Scale AI launches enterprise model customization with NVIDIA, citing insurance-call gains

Scale AI opened its enterprise model-customization service on October 15, 2026, alongside a collaboration with NVIDIA. Its insurance example shows both the promise and the remaining work: Scale says specialized training improved accuracy and cut running costs, but the French-language call copilot has been evaluated on synthetic conversations and is not yet in production.

The offering, Scale Post-Training, adapts existing open-weight models—including NVIDIA’s Nemotron 3.5 Super VL—for particular business jobs. Scale measures a customer’s current system, develops task-specific training and evaluation data, then tests successive versions within the intended workflow. The models can run inside a company’s own infrastructure.

Three jobs during a roadside-assistance call

The unnamed insurer’s copilot is designed to help representatives as French-language calls unfold. Scale’s technical case study separates the work into three tasks, evaluated individually:

• Extract case facts: identify the event, breakdown reason and passenger count, leaving fields blank when information is insufficient.
• Recommend coverage: match facts and policy rules to benefits, requesting missing details rather than assuming coverage.
• Check promises: flag commitments beyond the policy, such as reimbursing an uncovered taxi fare.

Legal approval is required before the insurer’s call data can be used for training, so Scale generated French conversations from sampled facts and policy scenarios. Scoring requires complete matches: all three extracted fields must be right, and coverage recommendations fail if even one benefit is missing or extra.

Training improved two tasks by different amounts

Scale first trained each task on successful reasoning and tool-use examples generated by the model. It then used reinforcement learning, rewarding answers that exactly matched verified results. Training compute cost about $890 for fact extraction and $230 for coverage recommendations.

Cheaper answers still have to arrive in time

Scale puts fact-extraction running costs at about $3 per 1,000 queries, versus $12.10 for the production model. Coverage recommendations cost about $1.40 per 1,000 queries, compared with $4.80.

A lower reasoning setting offered another tradeoff. Fact-extraction accuracy fell to 91.9%, while cost dropped to $1.80 per 1,000 queries and response time fell from 26 seconds to 11. Scale still plans to measure latency inside the live workflow and evaluate approved real calls before deployment.

The promise-checking guardrail also remains unfinished. Violations occur in only about 4% of checks in Scale’s evaluation. The guardrail must catch those cases without raising false alerts on the rest—a separate challenge from extracting facts or choosing benefits correctly.