Building a custom AI voice agent for a single, well-defined job (answering calls, booking appointments, qualifying leads) typically starts at $8,000 for a fixed-scope build delivered in about 4 weeks, as an AI Assistant engagement. Running it afterward is a separate bill: usage-based platforms charge roughly $0.05 to $0.14 per minute of call time, before you add telephony, LLM tokens, and any human handoff. Most teams budget for the build and forget the second number, then get surprised three months in.
This is the question that actually matters when you're scoping a voice agent project: not "what's the sticker price" but "what does this cost to build once, and what does it cost every month after that, forever." Those are two different budgets with two different owners, and vendors selling you the platform have an incentive to make the second one look small.
Build cost and running cost are not the same line item
The build cost is what you pay once to get a working voice agent: call flow design, integration with your phone system and CRM, prompt and knowledge base setup, testing against real call scenarios, and handoff. A fixed-scope build for a receptionist or appointment-booking agent runs from $8,000 as an AI Assistant engagement, delivered in about four weeks, with the client owning 100% of the code. The running cost is what you pay every month after launch, for as long as the agent takes calls: per-minute usage fees to whichever platform handles speech-to-text, the language model, and text-to-speech, plus phone number and carrier charges, plus the cost of a human stepping in when the agent hits something it can't handle. This cost scales with call volume, so it's the number that determines whether the project pays for itself at scale, not the build invoice.
What voice agent platforms actually charge per minute
Published rates vary by provider and by what's bundled into the quoted number. As of 2026, CloudTalk's pricing breakdown puts Vapi's platform fee starting around $0.05 per minute and Bland AI around $0.09 to $0.14 per minute depending on plan, with both noting that third-party voice, model, and telephony charges get added on top of the base rate. Retell AI's own comparison lists its pay-as-you-go rate at $0.07 per minute and estimates a typical full deployment, once telephony and model costs are layered in, lands somewhere between $400 and $1,200 a month for a moderate-volume line. None of these numbers include what happens when a call needs a transcript saved, a ticket filed, or an appointment written into your calendar system. That's integration work, and it's either built into the per-minute platform fee (rare) or billed separately (common).
What's actually included in that per-minute price?
Ask any vendor quoting a per-minute rate to break it into its parts: speech-to-text, the language model call, text-to-speech, and telephony. These are usually metered separately behind the scenes even when you see one blended number. A quote that doesn't separate them is a quote you can't audit later when your bill doubles. Also ask what happens on a failed or dropped call, since several platforms bill a minimum charge even when the call doesn't complete.
Off-the-shelf platform or custom build — which do you actually need?
If your use case is a single script with a handful of branches (confirm an appointment, take a message, route to a department), an off-the-shelf platform with a visual builder can get you live in days, and the per-minute rate above is close to your full cost. If the agent needs to pull live data from your systems, follow business logic that isn't a simple decision tree, or hand off cleanly to a human with full context, a custom build, starting at $8,000, earns its price by avoiding the workarounds and glue code that off-the-shelf tools force on anything non-trivial. We've shipped 20-plus production systems over five-plus years of doing this kind of integration work, and the pattern holds: the simpler the call flow, the less a custom build buys you.
What makes the running cost spike without warning
Three things reliably blow up a voice agent's monthly bill: call volume growing faster than expected, calls running long because the agent loops or the caller rambles, and silent retries where the model calls itself multiple times per turn to get a usable response. None of these show up in a demo. All of them show up in a bill after 10,000 real calls. Before you sign a contract with any vendor, ask for the cost at your current volume and at 5x that volume, in writing.
When not to build a custom voice agent
If you're testing whether voice AI even fits your business, don't start with an $8,000 build. Use a platform's free or low-volume tier for a month, see where it breaks, and let that tell you whether the fixes are configuration (stay on the platform) or architecture (move to a custom build). A lot of "AI receptionist" needs are genuinely solved by configuring an existing tool well, and no engineer should talk you out of the cheaper option just because it's less interesting to build.
Where a custom build pays for itself is volume and integration depth: once you're past a few hundred calls a week, or the agent needs to actually change state in your systems (book the slot, update the record, escalate with the right context), the per-minute savings and the reduction in handoff failures tend to cover the build cost within a few months.
What a realistic budget looks like
| Item | Typical range |
|---|---|
| Fixed-scope custom build | From $8,000, ~4 weeks |
| Platform per-minute fee (usage-based) | $0.05–$0.14/min |
| Full deployment incl. telephony, model, handoff | Roughly $400–$1,200/month at moderate volume |
These numbers are sourced from current vendor pricing pages linked above, not quotes we're making up to sound precise. Your actual number depends on call volume, call length, and how much of the conversation needs a model call versus a scripted response. If you want to plug in your own volume and see where you'd land, our cost calculator does the arithmetic for you.
If you want a fixed price for your specific call volume and use case rather than a per-minute estimate that could go either way, book a call and we'll give you a quote within 24 hours, scoped to what the agent actually needs to do.