LiveKit and Pipecat: How telephony and conversation management are distributed

What belongs to LiveKit and what belongs to Pipecat?

LiveKit is the border between telephony and real-time media, and Pipecat is the conversational data flow and process management layer. Separating their responsibilities reduces the clutter of media, instructions, and business logic.

Why does a node-based process increase accuracy?

Putting every task into one large instruction increases uncertainty. A node-based flow gives the language model only the context, expected fields and tools required at that stage, while each state transition is validated separately.

Pipecat's official Flows model describes a conversation as a graph. This approach is particularly useful for bookings, sales, surveys and identification, where conversation is natural but the boundary of action must remain clear.

Node is not just a textual question. It has entry condition, expected data, check, tool contract, success and recovery transitions. Therefore, the editing process in the panel and the actual behavior of the sound environment must be based on the same versioned model.

How does the system scale across multiple companies?

Multi-company voice AI scales when the number, SIP channel, process version, secret data, records and billing are separated by each workspace key. A voice handler can be launched from a common system image, but each call context is company-specific.

• SIP data is stored encrypted and never transmitted to the browser.
• A distribution rule associates a call with a specific workspace, number, and product.
• Processor concurrency is measured against processor, network, and vendor limits.
• Record, transcript and tool output do not cross company data boundary.
• Release version and process version per session are fixed for diagnostics.

Which protection rules should be built in?

Critical security rules must also exist outside of the agent's textual instructions: structured input, server authentication, data types, replay-safe execution, frequency limit, public address checking, audit event, and human-safe switching.

The NIST AI RMF considers risk management as part of the entire lifecycle, while the OWASP GenAI guidance focuses on the risks of instruction substitution and tool misuse. In voice AI, this means that what the user says is data, not system instruction.

OMO's approach to public safety does not claim to completely eliminate risk. The goal is risk visibility, a limited action surface, and that an uncertain response cannot create an unproven business action.