Oracle has described an architecture that pairs the LiveKit Agents framework with Oracle Cloud Infrastructure to automate real-time voice calls. The design links telephony, AI models and business services, while placing internal components in private subnets. For SMEs, the central question remains integration, security and running costs.
Editor’s note. This Oracle post is an architecture overview published by the vendor, not the announcement of a turnkey service nor an independent assessment of its performance.[1]
Oracle presents an architecture blueprint, not a ready-to-use deployment
Published on 26 September 2026, Oracle’s post explains how LiveKit Agents can be combined with OCI services to build outbound voice agents. The use case describes automated calls, for example to confirm a delivery, qualify an enquiry or schedule an appointment. The LiveKit framework, available as open source and self-hostable, orchestrates the voice session; OCI provides the infrastructure and, depending on the set-up chosen, AI, storage and database services.[1], [2]
The distinction matters: Oracle is documenting one possible architecture. The post alone does not demonstrate conversation quality, cost savings or the reliability achieved by a business in production. Those outcomes depend on the models chosen, the telephone network, business integrations and the level of oversight.
From speech to action in the information system
A voice agent combines several stages. The system first detects when someone is speaking, converts the audio into text, passes the content to a model that prepares a reply or triggers a tool, then converts that reply back into audio. LiveKit Agents coordinates these exchanges within a real-time session. LiveKit also documents outbound calls via a SIP trunk, the service that connects the application to a telephone carrier and the public network.[2], [3]
The agent can, for example, query a business tool to look up an order status or record the outcome of a call. This requires properly defined application interfaces and permissions. Voice then becomes a gateway to actions in the information system, with the same control requirements as an API or a user account.
A network boundary that remains partly exposed
In Oracle’s design, the media entry point and the connections required for SIP or WebRTC sit in the externally accessible part. The agents, tool servers, databases and secrets, on the other hand, can reside in private subnets. The post notably cites OCI Generative AI, OCI Speech, Autonomous Database, Object Storage and Vault among the services that can be used.[1]
This separation limits the direct exposure of internal components, but it does not make the whole call private. The stream passes through a telephone provider and may rely on external AI services depending on the configuration. OCI network security groups are used to define the permitted flows to the relevant resources; configuring them, managing identities and logging remain the organisation’s responsibility.[4], [5]
What this architecture changes in the IT ecosystem
The project is not purely a matter of cloud or AI. It requires coordinating telephony, networking, cloud services, business data, identity management and monitoring. A failure or misconfiguration in any one of these links can degrade the call or give the agent overly broad access. Success therefore depends on the coherence of the entire architecture, not just on the choice of voice model.
For SMEs, this approach can reduce the need to build each stage of the conversation separately, but it does not remove the integration work. The system has to be connected to the applications already in use, the operations the agent can perform must be restricted, the cases where an employee takes over must be defined, and component maintenance must be planned. A trial on a simple flow, such as confirming an appointment, makes it possible to assess the value before scaling up.
Analytical reading of the architecture
The table summarises the benefits described by Oracle and the questions an organisation must address before applying them to its own system.
| Component | Operational effect | Control question |
|---|---|---|
| Telephony and media access | Connects the agent to an inbound or outbound call via SIP and the telephone network. | Who provides the trunk, which connections are public and how are they filtered? |
| LiveKit Agents | Coordinates audio, transcription, the model, speech synthesis and tool calls. | What latency and quality are measured on real lines, in real languages and environments? |
| OCI services | Hosts the agents and can provide AI, storage, database and secrets management services. | Which services process each piece of data, in which region and at what usage cost? |
| Business tools | Allow operational data to be viewed or modified during the call. | Are permissions limited to each task, logged and reversible in the event of an error? |
| Monitoring and continuity | Logs and events help track sessions and diagnose incidents. | When is the call transferred to a human, and what happens in the event of an outage or uncertainty? |
Before deployment, test use cases, data and costs
The first check concerns quality in real-world conditions: background noise, accents, interruptions, response times, transcription errors and the rate of transfer to an adviser. The full cost also needs to be measured, adding up infrastructure, phone calls, voice services, models, integration, monitoring and maintenance. Oracle’s post describes the components, but provides neither an independent comparison of these costs nor generalisable production results.[1], [2], [3]
Recordings and transcripts may contain personal data, including the voice itself and information shared during the call. The business must therefore decide what is retained, for how long, with what access and for what purpose, then inform the people concerned where the applicable rules require it. The CNIL points out that a voice constitutes personal data and regulates how callers must be informed when calls are monitored or recorded.[7], [8]
If the calls are used to prospect private individuals in France, the rules applicable since 11 August 2026 must also be checked: the CNIL states that prior consent is now required, save for specified exceptions, notably in the context of an existing contract. The applicable regime may differ depending on the target and purpose of the call; a business must therefore classify its use case before automating it.[6]
The key IT skill: integrating services into a controlled architecture
The topic highlights a core skill: knowing how to integrate cloud, telephony and business services into a secure network architecture. This means understanding data flows, restricting access, monitoring exchanges and planning for degraded modes. It is precisely a cross-cutting view of IT, going beyond simply developing a conversational agent.
