Guides

Tasks to delegate to AI agents: lessons from Uber's 4 rides

AI agents booked four Uber rides in a week of 320 million trips. Here's what that says about which tasks are worth handing to an agent, and which to keep.

S

SelfAgentic Team

7 min read
A hand writing a to-do checklist with a pen in an open grid-paper notebook.

Uber says AI agents booked four rides during its busiest week of October. Across Uber and Uber Eats, agents placed about 500 of the roughly 320 million trips and orders that week.

That number is a useful guide to which tasks to delegate to AI agents. Booking a ride takes a person under a minute, so an agent has almost nothing to save. Agents pay off on work that is slow, repetitive and spread across several tools, the kind of work a small team keeps meaning to hire for. The rule of thumb: hand off the hour-long chore you do every week, and keep the 50-second tap.

What Uber's numbers say about everyday AI agent use

The figures came out of third-quarter earnings calls, and Fortune pulled them together. Uber CEO Dara Khosrowshahi said the median customer takes 50 seconds to order a ride and 90 seconds to order food. DoorDash CEO Tony Xu put agent orders at "maybe several hundred orders in total" since launch. Airbnb's Brian Chesky called agent bookings "a very non-material number" against nearly half a million bookings a day.

None of the three is giving up on agents. Khosrowshahi said "every single AI system company is talking to us." The early pattern is still clear: when a task is short and the app already works well, people do it themselves. Chesky gave a second reason. He said about 80% of travelers enjoy researching their trips, and people rarely want to delegate the part of a job they like.

Which tasks to delegate to AI agents: five tests

Office work looks different from ordering dinner. Anthropic's January 2026 Economic Index report found that more complex tasks produced bigger time savings, and about three-quarters of the business (API) interactions it studied were classified as automation. The same report found success rates fall as tasks get longer. For API calls they dropped from around 60% on tasks under an hour to roughly 45% on tasks that would take a person five hours or more.

So the useful range sits between the 50-second tap and the week-long project. A task is a good candidate for an AI agent when most of these are true:

  1. It takes you somewhere between 20 minutes and a couple of hours each time. Below that, checking the agent's work costs more than doing it. Above that, split it into smaller steps.
  2. It repeats on a schedule: daily, weekly, or every time a new lead comes in.
  3. It pulls from more than one place, such as the web, a spreadsheet, Slack and meeting notes.
  4. You can check the result faster than you could produce it. A one-page brief with sources is quick to review. A pricing decision is not.
  5. A mistake can be caught before it reaches anyone outside the team.

The first four tell you whether delegating saves time. The fifth tells you whether it's safe to try. Our time range in test 1 is a judgment call, so adjust it to how fast you review.

A worked example: a three-person team's weekly chores

Say a three-person B2B startup lists what it does in a normal week and runs each item through the tests. The times below are made up for the example.

Task

Time per run

Delegate?

Monday competitor brief from websites, pricing pages and news

About 90 minutes

Yes

Daily scan for prospects that match the ideal customer

About 45 minutes

Yes, with outreach drafts held for review

Turning sales call notes into follow-up tasks

About 20 minutes per call

Yes

Booking the meeting room

Under a minute

No

Setting next quarter's pricing

Days, mostly judgment

No, but an agent can gather the inputs

The competitor brief is the clearest win. A research agent reads the pages overnight, a writer agent turns the changes into a one-page summary, and a reviewer agent checks each claim against its source. On Monday the founder spends five minutes reading the brief instead of ninety building it. Over a 13-week quarter that is about 18 hours back.

A single chatbot can write the summary if you paste in the pages, but someone still has to collect them, remember what changed since last week, and post the result. In SelfAgentic that job can run as a scheduled workflow: one agent researches, another drafts, a third checks the claims, and the agents remember what they found in earlier runs, so the brief can say what is new.

The prospect scan passes with a condition. Finding and scoring companies is low risk. Sending the first email is not, because a bad message lands in a real person's inbox. So the agent drafts and a human sends.

Pricing fails test 4, because you can't check the answer faster than you could reason it out yourself. An agent can still collect competitor prices and past deal sizes into one sheet, which takes the dull part off your plate.

What should you keep for yourself?

Quick tasks in apps that already work well are the Uber lesson. If the job takes a minute, an agent can't save you anything worth the setup.

Choices that are the point of the work stay with you too. Chesky's travelers enjoy picking the hotel, and you probably care which investor gets the first call. Delegate the research behind those choices and keep the choice.

Actions you can't undo, such as payments, signed contracts, public posts and ad budget changes, can still be prepared by an agent, but a person should approve the last step. Anthropic's Claude Duet app is built this way: it reads, drafts and prepares work on its own, then shows what it is about to do and waits for approval before it sends a message, makes a payment, signs a document or deletes anything. We covered how to decide what needs sign-off in our guide to human-in-the-loop AI agents.

SelfAgentic uses the same default. Publishing to LinkedIn or Instagram and changing Meta ad campaigns wait in an Approvals inbox until you decide, and each agent can only use the tools you have granted it.

FAQ about delegating tasks to AI agents

What tasks should you not delegate to an AI agent?

Skip tasks that take a minute or two, choices you want to own, and anything irreversible that touches money, contracts or the public without an approval step. An agent can still prepare the inputs for those decisions.

Are AI agents worth it for small teams?

They earn their keep on recurring, multi-step work that eats hours each week, such as research briefs, lead lists and meeting follow-ups. For quick one-off tasks, the time saved is too small to cover setting up and checking the agent.

How long should a task be before you hand it to an AI agent?

There's no fixed cutoff. Anthropic's data shows bigger time savings on more complex tasks but lower success rates once tasks run past a few hours, so a task of an hour or two, split into clear steps, is a sensible place to start.

Do AI agents need human approval?

For anything that leaves the team, such as an email to a prospect, a public post or an ad budget change, yes. Research, drafting and internal summaries can usually run without sign-off, as long as someone reads the output.

How to start delegating tasks to AI agents this week

  1. Write down every recurring task from the last two weeks with a rough time for each.
  2. Cross off anything under ten minutes and anything you can't check quickly.
  3. Pick the remaining task with the most hours per month and the lowest cost of a mistake.
  4. Run the agent alongside your own process for two or three cycles and compare the results.
  5. Once its output matches yours, let it run on schedule and switch from producing the work to reviewing it.

Expect the first runs to miss things. Agents fail more often on long, loosely defined tasks, so when the output is weak, narrow the task and spell out the inputs before you rewrite the instructions. Start with one weekly chore, and add a second only after the first one runs without you.

If you want to try this with agents that run on a schedule, remember earlier runs and ask before anything goes out, see how SelfAgentic works.

Keep reading

Ready to build your AI team?

See how a team of agents handles real work, and checks with you before anything goes out.

Free to start. No credit card required.