Self-Hosted vs Managed Voice AI: How to Choose

What managed and self-hosted mean

Every AI voice agent is built from the same parts: telephony or WebRTC to carry the call, speech-to-text, a language model with tools, text-to-speech, and orchestration that streams audio between them and handles turn-taking and interruptions.

Managed platforms such as Vapi, Retell, Bland and Synthflow run the orchestration for you. You configure prompts, voices, tools and phone numbers in their dashboard or API; they handle scaling, uptime and usually the telephony. You pay a per-minute platform fee on top of speech, model and phone costs.

Self-hosted frameworks, mainly LiveKit Agents and Pipecat, are open-source code your team runs. You write the agent, deploy it with its media servers in your cloud or data centre, connect your own phone system and choose every model. There is no platform fee, but you own the engineering and operations.

It's a spectrum: four realistic options

The middle options are often overlooked. Running LiveKit Agents or Pipecat on their makers' cloud services gives you open-source code you can take elsewhere, at a lower per-minute fee than most managed platforms, without operating servers.

Managed vs self-hosted, side by side

How cost changes with volume

Speech-to-text, the language model, text-to-speech and telephony cost roughly the same whichever way you host, because you pay those providers either way. The real difference is the platform or hosting layer.

Prices change often and depend on plan and volume; check each vendor's pricing page before budgeting. The pattern, though, is stable: a per-minute platform fee is small for a pilot and large at scale.

Two costs are easy to forget on the self-hosted side: the engineering to build and maintain the agent, and on-call operations. At a few thousand minutes a month they outweigh any savings; at hundreds of thousands of minutes they usually don't. Our AI voice agent cost guide works through per-minute costs component by component.

Data, privacy and compliance

For many organisations this decides the question before cost does. A phone call carries personal data, sometimes health or financial details, and rules on where it may be processed vary by country and industry.

• Check whether each provider signs the agreements you need, such as a HIPAA business associate agreement, and on which plans.
• Check where each provider processes data, and whether in-region hosting is available.
• Decide how long recordings and transcripts are kept, and who can access them.
• Remember that your own compliance team makes the final assessment, whichever option you choose.

Our guide to HIPAA requirements for voice AI agents and voice AI for banks go deeper on regulated industries.

Control, quality and telephony

For a closer look at the two main frameworks, see LiveKit vs Pipecat; for the managed platforms, see Vapi vs Retell vs LiveKit vs Bland.

What self-hosting actually takes

Self-hosting is not just deploying a container. A production voice agent needs the same care as any real-time service that customers depend on.

• Media servers and agent workers that scale with concurrent calls, with headroom for peaks.
• SIP connectivity to your PBX or carriers, with authentication, encryption and failover.
• GPUs or hosted endpoints if you run your own speech or language models.
• Monitoring of latency, errors and call quality, with alerts and an on-call rota.
• Logging, transcripts and retention rules that match your data policy.
• A test suite of real-style calls, run before every prompt, model or code change.
• Security updates for the framework, media servers and operating system.

This is where a partner earns its keep. Many teams keep the agent logic in-house and have a partner run the platform, or the reverse. Our self-hosted LiveKit deployment guide covers the infrastructure in detail.

Costs people forget on each side

Headline per-minute prices rarely tell the whole story. These are the costs that most often surprise teams, in both directions.

The fair comparison is total cost per resolved call over a year or two at your expected volume, including the people needed to run each option, not the per-minute rate alone.

Avoiding lock-in, whichever you choose

The best protection against lock-in is to keep your business logic in your own systems. If the agent books appointments by calling your own booking API, that API works the same whether a managed platform or a self-hosted agent calls it. Only the thin wiring changes.

# Self-hosted (LiveKit Agents): the tool simply calls YOUR booking API.
# A managed platform would call the same API through its webhook tools.
from livekit.agents import Agent, RunContext, function_tool
import httpx

BOOKING_API = "https://api.example.com/v1"

class BookingAgent(Agent):
    def __init__(self) -> None:
        super().__init__(instructions="Help callers book appointments. Confirm date and time back to them.")

    @function_tool()
    async def check_availability(self, context: RunContext, date: str) -> dict:
        """Return open appointment slots for a date in YYYY-MM-DD format."""
        async with httpx.AsyncClient(timeout=5) as http:
            r = await http.get(f"{BOOKING_API}/slots", params={"date": date})
            r.raise_for_status()
            return r.json()

• Keep prompts and call flows in version control, not only in a vendor dashboard.
• Expose business actions as your own APIs with clear inputs and permissions.
• Keep your phone numbers on your carrier or PBX, and route to the agent over SIP.
• Store transcripts and analytics in your own systems as well as the vendor's.

Questions to ask a managed vendor

• What is the all-in cost per minute at our expected volume, including speech, model, voice and telephony?
• Which features are paid add-ons, and which plans include them?
• Can we bring our own SIP trunk and keep our numbers on our carrier?
• Where is call data processed and stored, and for how long? Is in-region hosting available?
• Can we export prompts, call flows, transcripts and analytics if we leave?
• What concurrency and uptime commitments come with each plan?
• How often do you change models or pricing, and how much notice do we get?

How to decide

A common, sensible path: prove the use case on a managed platform for a few weeks, with business logic in your own APIs, then move to a self-hosted framework once the numbers justify it.

A real example: the hybrid route

CelloIP's AI voice agent for a North American contact centre sits in the middle of the spectrum. The agent runs on the open-source LiveKit Agents framework, connected over SIP to the contact centre's own FreeSWITCH platform, while speech-to-text, the language model and text-to-speech use hosted APIs. That gave the centre full control over telephony, routing and handover, kept the code portable, and kept the build fast by not running its own models.

What it costs to build

Ranges match our AI voice agent development page. Running costs are separate.

Frequently asked questions