# On-premise voice AI for banks: how to keep call data inside the bank - Source: https://www.cbot.ai/on-premise-voice-ai-for-banks/ For a bank to use voice AI without data leaving its environment, speech recognition, speech synthesis, the language model, enterprise knowledge and logs all need to run on-premise. If any of these layers depends on an external service, data crosses the bank's approved network boundary at that point. A bank that wants to run Voice AI agents in its contact center without data leaving its environment needs every layer of the call inside its own infrastructure: speech recognition, speech synthesis, the language model, enterprise knowledge and logging. In that setup, the caller's voice, the transcript, the model output and the transaction record all stay within the bank's approved network boundary. What matters is where each link in the chain runs. ## When does a bank need on-premise deployment? Not every banking workload needs the same level of control. On-premise deployment usually becomes the right choice when: - The call involves identity verification, account, card or loan data. - Internal policy or sector regulation requires data to remain in a specific environment. - The information security team will not accept a mandatory dependency on an external AI service. - The bank operates its own data center and GPU capacity and wants to plan that capacity itself. - Call recordings and transcripts must be retained internally for audit. Deployment does not have to be one decision for the whole organization. In a hybrid architecture, non-sensitive workloads run in approved cloud environments while processes that touch customer accounts stay on-premise. ## What stays inside the bank? A voice call passes through several layers. In an on-premise architecture, each of them runs in an environment the bank controls: - **Speech-to-Text:** Converts the caller's voice to text in real time. - **Language model and understanding:** Identifies intent, evaluates context and decides which information or system the call needs. Model inference runs on the bank's own CPU and GPU resources. - **Enterprise knowledge:** RAG, vector storage and documents stay internal, so product, fee and procedure content is never sent out. - **Text-to-Speech:** Turns the response into natural speech and reads product names and abbreviations correctly using pronunciation dictionaries. - **Runtime and workflow:** AI agent orchestration, business rules, human approval and integrations. - **Records:** Voice recordings, transcripts, prompts, model outputs, analytics and audit trails. If any one of these layers sits outside, data crosses the boundary at that point. So "can the platform be installed on-premise" is not enough of a question. Buyers also need to confirm that model, speech and vector services run internally. ## How are regulation and data residency addressed? The regulations and internal policies that apply to banks differ by institution. Compliance is therefore not a product feature but the sum of architecture and operating decisions. In practice, these topics are handled together: - Where data is processed and where it is stored. - Role-based access, integration with enterprise identity systems and customer-managed keys. - Data masking, anonymization and configurable retention. - Which models are approved for which tasks, and control over external endpoints. - Audit trails, version control and rollback. - Human approval for high-impact actions. Final compliance depends on the use case, configuration, data and applicable regulation. Security, compliance and infrastructure teams should run the assessment together before deployment. ## IVR and contact-center integration Voice AI agents do not replace the existing contact center. They connect to it. Typical integration points are telephony and SIP infrastructure, the IVR environment, the contact-center platform, CRM, core banking and payment systems, and agent desktops. A traditional IVR walks callers through fixed menus. A voice agent lets the caller describe the need in their own words, completes the task or routes the call to the right person. During a handover, the detected intent, collected information, transcript and a call summary go to the human agent, so the caller does not have to start over. ## Evaluation criteria When assessing an on-premise voice AI solution, a bank should ask: 1. Does the full stack, including speech recognition, speech synthesis and the language model, run on-premise, or does any layer depend on an external service? 2. How does speech recognition perform on Turkish contact-center audio, noise and banking terminology? Is it tested on the bank's own call samples? 3. Does speech synthesis read product names, amounts and abbreviations correctly, and can the bank manage the pronunciation dictionary? 4. Is end-to-end latency suitable for a real-time phone conversation? 5. Do services scale independently with traffic, and are high availability and disaster recovery supported? 6. Can the solution be deployed on the container platform the bank already uses? 7. How are infrastructure, monitoring, security and update responsibilities shared between the bank and the vendor? 8. What level of audit trail, human approval and version control is available? ## What CBOT offers CBOT is an enterprise AI platform that can run fully on-premise with its own language models and speech technology. It supports SaaS, private cloud, hybrid and fully on-premise deployment. See [deployment options](/deployment-options/) for details. - **Proprietary speech technology:** CBOT develops its Speech-to-Text and Text-to-Speech technologies in-house under CBOT Speech, optimized for Turkish language structure, contact-center and mobile audio, noise and domain terminology. The [Voice AI](/cbot-speech-overview/) page covers inbound and outbound call scenarios. - **On-premise language models:** [CBOT LLM](/cbot-llm/) can be deployed on-premise. CBOT models, supported open-source models and customer-hosted models run without a required dependency on an external commercial model API. - **Data boundary:** In a fully on-premise architecture, prompts, conversations, voice streams, documents, embeddings, model outputs and analytics records can remain within the approved enterprise environment. - **Infrastructure:** Containerized deployment is supported on Docker, Kubernetes and Red Hat OpenShift. Speech-to-Text, Text-to-Speech, language-model inference and runtime services scale independently according to traffic. - **Governance:** Role-based access, data masking, audit trails, version control and human approval are part of the platform, described on the [AI governance and safety](/ai-governance-safety/) page. - **Integration:** CBOT connects to telephony and SIP infrastructure, IVR environments, contact-center platforms, CRM and core banking systems through approved integrations and APIs. Two banking examples involve voice. In the [Halkbank case study](/case-study/client-story-halkbank/), CBOT's Automatic Speech Recognition (ASR) and Text-to-Speech (TTS) technologies are integrated into the bank's IVR system and used in the call center, and the deployment is listed as on-premise. In the İşbank Maxi case study, the assistant expands to voice assistants and serves customers in both text and speech. Sector capabilities are summarized on the [banking](/industries-banking/) page. ## Conclusion On-premise voice AI lets a bank modernize the phone channel while keeping control of its data and infrastructure. The right architecture follows from the sensitivity of each workload and the bank's operating model. To assess your own scenario, [talk to the CBOT team](/contact-us/). ## FAQ ### Does on-premise voice AI require GPUs? GPU requirements depend on the selected models and workloads. Language model and some speech workloads may need GPU capacity, while other services can run on CPU. Detailed sizing is completed during solution design. ### What is the difference between private cloud and on-premise deployment? Private cloud usually runs in an isolated cloud environment. On-premise deployment runs on infrastructure the bank controls directly, such as its own data center. How infrastructure and operating responsibilities are shared depends on the chosen architecture. ### Can a bank still use commercial language models with an on-premise setup? Yes, in a hybrid architecture. Access to approved commercial models is managed through external or private endpoints the institution defines, while sensitive workloads stay on models that run internally. ### Can voice AI agents handle outbound banking calls too? Yes. Payment reminders, collections, application follow-ups, customer notifications and survey calls are typical outbound scenarios. Inbound and outbound calls share the same enterprise knowledge, systems and governance rules. ### How long does it take to launch on-premise voice AI? Timing depends on the use case, integrations, conversation complexity, deployment architecture, security requirements and test scope. The implementation plan becomes clear after the solution and infrastructure assessment.