AI personal agents: A tale of 4 Uber rides and 190K OpenTable diners
In September 2026, OpenTable reported 190,000 seats reserved through its native OpenAI integration. Uber reported four agent-originated rides during Halloween week. One sounds like a new channel taking off. The other sounds like almost nothing.
Both can be true.
The apparent contradiction comes from what they measure: different transactions, periods and experiences. OpenTable counts diners served; Uber counts fulfilled rides. More important, confirming a reservation in chat is different from delegating a task to a third-party agent. For CX leaders, the question is what comes next. What happens when customers ask their AI Personal Assistant (PA) to check an order, resolve a double charge, or cancel a service on their behalf?
A native AI integration exploded
On Booking Holdings’ third-quarter earnings call, CFO Ewout Steenbergen called OpenTable’s native OpenAI integration one of the brand’s fastest-growing channels and reported those 190,000 seated covers. The ChatGPT reservation flow includes a person asking for a table, reviewing options, and confirming the reservation details. Depending on the provider, completion happens in ChatGPT or at the provider.
Booking separately reported that AI referrals accounted for significantly less than 1% of Booking.com room nights; that is a different business and denominator.
Third-party agents barely made a blip
The other half of the story came from three earnings calls within about a week.
On November 4, Uber CFO Balaji Krishnamurthy reported around 320 million combined Uber and Uber Eats trips and orders during Halloween week. About 500 Eats orders came from AI agents. The rides business fulfilled four agent-originated trips.
Four rides. Including Eats, that’s roughly one agent-originated transaction for every 635,000 combined trips and orders!
On October 29, DoorDash CEO Tony Xu said third-party agents had generated several hundred orders across the time they had been operating. For scale, DoorDash separately reported 993 million total orders in the third quarter. The timelines don’t line up, but this is another signal of personal agents accounting for significantly less than 1% of traffic.
Then on November 5, Airbnb CEO Brian Chesky described bookings from Muse and Instinct as very non-material, against nearly half a million total bookings a day. He said traffic from those agents had peaked in early October and been roughly stable over the following four weeks.
The reported agent activity barely registered against the scale of those platforms.
Chat commerce could influence PA adoption
Put the two halves side by side and the distinction becomes clearer: Human-confirmed chat commerce has demonstrated use, while personal agent (PA) traffic remains a tiny share of overall traffic.
These numbers don’t yet demonstrate mass-market adoption of PAs. But I believe comfort with chat commerce will carry forward into personal agent use. Asking an assistant to resolve a double charge seems like a plausible next step for someone already comfortable asking it to arrange dinner.
The timing, scale and boundaries remain open. Customers may happily delegate status checks while wanting to approve a refund settlement or a flight change themselves.
CX leaders don’t need certainty about adoption to prepare. They need to decide what a representative may request, how authority is established and which decisions their own systems can safely make.
Four Uber rides is not a crisis. It is also not a reason to leave those questions unanswered.
One inbound request makes the policy question concrete
At Crescendo, we’ve researched PA usage across our customer base using the out-of-the-box Crescendo Insights dashboard to identify PA conversations with our Concierge AI or human agents. Within our installed base, Personal Agent activity is too low to be statistically relevant. Specifically we found less than 1 in 100,000 customer service conversations are Personal Agents.
Despite the low volume, the conversations are striking. We listened to our Concierge AI handle a customer account balance dispute politely and respectfully. Our Predictive CSAT gave its handling high marks. The PA was clear about what it needed, and our Concierge AI was clear in its answers.
We’ve heard some panic from the market, including customers asking, “How do we stop this?” But looking at the actual interactions, it’s not clear that PAs are necessarily a problem to stop.
When PAs reach live customer service reps, the calls can start awkwardly. The reps aren’t sure how to approach talking to an AI. We’ve seen them fall back on their training, speak as they would to a customer, and handle the conversation well.
An example was one inbound chat in which the sender identified itself as an AI assistant acting for a separate account holder and requested cancellation and a refund. Customer support explained the refund-request process to the PA and the exchange ended amicably.
Prepare for personal agent activity to scale
Don’t panic, but I believe you should prepare for PA activity to scale, potentially very rapidly.
I recommend revisiting your authentication approach and confirming when it’s acceptable to engage with PAs. For many ecommerce businesses, existing checks such as confirming an order number may be sufficient. In industries such as financial services, regulatory constraints or financial exposure may warrant declining or handing off PA requests to people. I believe this should be the exception, but the topic deserves a deliberate policy decision.
Most early-customer wins from PAs involve service operations with high friction. Keep focusing on reducing that friction. AI analysis can help: Crescendo Insights, for example, can rank opportunities to improve the customer journey. If returns or cancellations are intentionally difficult, expect more PA interactions and potentially higher CX operating costs. Watch PA volume and revisit those journeys as it grows.
The lesson in 190,000 diners and four rides? Distinguish today’s adoption from tomorrow’s possibilities. Use the current low volume to prepare.

