Key Takeaways:
- OpenAI's Codex is now used by nearly all employees weekly for knowledge work
- Google's finance agent saves $200 million a year on invoice overpayments
- Only 13% of companies have adequate AI-agent governance, Gartner found
Key Takeaways:

OpenAI, Google and Anthropic are using their own AI agents to automate complex internal workflows, revealing both the productivity gains and the governance challenges that await the broader enterprise.
The companies building the most powerful AI tools are also their most aggressive internal customers, deploying autonomous agents for tasks from invoice validation to contract review — and discovering that productivity gains create their own problems.
"People started to use it for other things," Kelsey Pedersen, a Codex deployment engineer at OpenAI, said of the coding tool that nearly all employees now use weekly for general knowledge work.
Google's finance team deployed an invoice-validation agent last year that compares vendor invoices against contract terms, allowing the company to review five times more invoices. The agent is on track to save $200 million annually on overpayment issues, according to Kristin Reinke, a vice president leading AI implementation in Google's finance organization. At Anthropic, marketing operations staff use Claude agents to automate data imports that previously took 15 minutes to an hour per task.
Yet agentic workflows introduce new frictions. Gartner estimates the average Fortune 500 company will run more than 150,000 AI agents within two years, but only 13% of organizations have adequate governance in place. Google's invoice agent worked so well it created a backlog of flagged discrepancies that now requires another agent to contact suppliers — a cascading automation cycle that executives say will define the next phase of enterprise AI adoption.
From Coding Tool to Corporate Workhorse
OpenAI's Codex, originally designed for software developers, proved intuitive enough for nontechnical teams including marketing, recruiting and legal to adopt. Ashton Summers, an account director on OpenAI's go-to-market team, used Codex to investigate a customer billing issue — work that previously required the billing and operations team. The agent built a daily-updating customer dashboard, created product demos for prospective clients, and generated a transition document for a new hire by scanning 30 minutes of emails and Slack messages.
Nicole Diaz, associate general counsel at OpenAI, has Codex perform work a junior associate would typically handle, including analyzing conflict-of-interest disclosures from new employees. The agent might flag that someone remains on another company's board or has a relative working at a competitor. Diaz said she is still hiring junior associates — in part to review Codex's output.
The 10X Problem
Partha Ranganathan, an engineering fellow at Google, described what he calls the "10X problem": when one workflow accelerates tenfold, something else in the system breaks. Google's invoice agent flagged so many discrepancies that the operations team could not keep up with supplier follow-ups. The solution is another agent that will initiate supplier contact automatically.
Jeremy Korst, a partner at the AI advisory firm Mindspan Labs, said smaller companies have been the most nimble with agents, while cross-team coordination creates friction at larger organizations. A legal team may not want the sales department using AI to conduct its own contract reviews, he said. "There is real friction," Korst said. "This is a very common conversation."
The Investment Angle
For investors, the internal deployment patterns at OpenAI, Google and Anthropic offer a preview of the enterprise AI market's trajectory. Google's parent Alphabet trades at about 22 times forward earnings, and the $200 million in annual savings from a single finance agent — while modest relative to its $350 billion market cap — demonstrates the cost-reduction potential that chief financial officers are beginning to quantify. Gartner's finding that 23% of tech leaders spend $200 to $500 per developer per month on AI coding tokens suggests the addressable market for agentic tools is expanding rapidly, even as token-based pricing introduces cost volatility that could reshape enterprise budgets.
This article is for informational purposes only and does not constitute investment advice.