few weeks someone asks me some variation of: "Vapi is $0.05 a minute, why would anyone pay hundreds a month for a managed voice agent?" It is a fair question with an unfair framing, because the two numbers are not prices for the same thing. Here is how I would walk a technical founder through the decision.
What the platform rate actually buys
Vapi, Retell, and Bland sell excellent infrastructure. But the advertised per-minute rate is mostly the orchestration layer. On Vapi ($0.05/min platform fee) the language model, speech-to-text, voice synthesis, and telephony are all billed on top, typically through your own provider keys. Computed at published provider rates that stack lands around $0.11 to $0.20 per minute. Retell's own component prices compute to about $0.13/min with a mid-tier LLM, more with premium voices. Bland's Start tier works out near $0.155/min with telephony. None of these numbers are scandalous; they are just not the headline.
The line item nobody prints: you
The bigger omission is the agent itself. On a DIY platform, the prompts, the tool integrations, the booking-system connection, the fallback behavior, the monitoring, and every production edge case (echo, barge-in, answering machines, callers switching languages mid-sentence) are your engineering work. A realistic initial build for a production-grade agent is measured in weeks, not days; the worked comparison published by Canarics puts it at roughly €8,000 of engineering (about 16 days at €500/day) plus ongoing tuning, which is consistent with what I would quote for that scope. At SME call volumes the build cost dominates the per-minute rate for the first year or two.
And the maintenance is not hypothetical. Your client changes their opening hours, their price list, their booking system. The model provider deprecates the model you tuned against. The carrier changes media behavior. Someone has to own that.
When building is right
- You have in-house engineers who will own the agent as a product, not a side quest.
- Your requirements are genuinely unusual: custom telephony topology, on-premise constraints, deep proprietary integrations.
- Call volume is large enough that per-minute economics beat any subscription, and you can amortize the build.
When buying is right
For a company that wants a working phone agent without an engineering project, the calculation collapses to subscription models. There the market splits again: metered European vendors (a base fee plus €0.30 to €0.40 per minute, where the meter is the real price) versus flat plans with minutes included. As a reference point, Canarics sells €249/mo inbound or €349/mo outbound with 1,000 minutes included and no setup fee; against a typical metered offer the flat plan wins above roughly 250 minutes a month. Synthflow, at the other end, moved to enterprise contracts from $30,000 a year, which answers a different question entirely.
The evaluation shortcut I give non-technical buyers: ask every vendor what happens on a mishear, who is allowed to end a call, and where the data lives. Vendors who publish production numbers with sample sizes are taking the problem seriously; this engineering writeup is a good example of what that looks like, and its numbers double as an acceptance checklist for anyone's system, built or bought.
The honest summary
Building on Vapi or Retell is buying a very good engine and committing to build the car. Buying a managed agent is buying the car. Both are rational; pretending the engine price is the car price is not.