Uber Cut 10% of Its Support Staff in July and Cited AI. Then It Gave 30,000 Employees an Agent

OpenAI said on Oct. 22 that ChatGPT Work, the agent it launched on July 9, would go to more than 30,000 Uber Technologies employees worldwide, expanding a pilot of 1,000 people that ran earlier in the year. About 80% of the pilot's early users were not engineers, and the work they did was operations work: analyzing operations data for trends, localizing training materials by market and language, building daily dashboards. Two design choices carry most of the information. The default model inside ChatGPT Work is GPT-6.1 Sol, which OpenAI positions as balancing frontier performance against cost efficiency, and developers keep using Codex separately — so the broad rollout deliberately sits on the cheaper end of the model line while the expensive, engineering-grade usage stays where it is. Nobody has published what the tokens cost

Here is what makes Uber Technologies a case study rather than a press release. In the first half of 2026, revenue rose 13.3% to $27.4 billion while cost of revenue — the line that carries courier and carrier payments, incentives and payment processing — rose 3.6%. Corporate overhead moved the other way: general and administrative expense rose 30.7%, sales and marketing 25.4%, research and development 20.5% and operations and support 14.9%. The variable cost of serving a trip is falling while the cost of running the company is rising

The headcount linkage is not theoretical and it is not new. On July 22 Uber Technologies cut 10% of the jobs in its customer service organization, telling the team it was working 'to simplify operations, strengthen in-person collaboration, and continue to embrace AI' — the first time the company has tied a layoff to an AI efficiency program, and the second round of cuts in under two months. The October rollout is that same bet extended from one function to every knowledge worker. The measurable output gains sit in two other places: engineering and the marketplace. Agents now account for more than 70% of the company's code-change submissions, engineers run more than 30,000 agent tasks a day, and the chief financial officer said on the second-quarter call that code output per engineer has doubled. The productivity is measurable; the labor saving is still a promise

Uber's own experience is the best available evidence on what a company-wide agent deployment costs, and it is a cautionary tale rather than a case study. Chief technology officer Praveen Neppalli Naga confirmed that the company's 2026 AI budget was exhausted in about four months, with the overrun driven by runaway use of AI coding assistants including Anthropic's Claude Code. By late August the picture had inverted: weekly agent requests were running 9.4 times their February level while total AI spending stayed roughly flat, the cost per 1,000 requests had fallen about 34% from the April peak and the cost per session was down 52% from the June high. Agents now handle more than 70% of code-change submissions. The October rollout is that lesson applied at scale — the same tools, a cheaper default model, no new headcount, and token caps instead of a budget. The seat fee is the small part of the bill

Each of those 30,000 seats is a metered inference order, and the chain that fills it is visible in public filings. OpenAI has spread capacity across Microsoft's Azure, Amazon's Web Services, Alphabet's Cloud and CoreWeave. CoreWeave discloses the commercial shape of that demand: 98% of revenue comes from committed take-or-pay contracts, and an OpenAI order form signed in September 2025 commits the lab to pay up to about $6.5 billion through May 2031. Second-quarter revenue of $2.6 billion was more than double the $1.2 billion of a year earlier, and concentration is loosening — no single customer crossed 10% of revenue and the three largest together were about 36%, against roughly 71% for the largest customer a year earlier. Enterprise agents are the first customer base that scales with seats rather than with data centres

The other end of the rollout matters more over a few years. Uber Technologies's agents do not sit in a chat window; they read Google Drive, Gmail and Google Calendar plus Slack, then return finished work. That turns the app layer into plumbing and the seat into an optional unit. Salesforce, whose workspace product sits inside the rollout, is already repricing the transition in public: Agentforce annual recurring revenue passed $1.5 billion, up more than 240%, and 3.2 billion agentic work units were delivered in the quarter, up 97% quarter over quarter, while total revenue rose 11% to $11.3 billion. Microsoft faces the same arithmetic on a far larger base, with commercial productivity revenue of $102.0 billion in fiscal 2026, up 16.2%. Agents do not cancel a seat; they change who is worth charging for

The template Uber Technologies is testing — work that moves inside a company and never comes back as a billed hour — is the risk for anyone selling outsourced operations. Genpact is the cleanest listed proxy for that labor: second-quarter revenue of $1,343.4 million, up 7.1%, average first-half headcount of about 144,300, down 0.4% year over year, and management now citing agentic solutions among its demand drivers. No evidence links Genpact to this rollout; the read-through is the model, not a contract. The asymmetry is plain — a platform can hand an agent to 30,000 employees in a day, while the same workflow has to be sold, staffed and priced one client at a time. Work that moves inside an enterprise never comes back as a billable hour

The market is paying for trips and cash, not for agents, and it has been paying less for the trips than it was a year ago. Uber Technologies reported second-quarter gross bookings of $58.0 billion, up 22% on a constant-currency basis, trips of 3.9 billion, adjusted EBITDA of $2.8 billion at 4.9% of gross bookings against 4.5% a year earlier, non-GAAP operating income of $2,143 million, up 40% at 3.7% of gross bookings, and free cash flow of $2,792 million that pushed the trailing twelve months above $10 billion for the first time. Guidance for the third quarter calls for gross bookings of $58.25 billion to $60.25 billion and non-GAAP earnings per share of $0.84 to $0.88, a midpoint that fell just short of consensus and was the reason the shares fell on the print. The market is underwriting volume and cash, and it marked down the shares32% over twelve months while bookings accelerated