Oracle got AI working on itself only this year

Over the past two years, Oracle has invested tens of billions of dollars to build the infrastructure that powers other companies’ AI. But it wasn’t until April that its own employees received a broadly useful AI tool.

Co-CEO Clay Magouyrk admitted, “I don’t think we figured out how to make AI really that useful for ourselves.”

He was describing Oracle’s position a year ago at an internal town hall, according to Business Insider. At that time, the company had introduced AI to customer support, but not yet to its developers, finance, or sales teams.

This changed in April and May, when Oracle launched ChatGPT Enterprise and OpenAI’s coding tool Codex, according to Jae Evans, the company’s global chief information officer.

Her team supports about 160,000 employees, but before launching the tools, Oracle established corporate standards, security controls, and internal policies. Within three months, 80% of staff were using the new tools.

But when the bills came in, the company was surprised by the costs, Evans said, because the tools were so easy to start using.

Now, Oracle shows employees which AI models they use and how much each one costs. For example, GPT-6 Astra is about two and a half times more expensive than cheaper options, so routine work is directed to models like GPT-5.6 Terra.

JPMorgan has taken a similar approach, limiting engineers to $2,000 per month for Claude usage.

The productivity numbers are impressive. Evans said developers now produce in a week what used to take a team two to three quarters. However, this hasn’t sped up delivery to customers.

Magouyrk explained that writing code faster “doesn’t mean that suddenly everything is 1000 times faster.”

Oracle is still updating its testing, validation, deployment, and release management processes to keep up.

A similar pattern has emerged in security. Oracle was one of the partners given early access to Claude Mythos Preview, a model that Anthropic announced in April but released only to a select group through Project Glasswing.

According to Anthropic, the model can identify and exploit zero-day vulnerabilities in all major operating systems and browsers.

When Oracle tested it on its own code, it found more potential vulnerabilities in two weeks than it had in the previous year, Evans said.

However, with a false-positive rate of 60% to 70%, Oracle had to create a verification step before engineers could start fixing issues.

So the bottleneck has shifted rather than disappeared. The AI model writes code faster than the company can release it and finds flaws faster than anyone can verify them.

The more advanced model also costs two and a half times as much as the cheaper one. Oracle stands out for discussing these challenges openly with staff, even as it cuts jobs and spends more on data centers than it earns in a quarter.