
Balaji Krishnamurthy gave the most useful number of the week on Uber's earnings call, and the number was four.
Uber and Uber Eats handled "around 320 million trips" in Halloween week, he said on the call. From AI agents, Uber Eats saw "around 500 orders total." And rides? "It was a grand total of four trips that we fulfilled from agents." DoorDash had said much the same a week earlier. As Mia Osmonbekov reports at Fortune, Tony Xu put agent orders at "maybe several hundred orders in total across the time that they have been up and running."
Six days after Uber's call, OpenAI opened Businesses in ChatGPT in beta. As Luis Rijo reports at PPC Land, a company can now claim its profile in ChatGPT, add a product feed with prices and stock levels, publish a Business Agent that answers questions in its name, and see how often ChatGPT mentions it. One early dashboard Rijo cites, posted by Conrad O'Connell for a business called BuildUp Bookings, showed 690 mentions and 24 clicks over 28 days.
Different companies and different windows, so that's no ratio. But the direction is hard to miss. Assistants aren't buying much yet. They're already talking. And from this week the talking has an editable profile, a feed, and an agent with your name on it.
Somebody in your company is going to own those. Decide who this quarter, before it's decided by whoever happened to have the login.
This is part 9 of The CPO Transition, part of The Product OS. Part 8 gave the CPO one outcome number. This part is about the first place outside agents touch that number, which is the story.
The short version
- Assistants tell your story long before they transact.
- Uber counted four agent-booked trips in a week of around 320 million, and DoorDash counted a few hundred agent orders in total, while OpenAI's new Businesses in ChatGPT beta gives a company a claimable profile, a product feed, a Business Agent, and a count of mentions and clicks.
- So the agent-facing surface that matters in 2027 is the description, and it's made of product artifacts: a feed generated from the catalog, an agent with instructions and a test set, and a profile that has to match what shipped yesterday.
- The CPO's decision is to make all of them outputs of one positioning file inside the product pipeline, written by product marketing, released and tested like any other agent, and measured from mention to a verified job.
- Hand them to four teams and the assistants get four versions of the company. The terms each assistant gets, what it can read, what it can do, and what comes back, are a product decision too, and the CPO makes it with the CRO.
Agents describe before they buy
Dara Khosrowshahi explained the four himself. Booking a ride in the app is fast: "It's actually 50 seconds." An agent can't save you much on a 50-second job. So Uber is "building our own AI experiences," he said baskets built with its cart assistant run roughly twice the average, and he's keeping the customer: "Those direct consumer relationships are central to our approach to AI."
That logic holds for a ride. Software isn't bought in 50 seconds. A buyer compares, asks what it costs, and asks whether it does the one thing they need, and more of those questions go to an assistant first. The Product Marketing Transition has been tracking how far that has moved, so I'll leave the data there.
The sequence I expect: describe first, transact later. Transaction counts stay small for a while. Description counts are already large. Look at that dashboard again. ChatGPT brought the business up 690 times and people clicked through 24 times. Most of the time, what the assistant said was all the person saw.
4
Kevin Indig called Business Hub "a sort of Search Console for ChatGPT," and Rijo is careful about its limits. No prompts, only the top discovery topics, and OpenAI's own help note says clicks "don't confirm website visits or conversions." It's a thin instrument today. Thin instruments on a channel this size don't stay thin.
Every field in the profile is a product artifact
Read the list of what a company fills in, as a product leader would.
The product feed carries prices and stock levels. For a software company that's plans, limits, and packaging. It changes when you ship.
The Business Agent is something you build, test, and publish, and for now it's limited to eligible US ecommerce and online service businesses. That's an agent with instructions, a knowledge source, and a test set. Your company already has a pipeline for agents like that, with prompts in the repo and evals at the gate. It lives in product.
The profile is the positioning, compressed to a description and a category. And the insights tab is the first outside read on whether the story landed.
Nothing on that list is a campaign. It's a release target, one more surface every release has to update, the way docs and the changelog are. Part 4 of The Product Marketing Transition argued that messaging should be one versioned context file every customer-facing agent loads, with an eval suite behind it, and I won't restate it. The CPO's piece is smaller and harder. The profile, the feed, and the outside agent are three more readers of that file, and they update in the same motion as the product. The gap between shipping a change and the assistants describing it correctly is the story lag that part 3 of that series measures. This is where it gets closed or doesn't.
One story needs one owner
Here's how it goes when nobody decides. Marketing claims the profile, because it looks like a listing. Operations uploads the feed. Support builds the Business Agent, because it looks like a chatbot. Product owns the agent inside the app. Four teams, four descriptions of one product, each updated on its own calendar, and the buyer hears whichever one is stale.
This is the one-function case in its plainest form. The loop runs customer, spec, build, story, adoption, and the story is a stage of it. When a release changes what the product does, the story has to change with it, and only the org that ships the release can make that happen the same day. So product marketing writes the file and owns the claim, inside the product org, and the release pipeline carries it to every surface that reads it.
The CMO keeps plenty. Brand, demand, and the paid side all stay there, including the option Rijo describes to sponsor a Business Agent as the destination for ChatGPT ads. What moves to the CPO is narrower: that what the assistants say about the product is true as of today.
