Should you build or buy an AI voice agent?
Buy a finished agent when your calls look like other businesses' calls: answering, booking, messages, transfers and reminders. Build, on a developer platform or a framework, when the conversation itself is your product, when you need logic no product offers, or when volume is high enough that the per-minute saving pays for engineers. On list prices read November 12, 2026, platform fees for 1,500 minutes a month run from about $120 to $210 on each platform's cheapest plan and a finished plan is $299, so the monthly fee rarely decides it. What decides it is who builds and maintains the agent. Botcadence's AI voice agent is the buy option; this guide is honest about when it is the wrong one.
Key takeaways
- There are three routes, not two: a framework you host, a developer platform, and a finished product.
- At 1,500 minutes a month the platform bills are close together. Engineering time is the large number.
- Build when the conversation is your product, your logic is unusual, or your volume is very high.
- Buy when your calls are common ones and nobody on the team wants to own a real-time system.
- Whichever you choose, keep your phone number with your own carrier so you can move.
The three routes
| Framework you host | Developer platform | Finished product |
|---|
| Examples | LiveKit Agents, Pipecat | Retell AI, Vapi, Bland AI | Botcadence, Rosie, PolyAI's self-serve plans |
| You write | The agent, in code | Prompts, flows and tool definitions | Your business information and rules |
| You run | Servers, scaling, monitoring | Nothing | Nothing |
| Control | Complete | High | Within the product's settings |
| Time to a production agent | Months | Weeks to months | Days |
| Who fixes it when a model changes | You | You, with the platform's help | The vendor |
What a month costs at 1,500 minutes
List prices, read on each vendor's page on November 12, 2026.
| Route | Product | Monthly fee at 1,500 minutes | Not included |
|---|
| Developer platform | Retell AI | $165 at its calculator's default of $0.11 a minute | Building and maintaining the agent |
| Developer platform | Vapi | $75 hosting plus model, speech and carrier at cost; its own estimate is $82 to $129 per 1,000 minutes | Building and maintaining the agent |
| Developer platform | Bland AI | $210 on Start; $479 on Build | Building and maintaining the agent |
| Platform with minutes | ElevenLabs Agents | $120 on the Pro tier with extra minutes, plus model and carrier | Building the agent |
| Finished product | PolyAI | $299 per location | Your setup time |
| Finished product | Rosie | $299 for 2,000 minutes | Your setup time |
| Finished product | Botcadence | $299 | Your carrier |
The spread across every row is a few hundred dollars a month. That is small next to the cost of an engineer's time, in either direction.
The cost that decides it
Building is not a one-off. A production voice agent needs someone to:
- design the conversation and write the prompts;
- connect and maintain the calendar, CRM and other tools;
- tune turn-taking, noise handling and voices on real calls;
- handle provider outages, slow responses and failed tool calls;
- re-test after every model, prompt or provider change;
- keep up with consent, recording and sector rules;
- read calls and fix what they show.
We won't put a dollar figure on that, because it depends on your team. Estimate it honestly: the hours a week someone will spend for the first three months, and the hours a week after that, at what that person costs you. If the answer is "a developer's spare time", the agent will degrade.
A finished product moves most of that work to the vendor. You still spend time on your script and on testing, and you accept the product's limits.
Signs you should build
- The conversation is your product. You sell an AI phone experience to your own customers and need to own how it behaves.
- Your logic is unusual. Multi-step workflows across your own systems, with rules no receptionist product has.
- Volume is high. At tens of thousands of minutes a month, a few cents a minute pays a salary.
- You have real-time engineers who want to own it, not a developer borrowed from another project.
- You need control of data flow: your own cloud, your own model contracts, specific regions.
Signs you should buy
- Your calls are common ones: answer, answer questions, book, take a message, transfer, remind.
- Nobody will own it. No one on the team has time to tune and maintain a voice agent.
- You need it this month.
- Predictable cost matters more than the lowest cost per minute.
- You want the vendor to carry model and provider changes.
A middle path
Many teams buy first and build later. A finished agent on the main line proves what callers ask and how many calls you get. If the limits start to hurt, you know exactly what to build, and you have the transcripts to design it from. The reverse order, building first and discovering after three months that a product did the same thing, is the expensive one.
Questions to settle before you commit
- What must the agent do that a finished product cannot?
- Who will own it, and how many hours a week do they have?
- What is our volume now, and in a year?
- Where does the phone number live, and can we move it?
- What happens to callers when the system is down?
- Which rules apply to us: outbound consent, recording, HIPAA?
- How will we test changes before callers find the problems?
Limits of the buy option, stated plainly
A finished product cannot do everything a build can. On Botcadence in particular: there is no public API beyond starting calls, messages and automations; a new account answers one call at a time until the limit is raised; the agent does not read keypad presses; opening hours are handled by forwarding or by instructions, not a schedule; and it does not offer a BAA. It does connect to custom tools through HTTP actions, MCP servers, Zapier and webhooks, which covers more than most teams expect. If your requirements sit outside those limits, a developer platform is the better answer.
For the build side, see how to build an AI voice agent. For costs, see AI voice agent cost per minute, and for testing either route, how to test an AI voice agent. The AI receptionist for small business page shows what a finished agent does.