Scale's Felix Su: The Importance of Architecture for Building Enterprise AI

Successful enterprise AI goes beyond simply adopting the latest technology; the key lies in building the right architecture and control mechanisms on top of AI primitives. Off-the-shelf LLMs are powerful, but they have limits when it comes to perfectly following a company's specific business logic, brand guidelines, and safety principles.

Felix Su, Head of GenAI Platform Engineering at Scale, explains that in enterprise environments, it's more practical to directly design and control systems rather than relying solely on the potential behavior of LLMs. In particular, combining ASR (speech recognition), LLM, and TTS (speech synthesis) offers greater control over inputs and outputs compared to multimodal models, and is advantageous for applying custom guardrails.

TIME AI was implemented using this approach. Scale indexed TIME's article corpus as a knowledge base and incorporated business logic into the system prompt, while building safeguards to prevent hallucinations and brand guideline violations. ElevenLabs' voice technology was combined with this to complete a more human-like conversational experience.

This architecture is applicable not only to journalism but also to a variety of enterprise use cases, including customer service. It stands to become a key solution for meeting customer demand for more immersive and human-like interactions that go beyond simple chatbots.