The terms of trade are a product decision
Khosrowshahi said one more thing that deserves a CPO's attention. "Every single AI system company is talking to us," and "we're really now negotiating the terms of trade." He's counted eight assistants so far and expects more than ten.
Terms of trade sounds like a business development phrase. Unpack it. What can the assistant read: the feed, prices, availability? What can it do: quote, book, change an order? What comes back: the customer's identity, the conversation topics, attribution? And what does it cost? The first three are an API and a permission model. All four have a revenue consequence.
So the CPO and the CRO write it together, once, as a policy, the same way part 7 argued for the knowledge that comes back from the field. Uber's answer is already clear: keep the direct relationship, make its own app the best path, and work with the third parties anyway. That's a product strategy, said out loud by a CEO. With ten assistants on the way, a company that answers deal by deal ends up with ten different products. A later part takes the line between the CPO, the CRO, and the CMO in full.
What I expect
These are my predictions, not facts.
My prediction: within 12 months, every major assistant offers a claimable business profile with a feed and an analytics tab, and "update the assistant listings" becomes a line on the release checklist.
My prediction: by the end of 2027, at most AI-native companies, the company's agent on outside assistants ships through the same eval and release pipeline as the in-product agent, and it's owned in product.
My prediction: within two years, public companies report assistant mentions and agent-completed orders as a channel, the way Uber and DoorDash just did by accident, and analysts start asking for both.
My prediction: by 2028, a written terms-of-trade policy for assistants is as standard as an API terms page, and at the companies that do it well the CPO signed it.
One thing to try this week
Open three assistants. Ask each one what your product does, what it costs, and who it's for. Write down every answer that was true three months ago and isn't now. Then, for each wrong answer, find out who in your company could fix it, and count the teams. If the count is more than one, you've found the decision. Bring the list to your next staff meeting and name the owner.
Sources: Uber, Q3 2026 earnings call transcript, Dara Khosrowshahi and Balaji Krishnamurthy, Uber investor relations, November 4, 2026. Luis Rijo, "Businesses gain ChatGPT click and mention counts in OpenAI's Business Hub," PPC Land, November 11, 2026, with Kevin Indig and Conrad O'Connell, product details as he reports them. Mia Osmonbekov, "AI companies want agents to handle your everyday errands, but Uber, DoorDash, and Airbnb say barely anyone is using them for that," Fortune, November 10, 2026, with Tony Xu from DoorDash's October 29 call.
Part of the running argument on Product Leadership: the product org chart was calibrated to a build cost that collapsed, and in the AI era the whole product outcome, from customer signal to adoption, belongs to one function under the CPO.
Related answer: Who owns the ChatGPT business profile, the CPO or the CMO?
Frequently asked
How many Uber rides did AI agents book?+
On Uber's Q3 2026 earnings call on November 4, 2026, CFO Balaji Krishnamurthy said Uber and Uber Eats handled around 320 million trips in Halloween week, that AI agents placed around 500 Uber Eats orders, and that Uber fulfilled a grand total of four trips from agents. CEO Dara Khosrowshahi said booking a ride in the app takes 50 seconds and called agent traction very early.
What is Businesses in ChatGPT?+
A beta OpenAI opened on November 10, 2026. As Luis Rijo reports at PPC Land, a business can claim its profile in ChatGPT, add a product feed with prices and stock levels, build, test, and publish a Business Agent, and see insights in Business Hub: mentions, website clicks, and discovery topics. OpenAI doesn't share user prompts, and Business Agents are limited to eligible US ecommerce and online service businesses for now.
Why do assistants matter if agents barely transact yet?+
Because they describe a company long before they buy from it. Uber counted four agent trips in a week, while one early Business Hub dashboard, posted by Conrad O'Connell, showed 690 ChatGPT mentions and 24 clicks in 28 days. Most of the time, what the assistant said was all the person saw. The description is the agent-facing surface that's live today.
Who should own the company's profile, feed, and agent on outside assistants?+
Falk's answer is the CPO's org, with product marketing writing the claim inside it. The feed is the catalog and pricing, the Business Agent is an agent with instructions and a test set, and the profile is the positioning in two sentences. All of them go stale when the product ships, so they should be outputs of one positioning file in the product pipeline. The CMO keeps brand, demand, and paid placement.
What are the terms of trade with AI assistants?+
Dara Khosrowshahi's phrase for what Uber is negotiating with every AI system company. In practice it covers what an assistant can read, what it can do, what comes back to the company, and what it costs. Falk argues each of those is a product decision with a revenue consequence, so the CPO and the CRO set it once as a policy.
What does Falk predict about assistants and the product org?+
His predictions, labeled as such: within 12 months every major assistant offers a claimable business profile with a feed and analytics; by the end of 2027 the company's agent on outside assistants ships through the same eval and release pipeline as the in-product agent at most AI-native companies; within two years public companies report assistant mentions and agent-completed orders as a channel; and by 2028 a written terms-of-trade policy for assistants is standard.


Sign in with LinkedIn to leave a comment.
Sign in with LinkedIn