LegalOn Technologies, which provides AI‑assisted legal and business services, reduced its estimated daily AI spending by about 65% while keeping development speed intact. The company achieved this by selecting among GPT‑6 Astra, GPT‑6.1 Sol and GPT‑6 Luna according to task complexity and development stage, restricting Fast mode by default, and applying budget caps differentiated by business stage.
The challenge
LegalOn initially allowed developers unrestricted access to its primary model, GPT‑5.5 in Fast mode, and expanded AI use across design, implementation and everyday work. This broad adoption threatened to push annual AI expenditures beyond planned budgets, creating a need to control costs without negating AI‑driven productivity gains.
The approach
The company’s AI‑powered Development Center of Excellence (AID CoE) developed guidelines for model selection, ran tests and monitored usage. Managers communicated findings to teams so engineers could select the most appropriate model for each task.
Three measures formed the core of the response:
- Task‑based model selection: move away from using the highest‑capability model for every task and choose among GPT‑6 Luna, GPT‑6.1 Sol and GPT‑6 Astra according to task complexity and the development phase.
- Default restriction of Fast mode: Fast mode was turned off by default, allowed separately when needed.
- Budget caps: monthly usage limits were set for departments and individuals; established businesses were expected to improve cost efficiency by roughly 20%, while new businesses in launch phases received more generous budgets to encourage active AI use.
Role of each model
Internal testing refined task‑level criteria:
- GPT‑6 Luna: the lightest model, used for code implementation when requirements are clear, running routine automations, relatively simple analyses and delegating straightforward tasks.
- GPT‑6.1 Sol: suited for standard design, data analysis and document preparation; used for simple software design, routine numeric analysis and tasks where faster turnaround than Luna is desirable.
- GPT‑6 Astra: the most capable model, applied to advanced judgment tasks such as architecture design, complex analysis and coordination of agents.
Teams typically start with a lighter model and escalate to more capable models as complexity rises.
Results
Combining model optimization, Fast mode restrictions and tiered budgets reduced estimated daily costs by approximately 65% compared with prior GPT‑5.5 usage, while maintaining development performance. The policy also aimed for about a 20% efficiency improvement in mature businesses, with active investment channeled toward new business lines.
Despite initial concerns, development speed remained stable through team practices like running tasks in parallel when Fast mode was unavailable by default.
Measuring value, not just speed
LegalOn is moving beyond metrics that measure only development speed or AI usage volume. Yuta Tokitake, Senior Engineering Manager at LegalOn Technologies, noted that although AI accelerates development, the key question is whether that acceleration creates real customer value. He said: “When using AI, we can track the cost of individual tasks, but the total cost of a complete piece of work—a feature release—often remains a black box. So we designed the measurement around each feature release as a single unit. By linking the customer value that release delivers to the actual AI costs invested in it, we aim to make the return on investment visible and accurate.”
The company is building a pipeline to compute this metric so that ROI can be assessed in terms of customer value delivered per feature release.
Spreading know‑how and organisational change
LegalOn’s next objective is to turn individual engineers’ AI expertise into a company‑wide capability. The firm plans to create a knowledge base that captures practical experience—such as best model combinations for design, implementation and review—so that successful local practices can be rapidly turned into organisation‑level standards.
These operational changes are also affecting hiring and organisational planning. As AI raises productivity, LegalOn is rethinking task allocation between people and AI and is prioritising AI skills in recruitment.
Governance and risk
AID CoE and the security team collaborate to build a flexible governance framework that balances performance, cost and risk while allowing teams room to experiment. The aim is an approach that preserves team autonomy in practical testing but ensures economical and secure AI usage at scale.
Conclusion
By tailoring model choice to task complexity, restricting high‑cost features by default, and aligning budgets with business stage, LegalOn cut AI costs substantially without sacrificing development speed. The company is now focused on measuring AI investments’ real customer value and institutionalising engineers’ AI knowledge across the organisation.
