Uber ChatGPT Rollout: 3 Hidden Work Costs of AI [2026]

Uber is introducing ChatGPT Work to more than 30,000 employees, according to the company's own announcement. The rollout follows a trial involving 1,000 people and was announced on October 22, 2026. Just days earlier, an internal study by Anthropic, based on interviews with 17 employees, raised an important question for business leaders: could AI create more work than it removes?

This question connects several business technology stories from the week. Meta added administrator tools to Forum, Google brought Gemini into more advertising products, and OpenAI shared a restaurant owner's story about reducing a three-hour task to a few minutes. These changes come as new AI models reach customers within weeks of one another. Our comparison of GPT-6.1 Sol, Sonnet 5.5, and Gemini 4 Argon examines this fast-changing market.

There is an important limit to the available evidence. It does not show that AI is increasing workloads across businesses in general. It covers one company's internal study, one very large rollout, and several announcements from technology companies. This article separates confirmed facts from company claims and our own analysis.

The Week's Business AI News at a Glance

Five developments stand out this week. Each raises the same question in a different way: who does the work after an AI agent makes its first attempt? The table lists what the available sources confirm.

Three of these five stories come from companies describing their own products. Only the Anthropic item studies how people actually spend their time, and Anthropic also sells AI agents. Readers should consider both facts when judging the evidence.

Uber's ChatGPT Work Rollout: From 1,000 Pilot Users to 30,000

Uber has expanded its trial from 1,000 people to more than 30,000 employees. That is a 30-fold increase, announced in a single rollout. The reported plan includes connections to company tools, an internal library of reusable skills, and employee training. These details may matter more than the number of employees involved. They suggest Uber is treating AI adoption as a major workplace change, not simply as a software purchase.

A small trial often answers a simple question: do the first users like the tool? A companywide rollout must answer a harder question: can the whole organization use it successfully? Uber has not published productivity figures from its trial in the information available to us. The public case for expanding the rollout therefore rests on the company's stated goals, not on published results.

The size of the rollout still matters. When a company as large as Uber gives tens of thousands of employees an AI assistant, other large employers may start asking why they are still running small trials. Readers can follow TechCrunch's AI coverage for updates. Uber's move also gives OpenAI a named customer using its workplace product on a large scale.

What ChatGPT Work Does Inside a Company

Research for this report describes ChatGPT Work as an AI agent designed to handle tasks with several steps. These include research, meeting briefs, presentations, data analysis, and work involving several applications. At Uber, it is reported to connect with workplace tools such as Google Drive, Gmail, Google Calendar, and Slack. The main difference between a chatbot and an agent is what each can do. A chatbot answers a question. An agent works toward a goal, gathers information from different systems, and produces a finished result.

This ability also creates a real cost: people must check the results. Reading a chatbot's reply may take only seconds. Checking a ten-slide presentation made from information in three email inboxes and two file storage locations takes much longer. The person checking it also needs enough background knowledge to spot mistakes. The agent shifts some of the work from creating content to checking it.

Skills libraries and training

The internal skills library is a small but important detail in Uber's announcement. A shared library suggests the company expects employees to create repeatable workflows that other people can use. This could reduce repeated work. However, it also creates a new responsibility: someone must manage and test these workflows, and update or remove them when tools or company rules change.

What Uber and OpenAI Leaders Said

Uber's chief technology officer, Praveen Neppalli Naga, said the rollout was about giving employees the tools to use and shape the technology. In Uber's announcement, he said: “As AI transforms how we work, we want to give everyone the tools to innovate, solve problems and help shape Uber’s future.” He added: “Rolling out ChatGPT Work across the company is an exciting step toward that ambition, and I can’t wait to see what our teams build with it.”

OpenAI president and co-founder Greg Brockman, quoted in the same announcement, described the partnership this way: “Now teams across Uber can tap into our portfolio of frontier models and agents to take on more complex work, connect the tools they use every day, and get more done.”

