Trained on more than 4 trillion tokens of customer service data and run on Level AI’s own infrastructure, the Latitude family scores every customer interaction, keeps sensitive data inside a controlled environment, and gets sharper with every correction

MOUNTAIN VIEW, Calif.: Level AI, the AI-native customer experience platform built on owned models and controlled infrastructure, today announced Latitude, a family of seven AI models each trained for one core job of customer experience: transcription, redaction, summarization, intent detection, inferred customer satisfaction, quality assurance, and voice of customer. Latitude replaces the general-purpose LLMs most CX vendors rent from third-party providers, and targets three problems those models can't solve in production: quality reviews that cover a fraction of interactions, customer data leaving the enterprise for external endpoints, and systems that don't improve until someone else ships a new model.

The privacy stakes are measurable. In a Level AI analysis of 50 million customer conversations sampled between January and April 2026, 29% contained sensitive data such as account numbers, medical identifiers, or financial records. In financial services and insurance, the figure was 47.1%. Latitude's redaction model strips sensitive data before it reaches any downstream system, and its quality assurance model scores 100% of interactions.

"Most AI vendors route every task, from transcribing a call to scoring a compliance disclosure, through the same general-purpose model," said Ashish Nagar, Co-Founder and CEO of Level AI. "We built each model to do one job well, running on infrastructure we own end to end, so customer data never leaves our environment."

The seven models:

• Transcription: automatic speech recognition built for contact center audio, including accented and noisy call conditions.
• Redaction: dual-layer redaction across transcript and audio, removing sensitive data before it reaches any downstream model.
• Summarization: turns every conversation into structured, actionable intelligence.
• Intent Detection: natural language understanding that replaces manually maintained keyword libraries.
• Inferred CSAT: predicts customer satisfaction from the conversation itself, without post-call surveys.
• Quality Assurance: GenAI-powered scoring applied to every interaction, not a sampled 1-2%.
• Voice of Customer: daily classification of every contact into a three-level hierarchy of customer issues and themes.

In internal testing against frontier LLMs across customer experience tasks, Latitude models delivered accuracy on par with frontier models at up to 50x lower cost to serve, up to 3.5x higher throughput, and 4x lower latency. Comparisons used the same evaluation harness, prompts, and pre- and post-processing for both model families, with cost measured as Level AI's fully loaded inference cost against published frontier API pricing on identical workloads.

"Our CARE team runs 1,200 to 1,700 agents across five countries, and for years 'quality' meant something a little different in every region,” said Paul Harraghy, Senior Vice President, Global CARE Team at VistaPrint. “Level AI scores every interaction the same way, everywhere, which brought our calibration variance from over 20% down to 9% and took hundreds of thousands of dollars of unwarranted credits out of the system, without shortchanging a single customer who deserved one. A general-purpose model reading a sample of our calls was never going to do that."

Because Level AI controls the full stack, from hardware allocation through the model layer to the workflow application, corrections made by QA leads and supervisors feed directly back into the models. The system improves through daily use rather than waiting on a third-party release cycle. The seven models are trained on more than 4 trillion tokens of CX-specific data annually, and additional task-specific models for sub-tasks are in development.

Latitude is available now to Level AI customers. More on the philosophy behind Latitude is available in Level AI's whitepaper, The Operating Standard for Enterprise AI.

About Level AI Level AI is an AI-native customer experience platform that helps enterprises improve customer experience through conversation intelligence, automation, and operational insights. The platform analyzes customer interactions at scale to automate quality assurance, agent coaching, compliance, and operational performance across enterprise contact centers. Level AI is backed by Adams Street Partners, Battery Ventures, and Eniac Ventures, among others, and serves more than 100 clients processing millions of customer interactions monthly. Learn more at thelevel.ai.

Fonte: Business Wire