Oracle's Internal AI Rollout Lagged Years Behind Its Cloud Business

Oracle has spent the past two years constructing one of the world's largest AI cloud infrastructures, helping customers deploy large language models at scale. But the company's own internal adoption of generative AI tools lagged significantly behind that external buildout, according to remarks from executives at an internal town hall this week.

Co-CEO Clay Magouyrk told employees that as recently as last year, Oracle had not identified ways to make generative AI broadly useful across its own workforce. While the company had made some progress applying AI to customer support functions, deployment across core business units—including engineering, finance, and sales—remained limited.

The turning point came in spring 2025

That changed in April and May when Oracle rolled out ChatGPT Enterprise and OpenAI's Codex development tool companywide, Chief Information Officer Jae Evans explained. Rather than simply granting access, Oracle established corporate standards, security controls, and usage policies that drove adoption to 80% within three months.

The rapid uptake created unexpected challenges. "We made it so easy to use and adopt that we might have gotten a little bit of sticker shock," Evans said. The company now provides visibility into which models employees use and their associated costs. OpenAI's GPT-6 Astra, for example, costs 2.5 times more than alternatives like Terra that can handle routine tasks.

Speed gains create new bottlenecks

The impact on software development has been dramatic. Evans noted that developers now generate code in a week that previously required teams of engineers two to three quarters to complete. But Magouyrk emphasized that faster code generation hasn't translated to proportionally faster product delivery.

"When you make the actual act of writing the code quicker, it doesn't mean that suddenly everything is 1000 times faster," he explained. Oracle is still redesigning its testing, validation, deployment, and release management processes to match the new pace of development.

The company also gained early access to Anthropic's Mythos Preview model for security vulnerability scanning. In its first two weeks, the tool identified more potential issues than Oracle had found in an entire year. However, the model produced false positives at a 60% to 70% rate, requiring Oracle to build new verification processes before engineers could act on the findings.

Why it matters

Oracle's experience illustrates a pattern emerging across enterprise AI adoption: powerful models can dramatically accelerate specific tasks while simultaneously creating cost pressures and shifting bottlenecks to adjacent processes. Companies like JPMorgan have already begun setting cost limits for employee AI usage. The challenge for technology leaders isn't simply deploying AI tools—it's redesigning entire workflows to capture the productivity gains without overwhelming downstream systems or budgets.

These details were first reported by Business Insider.