Level AI bets specialised models can beat frontier AI on customer service

Level AI has launched Latitude, a family of seven models built for specific customer-service tasks, as it argues that enterprises do not need a frontier model for every workflow.

What happened

The model family covers transcription, redaction, summarisation, intent detection, inferred customer satisfaction, quality assurance and voice-of-customer analysis.

Level AI says the systems run on infrastructure it controls rather than relying entirely on rented general-purpose models. The company claims that the specialised models can match frontier-model accuracy on selected tasks at as little as one-fiftieth of the cost.

Those performance and cost claims come from the company and were not supported by enough independent benchmark detail in the accessible announcement.

Why it matters

Customer-service systems process high volumes of repetitive interactions, making latency and inference cost economically important. A smaller model optimised for one task may be easier to govern and cheaper to run than a broad model capable of many unrelated functions.

The trade-off is flexibility. Specialised systems may perform well inside a defined workflow but struggle when customer requests fall outside their training domain.

The bigger picture

Enterprise AI is likely to become a portfolio of models rather than a single universal system. Large frontier models may handle ambiguous or complex work, while smaller models perform high-volume tasks. Level AI’s launch reflects a growing market for domain-specific systems competing on reliability, privacy and unit economics—not headline benchmark scores alone.