See How AI Giants Are Using AI in Their Own Offices－By Katherine Bindley,WSJ
What does the future of work look like? The companies building the most popular AI tools are a good place to check.
While many companies across the U.S. have been struggling to get their employees to use artificial intelligence, OpenAI, Google and Anthropic are getting their workers to share complicated tasks with AI.
While many AI-use cases in professional settings might involve prompting AI to transcribe a meeting or write a memo, these companies are increasingly giving more work to “agents.”
Trusting an AI agent to go out and perform tasks is a leap of faith, and in some instances, it has led to mass email deletions, vanishing code and other mishaps . Even when things are going well, agents can create more work for humans.
But the companies that aim to market such tools to their customers say they are making measurable progress. Even the nontechnical departments are using agents to handle multistep tasks across applications. Humans then become “reviewers” of much of the work, effectively fact-checking the AI’s output.
Here’s how AI is changing work at the companies leading the charge:
OpenAI
The backbone of OpenAI’s automated workflow is Codex. Originally designed for software developers, the tool was intuitive enough for nontechnical teams like marketing and recruiting to pick up, the company says. Nearly 100% of employees now use it weekly.
“People started to use it for other things,” says Kelsey Pedersen , a Codex deployment engineer. “Now it’s just a general knowledge-work tool.”
Ashton Summers , an account director on OpenAI’s go-to-market team, got a customer note about incorrect billing. Previously, Summers would have gone to the billing and operations team to investigate the issue. Codex did all the legwork, which Summers then validated.
“This is removing a bunch of bottlenecks and removing reliance on other teams,” he says.
Codex also created a daily-updating dashboard with his customers and the metrics he cares about, like product adoption. Summers asks Codex to build demos for the companies he is trying to sell it to, turning it into a marketing tool for itself. It also spent about 30 minutes going through his emails and Slack messages and creating a transition document for a new hire.
Nicole Diaz , associate general counsel at OpenAI, has been having Codex do the work a junior associate otherwise would, including analyzing disclosures from new employees and drafting replies. Codex might flag that someone is still on the board of another startup or has a brother working at competitor Anthropic, for instance.
Diaz says she is still hiring junior associates, in part to review Codex output.
Alphabet’s Google likes to refer to itself as “customer zero,” the first to try out the products it will push to customers.
The finance team, consisting of a few thousand employees, last year launched an invoice-validation agent to compare vendor invoices with the contract terms to make sure everything checks out.
“Before, we had a team that actually just mind-numbingly went through the contract,” says Van Bui , head of Google Business Services. With the new agent, “we can review five times more invoices,” she adds.
Employees who used to do invoice validation now spend time doing higher-level audits and reviewing the AI’s work. “This is the new skill set that they’re learning,” says Bui. “They get to put on their résumés, ‘I do AI model training.’”
The agent is on track to save the company $200 million a year on invoice-overpayment issues, says Kristin Reinke , a vice president leading much of the AI implementation in Google’s finance organization.
The Google finance team has also been using AI to help manage cash across thousands of its bank accounts and suggest investments based on criteria including risk tolerance.
“What our agent will do is give a recommendation: Here’s what we think you should be investing the excess cash in and over what period of time,” says Reinke.
The treasury team reviews the ideas to make sure they’re sound. “Then another agent will go and execute that transaction,” she says.
The finance team has remained around the same size despite producing more, she adds.
Anthropic
Annabel Custer , who works in marketing operations at Anthropic, used to manually create event pages for campaigns and handle data imports herself. One data import could take 15 minutes to an hour. She now uses Anthropic’s Claude AI to automate both processes, freeing up time to focus on strategic work.
One Claude agent has been serving as a “builder” and another serves as an auditor, summarizing the work the first agent did. Custer reviews it all.
The biggest challenge when she got started using Claude for multistep tasks was knowing the right questions to ask and how to best guide it. Claude sometimes skipped steps or pushed back on her suggestions. At one point, it asked her if a task was a good use of its time.
Custer learned to be patient, sometimes reworking her instructions.
Once workers let go of their old approaches and ride with the AI, “it unlocks a lot of difference,” she says.
Agents, agents everywhere
Agents are starting to reach beyond the realm of tech. The average Fortune 500 company will run more than 150,000 AI agents in the next two years, according to estimates from market-research firm Gartner.
Yet implementing agentic workflows leads to other challenges, such as a sudden surge in productivity. Gartner found only 13% of companies believe they have adequate AI-agent governance in place.
Agents can trigger what is known as a “10X problem,” says Partha Ranganathan , an engineering fellow with Google. If you increase one workflow by 10X, something else in the system might get affected.
In the case of Google’s invoice-validation agent, the company says the system is working so well that there is a backlog of flagged discrepancies that the operations team needs to take to suppliers. The solution? Google is building another agent that will initiate the process of contacting the supplier.
Jeremy Korst , a partner with the AI advisory-consulting firm Mindspan Labs, says that smaller companies have been the most nimble with agents. Things get more complicated when managing across teams, he adds. The legal team might not actually want the sales team having AI do its own contract review.
“There is real friction,” says Korst. “This is a very common conversation.”