For SMEs, start with a controlled trial
The Oracle and LiveKit architecture shows how a voice agent can be connected to telephony and business tools while isolating part of its execution in the private cloud. It is not enough, however, to establish that a deployment will be cheaper, more reliable or suitable for every organisation. For SMEs, the next step is to test a limited use case, measure its quality and cost, then check access security, how conversations are handled and the hand-off to a human when the agent reaches its limits.
READ ALSO
To place the Oracle and LiveKit architecture in the context of current developments in cloud and enterprise agents, explore these four complementary analyses published on CCCLX.
Kimi K3, Mistral, DeepSeek: Oracle turns OCI into an AI model hub
Agentic Data Cloud: Google rethinks data for the age of AI agents
Enterprise AI Harness: Salesforce tackles AI agent chaos
With Lambda, AWS readies cloud infrastructure for autonomous AI agents
Sources
- 01
Oracle (26 September 2026). “LiveKit Agents Framework: Real-Time Outbound Voice AI on Oracle Cloud Infrastructure”. Primary source for the architecture.
- 02
LiveKit. Agents documentation. Real-time media orchestration and deployment.
- 03
LiveKit. Outbound calls. Telephone connection via SIP.
- 04
LiveKit. Self-hosted SIP server. Components and network requirements.
- 05
Oracle Cloud Infrastructure. Network security groups. Controlling flows to OCI resources.
- 06
CNIL. Prospection commerciale par téléphone. Applicable French rules.
- 07
CNIL. Définition d’une donnée personnelle. A voice can constitute personal data.
- 08
CNIL. Informing callers when calls are monitored or recorded.
AI propels Nokia: data centres are reshaping the networking market
The AI explosion is transforming the infrastructure market Nokia is reaping the full benefits of the rise of artificial intelligence and the proliferation of data centres that comes with it. In the second quarter of 2026, the company’s Network Infrastructure business...
Gemini 3.5 Pro delayed: Google under pressure in the AI race
The race for AI models is speeding up Google is going through a difficult period in the competition over advanced artificial intelligence models. According to several reports in the specialist press, the launch of Gemini 3.5 Pro has reportedly been pushed back by...
Project Glasswing: Anthropic goes on the offensive against software vulnerabilities
A new step for AI applied to cybersecurity Anthropic is pressing ahead with Project Glasswing, an initiative that makes its Claude Mythos Preview model available to specialist cybersecurity players in order to identify and fix vulnerabilities in critical software. The...
Cloud, AI, defence: AWS steps up on every front
AWS makes artificial intelligence a pillar of critical infrastructure At the AWS Summit Washington, D.C. 2026, Amazon Web Services unveiled several developments aimed at government departments, public bodies and critical sectors. The company is strengthening its...
MDASH: the AI helping Microsoft secure Windows
Microsoft hands part of the bug hunt to artificial intelligence Cyberattacks are evolving ever faster, forcing software vendors to detect vulnerabilities before they can even be exploited. To meet this challenge, Microsoft has unveiled MDASH (Microsoft Detection and...
Microsoft speeds up its cybersecurity with AI: defence moves to machine scale
Microsoft wants to regain the upper hand against cyberattacks Cybercriminals are now using artificial intelligence to discover vulnerabilities, automate their attacks and speed up their malicious campaigns. In response, Microsoft is pressing ahead with its Secure...
Free access to Claude Fable 5 extended as Anthropic holds off
A new generation of AI designed for specific uses OpenAI is changing its strategy with the launch of GPT-5.6, a new family of models made up of Sol, Terra and Luna. Rather than offering a single model capable of meeting every need, the company is now betting on a...
OpenAI bets on modular AI with Sol, Terra and Luna
A new generation of AI designed for specific uses OpenAI is changing its strategy with the launch of GPT-5.6, a new family of models made up of Sol, Terra and Luna. Rather than offering a single model capable of meeting every need, the company is now betting on a...
Anthropic’s AI is advancing at a pace that worries cybersecurity specialists
AI models cross a new capability threshold Barely a month after its launch, Claude Mythos Preview is already improving its performance. According to the latest tests published by the UK’s AI Security Institute (AISI), Anthropic’s model now beats its own results from...
Lumo AI 2.0 speeds up without sacrificing data protection
Proton wants to prove that a high-performing AI can also be private The artificial intelligence race is no longer played out solely on model performance. User trust and data protection are now becoming major differentiators. With Lumo AI 2.0, Proton is upgrading its...