Both statements promise new abilities, not proven savings. Naga talks about what employees will create, while Brockman talks about handling more complex work. Neither says the rollout will reduce the number of hours employees work. This is a reasonable reading of their words, and it fits with the Anthropic findings discussed below.

Anthropic's Study: Human Attention Becomes the Bottleneck

Anthropic's internal study used employment records, data about tool use and work output, surveys, and interviews with 17 employees. According to the available findings, employees took on more work coordinating and monitoring AI agents as they handled more of the agents' output. Fortune reported this statement from Anthropic: “This acceleration makes human attention the bottleneck for agent output.”

This statement captures the main concern. If AI agents produce work faster, the main limit may become the people who must read, approve, correct, and bring together their results. The amount of work produced may increase, but people still have only so many hours in a day.

There are two important limits to these findings. First, 17 interviews make up a small group, and the study was carried out inside a company that builds AI tools. Second, the findings we can verify do not show that the same pattern occurs at other employers. Treat the study as a credible warning sign, not proof that AI is increasing workloads across the whole economy.

Why Agents Can Add Coordination Work

Anthropic's separate research into systems that use several AI agents gives an example of how demand can exceed available capacity. The research states: “In one run there were 2.4 million job requests and only 117 jobs accepted.” This example concerns AI agents competing for work, not office employees. However, it shows how quickly the number of requests can grow beyond the number that can be handled.

The same basic problem can happen in a workplace. When each employee can start several AI agents, the number of drafts, summaries, and suggested actions can multiply. Someone must sort through them and decide what needs attention. Three things explain how AI can create more work even when each task becomes faster.

Review work grows with output

When AI creates drafts faster, it can produce more drafts than people can easily review. If a manager used to check four documents a day but now receives twelve, the review workload triples unless the manager spends less time checking each one. Our earlier report on AI cutting coding time by 52% while flooding open source with slop described a similar problem in software. Faster production can leave maintainers with more work to check and fix.

Handoffs between agents and people

When an agent works with Gmail, Slack, and Drive, someone may need to follow what it did if the result looks wrong. Assigning tasks, checking progress, and fixing failures are all coordination jobs. These activities may not appear in productivity forecasts, but they still take up real working hours.

Meta Gives Forum Administrators Admin AI and Admin Assist

Meta introduced new administrator tools for Forum, including Admin AI, Admin Assist, and a shared calendar for moderators. The calendar may sound less exciting than the AI features, but it reveals something important. Moderation requires people to take turns doing the work, and a shared schedule helps organize those people rather than automate the task. Meta is combining AI assistance with a feature designed to coordinate the people who remain involved.

Community administrators are a useful group to consider when studying changes in workload. Many do this work as volunteers or in part-time professional roles, so their time is limited. Meanwhile, the number of posts they must handle may not decrease. An assistant that answers routine questions and flags items for review could reduce tedious work. But it could also create a new list of flagged items that someone must check. Meta has not published measured results in the information available to us, so the overall effect remains unknown.

Google Pushes Gemini Into Ad Buying

Google announced business tools that bring Gemini into Display & Video 360 and Search Ads 360. It also announced a connection between LinkedIn's connected-TV advertising inventory and Google's advertising system. This gives advertisers another way to reach professional audiences through TV-style video ads while using the advertising tools they already work with.

Advertising operations are an area where extra review work can become expensive. Generative AI tools can plan campaigns, improve their settings, and make changes faster than a person can read the record of those changes. That speed makes checking the work more important. An advertising buyer is still responsible for the money spent, even if they cannot explain why an automated system changed a bid. The sources available to us do not provide figures showing how much time Google's tools save or how often they make mistakes.

The Restaurant Test: Ted Taylor's Cash Reconciliation

OpenAI's customer story about Ted Taylor provides the clearest example this week of work being reduced. Taylor cut the time needed to match monthly cash records from three hours to between 5 and 10 minutes across three restaurants. This process compares sales records, payment information, and cash deposits. It is a task with clear rules and an answer that can be checked.

