Build vs. Buy an AI Voice Agent: The Arithmetic Most Agencies Won't Show You

The short version

At the volume a single small business actually does — call it 500 to 2,000 minutes a month — building an AI voice agent on Vapi or Retell costs less in platform fees than buying a packaged receptionist. Somewhere between $60 and $250 a month against $149 to $449.

That is a real saving and it is almost never the right reason to build, because platform fees are maybe a third of the true cost. Build when the packaged products cannot do something you need. Do not build to save $100 a month.

What the platforms actually charge

As published on each vendor's own pricing page, read in the first week of September 2026.

Note what the middle column of the Retell row tells you: your model choice moves the per-minute cost by 50x. GPT-5 nano at $0.003/min and GPT 5.5 at $0.16/min are the same call, differently reasoned. Most receptionist work — take the details, check a calendar, book a slot — does not need the expensive tier, and the single biggest cost lever in a built agent is refusing to run a frontier model on a scheduling call.

A worked example

An HVAC shop takes 400 calls a month averaging four minutes: 1,600 minutes.

Built on Retell, mid-tier model, standard TTS, US telephony — call it $0.12/min blended. That is $192, plus concurrency. Say $250 all in.

Built on Vapi, bringing your own model key — $0.05/min hosting is $80, plus your own model and TTS spend, plus $10 per concurrent line. Call it $180.

Bought from Rosie at the Growth tier: $299/mo for 2,000 minutes. Bought from Goodcall at Growth: $129/mo, unlimited minutes, capped at 250 unique callers.

So the packaged product with unlimited minutes is cheaper than either build at this volume. The build only wins on platform cost against the minute-metered packaged products, and only by about $50 to $100 a month.

The line items the per-minute number hides

This is where build-vs-buy is actually decided.

Concurrency. Retell charges $8 per concurrency per month past the free 20; Vapi charges $10 per line per month. Fine at normal volume, not fine on the first hot day in July when everyone calls at once. Size for your peak hour, not your average.

Compliance. Vapi lists HIPAA at $2,000/month and zero data retention at $1,000/month as add-ons. Retell puts HIPAA and BAA terms under Enterprise with a custom MSA. If you are a dental practice or a med spa, that single line item ends the build conversation for most practices — a packaged vendor that already carries the BAA is dramatically cheaper. We wrote up what a BAA actually covers and who signs one separately.

The build itself. Prompt design, calendar logic, escalation rules, CRM field mapping, testing against real recordings. Whoever does this — you, a staff member, or an agency — spends real hours on it, and the estimate is always low because the hard part is the twenty-third edge case, not the first.

Month six. This is the cost nobody quotes. Your price book changes. A tech leaves. The calendar integration changes an API. The model provider deprecates a version. Somebody has to notice and fix it. When you buy, that person works for the vendor. When you build, that person is you, or an agency on a retainer that quietly costs more than the packaged product ever did.

Before you sign an agency build, ask one question: who owns the prompt in month six, and what does that cost per month? The answer tells you whether you are buying a system or renting a dependency.

When building is genuinely right

Four cases, and they are all about capability, not price.

• Nothing integrates with your system of record. If your CRM or scheduling system is not on any packaged vendor's list, the vendor's biggest advantage evaporates and you are building the connection either way.
• You are an agency running many clients. Per-minute platform economics across 30 client agents beat 30 packaged subscriptions by a wide margin, and you have the technical capacity by definition. This is the strongest case in the list.
• Your call logic is genuinely unusual. Multi-location routing with capacity rules, or a qualification tree that branches on something packaged products do not model.
• You need the transcripts and the data under your own control for reasons a vendor DPA cannot satisfy.

If none of those is true for you, buy.

When buying is genuinely right

• Single-location business under about 2,000 minutes a month.
• You are in a regulated vertical and need somebody else to carry the compliance obligation.
• Nobody on the team wants to be paged when the phone agent breaks.
• You want to be live this month rather than this quarter.

The agency case, priced out

The one build case that is not close: an agency running voice agents for many clients.

Thirty client agents at 600 minutes a month each is 18,000 minutes. On Retell at a $0.12 blended rate that is roughly $2,160 a month in platform cost for all thirty, plus concurrency. The same thirty clients on Rosie Scale would be $4,470 in subscriptions — and you would be reselling somebody else's product on somebody else's margin.

The bigger point is that you amortize the build once. The prompt architecture, the escalation pattern, the CRM mappings: you solve them on client one and deploy them thirty times. That is the real economics of building, and it is why agencies build and single businesses should not.

The trap in the agency case is support. Thirty clients means thirty people who call you when the phone agent says something odd, and you have quietly taken on a support obligation the packaged vendor used to carry. Price it into the retainer or it will eat the margin you just calculated.

What breaks in month three

The build costs that appear in no quote are all maintenance, and they arrive on roughly the same schedule.

The model changes underneath you. Providers deprecate versions and ship successors that behave differently. An agent tuned against one model does not perform identically on the next, and the change lands whether or not you were ready for it.

The integration drifts. A calendar API changes a field. A CRM adds a required property. Nothing errors loudly — bookings just start failing in a way nobody notices until a customer complains.

The prompt rots. Prices, hours, service area, staff names. Every one of those is a line somebody has to remember to edit.

Nobody is on call. At 2am on a Saturday your agent is answering calls. If it starts failing, who finds out, and how? Packaged vendors have a support rota. Your build has you.

None of this argues against building. It argues for costing it honestly: a built agent needs a named owner and a few hours a month, permanently. At agency scale that is a role. At single-business scale it is the reason the packaged product wins.

The objection you will hear

An agency quoting a custom build will tell you: "You'll own it instead of renting it."

Half true. You own the prompt and the call logic. You do not own the platform, the model, the telephony, or the voices — all rented, from vendors who can change terms and deprecate things. What you actually own is the obligation to keep the whole assembly working.

Ownership is worth paying for when the thing you own is a differentiator. A phone agent that books appointments is table stakes. Own what your competitors cannot buy, and rent the rest.

Where the platforms are heading

The build-side platforms are not fragile bets any more. Vapi reached a reported $500M valuation in May 2026 after Amazon's Ring picked it over roughly 40 competing platforms, per TechCrunch; ElevenLabs raised $500M at a reported $11B valuation in February 2026. Those are reported figures, not audited ones, but they answer the "will this vendor exist in two years" question well enough for most buyers.

That maturity cuts both ways. It means building is safer than it was, and it also means the packaged products sitting on top of these platforms are getting better faster than a one-off custom build will.

The honest summary

Next step

If you are buying, start with the AI voice receptionist workflow — it has the escalation design and the fields to capture. If you are building, the platform pages for Vapi, Retell, Bland and ElevenLabs have the current feature detail, and our Vapi vs Bland vs Retell comparison covers the platform choice itself.