Custom AI voice agent development, built on Vapi.
Vapi gives you the voice stack. Somebody still has to decide what the agent says, who it wakes up at 2am, what happens when a caller goes off script, and where the booking actually lands. That part is the build.
What building on Vapi actually involves.
Vapi is infrastructure, not a product you buy off a shelf. It supplies the parts that are genuinely hard to assemble yourself, and none of the parts that make the agent yours.
The hard parts Vapi handles: turning speech into text fast enough that the pause doesn’t feel wrong, handing that text to a language model, turning the answer back into a voice, and holding a real phone line open the entire time. Vapi publishes an average latency of under 500ms for that loop, and lists over 750,000 developers on the platform (vapi.ai, checked 29 July 2026).
What Vapi does not supply is a single one of the decisions. It does not know your hours, your service area, what you charge, which calls count as emergencies, who gets woken up for those, or how much a caller has to tell you before the agent should stop asking questions and just book them. Those decisions are the build, and they are most of the work.
Assistant or workflow: pick wrong and you lose weeks
Vapi splits the job into two shapes. An assistant is one agent with a prompt and a set of tools, holding a single continuous conversation and deciding for itself what to do next. A workflow is a visual flow chart: conversation nodes that pull specific values out of what the caller said, edges with conditions on them, transfer nodes, end-call nodes, and global nodes that can fire from anywhere in the call. Vapi’s own documentation draws the line at structure, and it’s the right line. Open-ended questions want an assistant. A five-question intake that has to end with a booked slot, a correct address, and a callback number wants a workflow.
A quick demo is almost always an assistant, because an assistant is fast to stand up and sounds impressive for ninety seconds. Once real callers hit it, the awkward paths have to be pinned down: either with a workflow, or with an assistant whose edge cases get handled one at a time. The gap between the demo version and the version that survives a Monday is where the build time goes.
Illustrative · The seat that answers the phone, and the hours nobody is in it
Three things get built. Everything else is detail.
The conversation
- The opening line, which decides whether the caller stays on
- Voice selection, tested out loud against your actual customers, not picked off a dropdown
- Your intake questions, in the order a human would ask them
- Variable extraction so the name, address, and callback number come out as clean fields, not a transcript to read later
- What the agent says when it doesn’t know, which matters more than what it says when it does
The handoff
- Transfer rules: which callers reach a person, and which get handled end to end
- Blind transfer where speed wins, warm transfer with a spoken summary where context wins
- After-hours behavior that is different from daytime behavior, because it should be
- Escalation paths for the caller who asks for a human three words in
- What happens when nobody on your side picks up
The connections
- Vapi tools that call out to a webhook mid-conversation when the agent needs to do something real
- n8n workflows on the other end of that webhook, talking to your calendar, CRM, and inbox
- Live availability lookups so the agent books a slot that actually exists
- Post-call routing: transcript, summary, and lead record land where your team already works
- Confirmation text or email to the caller before they’ve put the phone down
What Vapi itself charges, before anyone builds anything.
Vapi’s headline number is the hosting fee. It is not the whole bill. Speech to text, the language model, the voice, and the phone line are billed separately, which is why a “$0.05 per minute” quote and a real monthly invoice rarely resemble each other. Here is the published structure so you can budget against it honestly. Every line below is billed to you directly, by Vapi or by the provider named in the last column. Acuity does not resell any of it, does not mark it up, and does not add a charge of its own on top.
| Cost component | Published rate | Who bills you |
|---|---|---|
| Vapi platform | ||
| Call hosting | $0.05 per minute | Vapi |
| Chat and SMS messages | $0.005 per message | Vapi |
| Concurrent call lines | 10 included, then $10 per line per month | Vapi |
| Passed through at provider cost | ||
| Speech to text | At cost. $0 if you bring your own API key | Not Vapi. Your transcription provider |
| Language model | At cost. $0 if you bring your own API key | Not Vapi. Your model provider |
| Text to speech | At cost. $0 if you bring your own API key | Not Vapi. Your voice provider |
| Telephony and transport | Charged by the provider, not by Vapi | Not Vapi. Your telephony provider |
| Optional add-ons | ||
| HIPAA compliance | $2,000 per month | Vapi |
| Zero data retention | $1,000 per month | Vapi |
Two things worth noticing. The compliance add-ons are priced for companies that have a compliance officer, so if you run a clinic and HIPAA matters to you, that line belongs in the budget conversation on day one rather than after launch. And bringing your own API keys drops three of the four pass-through costs to zero on the Vapi side, which is a lever worth asking any quote about.