The story also mentions a possible annual cash-flow benefit of $50,000 to $100,000. However, this is a possible future benefit, not a confirmed saving. It should not be treated as a result that has already been achieved. The claim that the task now takes only 5 to 10 minutes also comes from a company-published story about one restaurant operator. It shows what may be possible, not what usually happens.

The contrast between the restaurant example and the Uber and Anthropic findings is important. Cash records can be checked against a clear standard: either the totals match or they do not. More open-ended work, such as presentations and research briefs, has no equally simple test. People must judge whether the result is good enough. AI reduces work most easily when checking the result takes little effort.

The Model Layer Behind the Rollouts

These workplace products rely on AI models that are changing quickly. Release trackers list many new models since late September. The dates below come from those trackers and our earlier model coverage. Where only a month is available, the table says so.

Google also released Gemini Nano Banana 2.1 for creating images and editing them through conversation. For companies choosing AI tools, the main issue is how quickly the technology changes. A company that chooses one AI agent today is relying on a product and a development plan that may change from month to month. Our report on Gemini 4 Argon's 1M-token context and paid-only access covers one part of this changing market.

Competitive Comparison: Three Vendor Pitches for Workplace AI

OpenAI, Anthropic, and Google all want their technology to become the main AI agent layer used by businesses, but their announcements this week show different approaches. OpenAI is showing that its product can be used on a large scale through a named customer. Anthropic is publishing research about possible problems caused by AI agents. Google is adding Gemini to products businesses already pay for. The table gives our view of the questions each approach leaves unanswered.

Anthropic's own workplace agent, covered in our article on Claude Duet and its 6,000-app pitch, competes in the same market. That makes the study unusual. An AI vendor is publishing evidence that challenges the simplest version of the sales story for its own type of product.

Historical Context: New Tools, New Chores

Past productivity tools show how technology can move work around instead of removing it completely. Email made sending messages easier, but it also increased the number of messages people had to read. Spreadsheets made building models cheaper, so managers began asking for more models. Chat apps made replies faster, but many teams ended up spending more time working and responding. These examples illustrate a possible pattern, not proven measurements. The sources for this report do not provide figures for any of them.

AI agents may create a different kind of extra work. Earlier tools often increased the amount of reading and writing. Agents can increase the amount of judging people must do because someone has to decide whether the agent's work is correct. The effects on jobs are also uncertain. Goldman Sachs has highlighted concerns about AI's effect on hiring and a possible shortfall of 10,000 jobs. This shows how the same technology may affect both the amount of work available and employers' demand for workers.

Market Impact for Enterprise Buyers

For companies buying AI tools, Anthropic's findings change how they should judge the business case. If a tool saves ten minutes on a task but adds twenty minutes of checking, it does not save time overall. Companies will increasingly need to ask vendors for figures showing the total time saved after checking and correcting the work, rather than relying only on how quickly the AI completes a task.

Cost is another concern. McKinsey's research on the value of AI, discussed in our report on a $2.52 trillion prize that shrinks by $850 billion if costs rise 30%, shows how higher costs can reduce the value of AI. Extra hours spent checking AI work are an ongoing cost that may not appear in a vendor's price quote.

Managing AI use is a third challenge. Workers already use AI tools that their employers have not approved, a trend we covered in our report on 45% of workers using shadow AI as firms push tools. Companywide rollouts such as Uber's may help bring this activity into systems the company can manage. At the other end of the spectrum, five large US banks block third-party AI agents, according to our report on banks blocking outside agents. This shows that willingness to adopt these tools varies sharply by industry. More information is available in our AI and machine learning coverage.

