“We no longer just choose an AI system; we choose an architecture: it is the architecture that ensures the solution’s performance, sovereignty, and sustainability.”
Behind every voice AI agent lies an architecture of varying robustness, which determines latency, the quality of user journeys, data sovereignty, and the ability to scale the solution over time.
Patrice Merrien, CTO of Telecom & Infrastructure at Zaion, explains why customer relationship managers need to focus on architecture just as much as on recognition or language models.
Why is the architecture for integrating AI just as strategic as the choice of the model itself?
A conversational AI solution is not a monolithic system; it relies on a chain of components ranging from telephony to ASR, NLU, language models, the orchestrator, and business systems. How these building blocks are integrated determines how errors propagate, the latency experienced by the customer, and the ability to maintain consistent quality at scale.
At a high level, what does a typical architecture for voice AI agents and assistant agents look like?
Simply put, the call passes through the contact center platform, is processed by an ASR system, analyzed by natural language understanding engines, and then routed to business services and, if necessary, generative models to summarize or enrich the information. The agent relies on a nearby workflow, featuring real-time transcription, classification of the call reason, and generation of structured summaries, all while adhering to real-time constraints.
What are the practical implications of latency on the customer experience and on team adoption?
Excessive latency causes pauses, overlapping, or unnatural responses that disrupt the flow of the conversation and undermine the perception of quality. From the Agent Assist perspective, if the system takes too long to transcribe or summarize, it becomes impractical for everyday use, which slows adoption and reduces its operational impact.
How does AI built into a CCaaS platform differ from a specialized provider like Zaion in terms of architecture and integration?
Native CCaaS AI solutions are integrated into the contact center platform and effectively address a range of basic, general-purpose needs, with “turnkey” telephony integration. A specialist like Zaion focuses on the core business of conversational AI, offering advanced integrations with CCaaS, CRM, and enterprise systems, the ability to build multiple models, and end-to-end AI governance.
How can we balance model performance with data sovereignty requirements (GDPR, sovereign cloud, regulated sectors)?
Data sovereignty comes into play at several levels: the nature of the data (transcripts, recordings), model training, inference, storage, and access control. By choosing models that can be deployed on a sovereign cloud or on-premises, anonymizing and encrypting data, and relying on certifications such as ISO 27001, organizations can balance performance and compliance.
Are large hyperscaler-type models compatible with a long-term strategy of sovereignty and independence?
They offer significant power and maturity, but they also create a risk of heavy reliance on a single provider, its roadmap, and its data localization decisions. A well-designed architecture allows them to be used where appropriate, while retaining the flexibility to combine them or replace them with other models that are more specialized or offer greater control.
What should a customer relationship manager expect from their technology partner in terms of AI governance and continuous improvement?
Beyond the initial project, structured AI governance is essential: continuous performance monitoring, business-driven QA, model version management, and continuous improvement processes. This also requires close collaboration between business units, IT, and the software vendor, with shared processes to ensure quality and prioritize updates.
How can a “well-designed” architecture prepare for the future (new AI building blocks, new use cases, new channels) without starting from scratch?
By placing the orchestrator and APIs at the heart of the system, the architecture can integrate new AI components or use cases without having to overhaul the entire system. This composable approach prevents lock-in to a single stack and allows organizations to leverage innovations while maintaining the stability of customer journeys.
The carbon footprint of AI is becoming a hot topic: how can architecture and model selection help limit this impact?
The carbon footprint of an AI service is primarily linked to training and model selection, but also to how models are deployed and shared. Using SLM for routine interactions and reducing model size while optimizing their use helps lower the impact per conversation, in terms of both CO2 emissions and water consumption.
If you had to name three architectural “red flags” to watch out for before launching a customer relationship AI project, what would they be?
Warning signs include total reliance on a single vendor with no room for negotiation, the absence of formalized AI governance, and uncontrolled latency for real-time use cases. Another indicator is the lack of clear answers regarding data location, anonymization mechanisms, and security certifications.
Patrice Merrien
CTO of Telecom and Infrastructure at Zaion