Worth being blunt about the shape of this: Vapi and the phone carriers bill by the minute, so a busier month costs more than a quiet one. Nobody removes that, and this page is not going to pretend otherwise. What Acuity does is keep its own fee off the meter, so the only thing that moves with your call volume is what the platforms charge you directly, at the rates they publish.
Want the number first?
Run your own call volume through the missed-call cost calculator. If the number bothers you, book a call. If it doesn’t, you probably don’t need this yet and we’ll say so.
Book a 30-minute callBuilt for you, owned by you.
Where the agent lives. In a Vapi account with your name on it and your card on file. If we set it up first, it gets handed over at launch. You are never renting your own phone agent back from us.
What we charge. A one-time build fee, custom-scoped, starting at $1,500. After that a care plan is optional, starting at $75 per month for monitoring, prompt tuning, and small changes. A higher, more active management tier runs closer to $300 per month if you want someone in it more often. Acuity does not bill you per call or per minute.
What the platforms charge. A real cost, and one you do pay. Vapi, your model provider, and your phone provider bill you directly at their published rates, the ones in the table above. That bill moves with your call volume. Acuity adds no markup to it and never touches it.
What happens if you fire us. The agent keeps answering. Every prompt, workflow, transfer rule, and n8n automation stays in accounts you control, documented well enough for someone else to pick up.
One-time build from $1,500 · optional care plan from $75/mo · platform usage billed to you at costThat last point is the whole reason the arrangement is shaped this way. A voice agent is the front door of your business. Nobody should be able to switch it off because an invoice bounced. If you want the full picture of how a build runs from first call to launch, the process page walks through it, and the pricing page covers what moves the number. If you run a clinic, the medspa page gets more specific about consult calls and no-shows.
The things people ask before booking.
Do you write code for this, or configure the platform?
Configure and build on the platform. Vapi supplies the voice stack: speech to text, the language model, text to speech, and the telephony leg. The work is deciding what the agent says, how the call branches, when it transfers, what data it pulls, and where the result lands. That gets built in Vapi’s own tooling, plus n8n workflows for the connections to your calendar and CRM. There is no custom application sitting underneath it for you to maintain.
Who owns the Vapi account and the phone number?
You do. The agent gets built in an account under your name and your billing, or we set it up and hand it over at launch if you would rather not deal with the setup. Built for you, owned by you. That also means the usage bill is yours: you pay Vapi and your phone provider directly, at their published rates. Acuity doesn’t mark that up, and doesn’t bill you per call or per minute.
What does a Vapi build cost through Acuity?
Builds start at $1,500 as a one-time, custom-scoped project. After that, a care plan is optional and starts at $75 per month, which covers monitoring, prompt tuning, and small changes. A higher, more active management tier runs closer to $300 per month if you want that level of involvement. Platform usage is a separate cost you do pay: Vapi, your model provider, and your phone provider bill you directly at their published rates, and that bill scales with how many calls you take. Acuity doesn’t mark it up and doesn’t bill you per call or per minute. Vapi publishes its rates at vapi.ai/pricing.
How do you know the agent works before it answers a real customer?
Vapi ships Voice Test Suites. An AI tester places a real call to your agent, follows a script that simulates a caller, and the transcript is scored against a rubric you define. That covers the paths you thought of. The rest gets tested by hand: the caller who interrupts mid-sentence, the one who gives half an address, the one who asks something the agent was never told about.
Can the agent hand the call to a real person?
Yes, and there is more than one way to do it. Vapi supports a blind transfer that simply connects the two parties, and warm transfers where the agent speaks a summary or a custom message to whoever picks up before the caller is joined. Which one you want depends on whether the person receiving the call needs context before they say hello.
Do I need n8n, or is Vapi enough on its own?
Vapi is enough to answer the phone and hold a conversation. It is not enough to book the slot in your calendar, write the lead into your CRM, and text the caller a confirmation. Vapi calls out to a webhook during the call when the agent needs to do something in the real world. n8n is what sits on the other end of that webhook and talks to your actual tools.
More general questions are answered on the FAQ page, and there’s a longer write-up on after-hours calls in this post.
Let’s find the first call you’re losing.
Thirty minutes. We’ll tell you what we’d build on Vapi, what it would cost to build, what the platforms will charge you to run it, and whether it’s worth doing at all.
Book a 30-minute call