What Is Confirmed and What Is Not

• Confirmed by the announcement: Uber announced on October 22, 2026, that it would roll out ChatGPT Work to more than 30,000 employees after a trial involving 1,000 people.
• Reported: The Uber rollout includes connections to company tools, an internal skills library, and employee training.
• Company study, limited sample: Anthropic found increased coordination and monitoring work, based on 17 interviews along with records and survey data.
• Vendor claim: Ted Taylor reduced the time needed for monthly cash reconciliation from three hours to 5 to 10 minutes.
• Not confirmed: The potential annual cash-flow benefit of $50,000 to $100,000 has not been confirmed as an actual saving.
• Not established: The available evidence does not prove that AI is creating more work across businesses in general, or provide a complete list of every story behind this week's headline.

The BBC's AI business hub and similar news outlets may add independent evidence in the coming weeks. Until then, the safest conclusion is that AI may increase workloads in some situations, but this possibility is still at an early stage and has not been measured across businesses on a large scale.

Five Predictions for the Next Two Quarters

These are our forecasts, not confirmed facts. Any of them could prove wrong.

1. Companies will measure total time saved instead of just task speed. Buyers will ask vendors to report how much time is saved after employees check the work. Better sales presentations will include these figures.
2. More companies will expand pilots by around 30 times. Uber's increase from 1,000 to 30,000 users may become a model for other large employers, which could announce similar expansions.
3. Tools for supervising AI agents will become a product category. Companies will offer dashboards for tracking agent tasks, sending work for approval, and keeping records of actions, either as built-in features or as add-ons.
4. Advertising buyers will demand records of changes. As Gemini automates more work in Display & Video 360 and Search Ads 360, advertising agencies will want clear explanations of changes made automatically.
5. Another AI vendor will publish a study of workload effects. Anthropic's decision to discuss these problems may encourage a competitor to publish its own data about the work needed to check AI output.

A Rollout Checklist That Accounts for Oversight

Companies planning their own AI agent rollouts can use these lessons before more studies become available. Before expanding a trial, measure the time employees spend checking AI results as well as the time the AI spends producing them. Limit how many agents each person can run at once until the team knows how much checking it can handle. Give every shared skill or workflow a clear owner. Finally, require a person to approve any result that leaves the company. Our guide to an AI agent sandbox that isolates MCP servers explains the technical side of limiting what an agent can access.

FAQ

Is AI creating more work for employees?

It can in some situations. Anthropic's internal study found that employees did more coordination and monitoring as they handled more AI agent output. The study included 17 interviews and other types of data. The available results do not prove that the same effect occurs in every business.

How many Uber employees will get ChatGPT Work?

More than 30,000 employees, according to Uber's announcement on October 22, 2026. The rollout followed a trial involving 1,000 people.

What can ChatGPT Work do?

It is designed for tasks involving several steps, such as research, meeting briefs, presentations, and data analysis. It can work across applications. At Uber, it is reported to connect to tools such as Google Drive, Gmail, Google Calendar, and Slack.

What did Meta announce for Forum?

Meta introduced administrator tools called Admin AI and Admin Assist, along with a shared calendar for moderators. The available sources do not provide measured results showing how much work these tools save.

Is the $50,000 to $100,000 restaurant saving real?

It is not a confirmed saving. The figure describes a possible annual cash-flow benefit. The reported result is that Ted Taylor reduced the time spent matching monthly cash records from three hours to 5 to 10 minutes across three restaurants.

Which new AI models matter for workplace agents right now?

Recent releases include GPT-6.1 Sol, dated September 29; Claude Opus 5.5, dated September 22; Claude Sonnet 5.5, dated September 28; Claude Haiku 5.5, dated October 7; and Gemini 4 Argon, which became generally available on October 13.

What should companies measure before expanding an AI agent pilot?

They should measure the total time saved after checking and correcting the results, not just how quickly the AI completes a task. They should track how long employees spend reviewing output, assigning work, and fixing failures. These are the extra costs highlighted by Anthropic's study.