Oracle executive admits company's AI adoption faced challenges

For years, Oracle has been investing billions of dollars to help its clients integrate advanced artificial intelligence into their businesses. However, the company’s own internal adoption of generative AI only gained traction much later. During an internal town hall this week, co-CEO Clay Magouyrk admitted that, as recently as last year, Oracle had not yet found a way to make generative AI broadly valuable for its own workforce. While some strides were made in areas like customer support, AI was not yet transforming work across diverse business units such as development, finance, and sales.

That situation began to shift this spring. According to Oracle Chief Information Officer Jae Evans, the company rolled out ChatGPT Enterprise, along with OpenAI’s coding tool Codex, to employees in April and May. Rather than provide open access to the tools, Oracle implemented corporate standards, strong security protocols, and clear internal policies. Within three months, the company reached an 80 percent adoption rate among employees. Evans described the transition as so seamless that Oracle was surprised by the associated costs. To improve visibility, Oracle has implemented tracking for which AI models employees are using and their respective expenses, she said.

The cost burden from high-end models quickly became clear. For instance, OpenAI’s GPT-6 Astra is 2.5 times more expensive than other models. Evans referenced Terra, a lower-cost alternative suited for routine tasks, as a way for teams to reduce spending. Despite these efforts, usage rapidly outpaced initial expectations, leading Oracle to educate teams on the varying expenses tied to different AI models.

Magouyrk and Evans acknowledged the real productivity gains achieved by the AI rollout, particularly in software development. According to Evans, code that formerly took teams two to three quarters to produce can now be generated by developers using AI tools in just a week. However, Magouyrk cautioned that this acceleration exposed bottlenecks elsewhere, particularly in testing, validation, deployment, and release management. Faster coding did not immediately translate to faster product releases, and Oracle is now focused on overhauling its software development lifecycle to match the new pace of code generation.

Oracle also received early access to Anthropic’s Mythos Preview model, an AI tool designed to scan code for security vulnerabilities. Within its first two weeks, Mythos flagged more potential security issues than Oracle’s teams had discovered over an entire year. While powerful, the model also produced a high number of false positives, with Evans estimating a 60 to 70 percent rate, requiring Oracle to build new verification processes before tasking engineers with potential fixes. Mythos, which was announced in April and is considered among the most advanced AI cybersecurity tools, has not been publicly released.

Oracle’s experience highlights both the promise and challenges of adopting generative AI at scale. While the technology can boost productivity for individual tasks, it can also shift bottlenecks and drive up costs, requiring companies to rethink both workflows and spending controls. As other enterprises race to implement similar technologies, many, including JPMorgan, have introduced cost limits and strict oversight for generative AI deployments. For Oracle, the journey of integrating AI internally remains a work in progress as it seeks to balance efficiency, security, and cost in a rapidly evolving landscape.