Summary

A leading global safety science company wanted to expand its use of generative AI without absorbing the high cost of sending every request to a single frontier model. Marlabs deployed PromptRouter®, its proprietary routing platform, to analyze the complexity of each request and direct it to the most cost-effective model capable of a strong response. On a large-scale data extraction project, the approach cut token usage by roughly 30% while preserving output quality. The engagement gave the organization a scalable, model-agnostic foundation for responsible generative AI.

Challenge

Generative AI has become central to knowledge work, but running it at scale is expensive. Like many enterprises, the client was sending every request to a single high-end model, whether the task was simple or complex. That pattern drove up token consumption and made a large data extraction workload costly to operate. The organization needed to reduce cost and token usage without sacrificing the quality of extracted results and to do so within its existing corporate and compliance guidelines.

Solution

Marlabs addressed the challenge with PromptRouter®, a proprietary platform that decides, in real time, which model should handle each request. Rather than treating every prompt the same, PromptRouter® evaluates complexity and routes work to the model best suited for it, reserving frontier intelligence for the tasks that genuinely require it. The engagement moved through four phases.

Use-Case Discovery and Assessment

The team began by studying the client's data extraction workload in detail. We mapped how requests were being handled and identified where expensive frontier models were being used for straightforward tasks. This assessment established a clear cost baseline and revealed the largest opportunities for savings.

Prompt Routing Configuration

With the baseline in place, the team configured PromptRouter® to score the complexity of each incoming request. Simple requests were routed to more affordable models, while complex requests continued to reach frontier intelligence. The routing logic was tuned so that cost savings never came at the expense of the response a task required.

Governance and Security Integration

Because responsible AI use was a priority, Marlabs embedded guardrails directly into the routing layer. Every request was checked against the organization's corporate and compliance guidelines before it reached a model. This gave the client confidence that cost optimization and governance were working together rather than in tension.

Optimization and Validation

Finally, our team validated the routing decisions against real workloads. We measured response quality, adjusted complexity thresholds, and tracked token consumption to confirm the system was performing as intended. Continuous tuning ensured the solution kept delivering savings while maintaining the accuracy of extracted results.

Results

By routing each request to the most appropriate model, the engagement reduced the cost of generative AI while preserving quality and control. Token usage on the data extraction project fell by roughly 30%, with no loss in output quality. Just as important, the organization gained a flexible foundation that can adopt new and cheaper models as the market evolves.

Impact

The engagement delivered measurable and strategic value:

• Reduced token usage by roughly 30% on the data extraction project, with no loss in quality.
• Lowered the cost of running generative AI by routing simpler requests to more affordable models.
• Preserved accuracy by reserving frontier models for the most complex requests.
• Kept every request within corporate and compliance guidelines through built-in guardrails.
• Positioned the organization to adopt newer, cheaper models as the landscape evolves.