Summary
Uber said that only four rides were ordered through an AI agent during its busiest week in October, while AI agents placed about 500 Uber Eats orders. The figures suggest that agentic consumption and delegated payments remain at an experimental stage despite intense industry interest.
A small number with a larger message
The market for personal AI agents is being built around a straightforward promise: software should do more than answer questions. It should be able to act on a user’s behalf, from finding a product to arranging a ride or placing a food order. That promise has attracted significant attention from technology companies, startups, and investors.
Uber’s latest disclosure offers a useful reality check. On the company’s third-quarter earnings call, Uber said that only four rides were ordered through an AI agent during its busiest week in October, the week that included Halloween. During that same week, Uber and Uber Eats processed roughly 320 million trips and orders combined. Uber’s food-delivery business saw more activity from agents, with about 500 Uber Eats orders placed through one, but that number was still small compared with the platform’s overall volume.
Chief Executive Dara Khosrowshahi described the current level of agent-initiated activity as “very, very early” and said the company was seeing “very limited traction,” while expressing hope that adoption would increase. The comments do not show that AI agents are incapable of booking rides or ordering food. They show that these capabilities have not yet translated into a meaningful everyday habit at a major consumer platform.
The difference between capability and authorization
An AI agent that can navigate an application is not necessarily an agent that users are prepared to trust with a transaction. The first task is largely about automation: locating an option, entering information, or moving through a workflow. The second requires the agent to interpret intent, make choices, and potentially commit the user’s money.
A ride request can involve a pickup location, a destination, vehicle type, timing, price, and changing availability. A food order may include substitutions, delivery instructions, promotions, fees, and other details. If an agent misunderstands even one of those inputs, the result can be a late arrival, an unsuitable order, or an unwanted charge. Users therefore need confidence not only that an agent can complete a task, but that it knows when to ask for clarification instead of acting automatically.
Payment raises the bar further. An agent that can recommend a restaurant or compare transportation options may not need access to a payment credential. An agent that can confirm an order does. That distinction introduces requirements around authentication, permission scope, confirmation thresholds, spending limits, reversibility, and dispute resolution.
For a consumer, the central question is whether using an agent is actually easier than opening the Uber or Uber Eats application directly. If the user must review every field, approve every step, and resolve occasional errors, the agent may add another layer rather than remove friction. If the agent is given broad authority, the convenience may come with an unacceptable risk of unintended transactions.
Platform integration is more than a user interface problem
Agentic commerce also depends on how platforms expose their services. An agent may try to interact with a website or mobile application as a human user would, but such workflows can be disrupted by authentication steps, dynamic interfaces, anti-automation controls, or changes to page structure. Formal integrations can be more reliable, yet they require platforms to define how an agent identifies itself, how a user’s authorization is represented, and which actions are allowed.
Transportation and delivery services add operational complexity. They do not simply process a payment; they coordinate real-time supply, location data, merchants, drivers, customer support, and fulfillment. A platform might welcome agents that help users discover or compare services while remaining cautious about allowing external software to finalize every order. The benefits of additional demand must be weighed against integration costs, fraud exposure, support volume, and disputes over orders placed by an agent.
This creates a two-sided adoption problem. Users need dependable agents, while platforms need dependable agent traffic. A provider may hesitate to open a broad set of transaction capabilities until it can distinguish legitimate delegated activity from abuse, automate appropriate safeguards, and establish responsibility when an agent acts incorrectly.
Why delegated payments need stronger controls
The same issues matter to wallet providers, custodians, and other payment infrastructure operators. When software acts for an individual or an institution, permissions should not be treated as a single all-or-nothing switch. Viewing information, creating an order, confirming a payment, changing account settings, and transferring assets are materially different actions.
A mature agentic payment system would likely need granular authorization, explicit transaction context, limits, step-up verification for sensitive actions, and records that show which user or organization granted permission. It would also need a way to revoke access and investigate activity after the fact. These requirements are relevant whether the underlying payment is made through a conventional account, a wallet, or another financial system.
Uber’s disclosure does not describe the payment architecture behind its AI-agent orders, so it does not establish which controls are currently in place. It does, however, illustrate the practical challenge: the existence of a technical pathway does not mean that users will routinely delegate real spending decisions to an automated system.
Early figures should be read carefully
The figures from Uber represent one company’s disclosure for one particularly busy week. They cannot, by themselves, describe the entire agentic-commerce market or establish how every AI product performs. Low usage could reflect limited awareness, restricted availability, an inconvenient entry point, or an authorization flow that users have not yet adopted.
Even with those caveats, the figures are informative because Uber and Uber Eats are services that naturally fit the agentic-commerce narrative. Ordering a ride or a meal appears straightforward from the outside, making these categories obvious candidates for automation. If agent-driven activity remains very small in those settings, the industry still has to validate several steps between a successful product demonstration and sustained consumer behavior.
Important measures will include repeat usage, cancellation and refund rates, the frequency of human intervention, user acceptance of automatic confirmation, and whether agent-generated activity creates enough incremental value to offset integration and risk-management costs. A single successful transaction is not evidence of a durable business model. Repeated, predictable use is a more meaningful test.
The next test is trust, not just model intelligence
OpenAI, Anthropic, Google, and other companies are now offering task-oriented AI agents, while newer personal-AI startups are attracting attention and capital. As these products reach more users, agentic consumption may grow. But adoption will depend on more than whether the underlying model can navigate a service or interpret a natural-language request.
The more difficult questions are operational and institutional. How much authority should an agent receive by default? Which actions require a second confirmation? How should an agent explain a price or a change in terms? Who is responsible when a request is ambiguous? How can a user distinguish an authorized agent from a malicious or compromised one? How can a merchant verify that the transaction was approved without creating unnecessary friction?
For payment companies, wallet operators, custodians, and merchants, the relevant shift is from asking whether an agent can act to defining the conditions under which it may act. Permission tiers, transaction limits, identity verification, anomaly monitoring, and clear dispute procedures are likely to become as important as the model’s ability to complete a workflow.
Uber’s comments therefore offer a measured conclusion about the current state of agentic commerce. Industry enthusiasm is high, and the technology is advancing, but real-world transaction volume remains limited in at least one large consumer ecosystem. The next phase will require more than impressive demonstrations. It will require users to trust agents with bounded authority and platforms to make that authority visible, auditable, and reversible.
Source: link