LegalOn Technologies, which offers Professional AI for legal work, adopted Codex across its development process and expanded the practice to make the organization itself AI-native. Developers initially got unlimited access to the default main model, GPT-5.5 in Fast mode, and used it for design, implementation, and everyday work. Unlimited use of high-performance models created a budget risk, while blanket restrictions would undermine the productivity gains. The company’s response was to stop using the most capable model for everything and instead select among GPT-6 (Astra, Luna) and GPT-6.1 Sol based on task complexity and development stage, while also aligning budgets with each business’s growth stage. That combination halved costs without slowing development.

The AI-powered Development CoE (AID CoE) built guidelines for model selection, tested and monitored models, and managers shared findings with teams so each engineer could independently choose the right model for each task. System controls followed: administrator settings with monthly usage limits for departments and individuals. Internal testing made the selection criteria more specific. GPT-6 Luna, the lightest model, handles code implementation with clear requirements, everyday automations, and relatively simple tasks, mainly as a subagent. GPT-6.1 Sol supports standard design, routine numerical analysis, and document preparation, with shorter completion times than Luna. GPT-6 Astra, the most capable model, handles complex analysis, complex software design, and orchestration of agents. The company also restricted Fast mode by default, allowing individual requests only when needed; teams maintained performance by practices such as running tasks in parallel. Budget caps were applied at department, group, and individual levels: the established LegalOn business was asked to improve cost efficiency by up to approximately 20%, while new businesses in the launch phase received generous budgets to encourage active AI use, because new businesses prioritized business speed over cost efficiency.

Together, model selection, feature restrictions, and budgets tailored to each business reduced estimated daily costs by approximately 65% compared with GPT-5.5, keeping overall spending under control while directing resources toward growing businesses. The company is now looking beyond development speed as the measure of AI success. Yuta Tokitake, Senior Engineering Manager at LegalOn Technologies, says the real question is whether faster development translates into customer value. LegalOn is building a metric around each feature release as a single unit, linking the customer value that release delivers to the actual AI costs invested in it, to make the return on investment visible and accurate.

Next steps include a knowledge base to turn individual engineers’ AI know-how into organization-wide capability, capturing practical experience such as the best model combinations for design, implementation, and review. AI-driven productivity gains are also prompting the company to rethink how work is divided between people and AI, and to emphasize AI skills in hiring. AID CoE and the security team are working together to optimize performance, cost, and risk from multiple perspectives. Tokitake: “Excessive restrictions through rules and budgets can undermine an organization’s momentum. What we need is a flexible operating model that effectively balances risk control with the speed teams need.”