In partnership with

Happy Sunday. In Halloween week, Uber handled about 320 million trips and orders, and AI agents were responsible for only 4 rides and around 500 Uber Eats orders. But that's not an AI agent problem. We'll talk about it. Also this week: Airbnb says AI helped cut its support cost per booking by more than 17%, but a closed ticket isn't the same as a happy guest. And a humanoid robot maker thinks it has come up with a way to measure robot efficiency, which sounds great until you notice who wrote the scorecard.

IN TODAY'S ISSUE

Plus, free: Try this week. Check whether one product page gives an AI shopper enough to buy, with two copy-and-paste prompts.

STORY 1 OF 3 · AGENTS

1. AI agents booked just 4 Uber rides and 500 Eats orders. Out of 320 million.

On Uber's third-quarter earnings call on November 4, CFO Balaji Krishnamurthy said the company handled around 320 million trips and orders across Uber and Uber Eats in the week of Halloween, its busiest week in October. Uber didn't split that total between rides and Eats orders. AI agents placed only around 500 of the Uber Eats orders. And four rides. Not four thousand. Four.

Uber didn't dress it up, calling it "very, very early and very limited traction." CEO Dara Khosrowshahi also pointed out that booking a ride in Uber's app takes about 50 seconds on average, and reordering a favorite meal about 90. Hard for an agent to beat an app that already knows the home address and last Friday's order. DoorDash, TechCrunch reports, said on its own call that it has received only a few hundred agent orders so far.

Retailers are building the doors anyway. On October 21, Walmart announced an Agent Gateway that lets agents shop through MCP, a standard way for AI tools to connect to other systems, starting with a beta for companies and developers. It also announced Agent Profiles, "launching in the next month," which give each agent its own profile inside a customer's account, with approvals and a spending limit, plus terms that agents must follow or get blocked. Walmart was refreshingly blunt about why: agents shopping on a website built for people is "slow and error-prone, and agents may get stuck along the way."

My take: four rides isn't proof that nobody wants agents. It's proof the plumbing isn't there yet. An agent shopping today mostly clicks through pages designed for human eyes: product names written for marketing, pack sizes hidden in photos, prices that change at checkout, login walls and CAPTCHAs. That's a product data and checkout problem, not a demand problem.

FROM OUR PARTNERS

The AEO work, without the AEO hire.

Most AI visibility tools show you where you rank, then leave the work to you.

ChatFeatured is an AI search agent that does it: finds the questions competitors win, drafts the content in your voice, and tracks the lift. 

Your team decides what ships. Month to month.

STORY 2 OF 3 · MEASURING VALUE

2. Airbnb's AI cut support costs 17%

Airbnb's Q3 results, out November 5, had a number most companies would love to put in a press release. About 40% of all customer support issues were resolved without a human agent in Q3, up from a third in Q2. That's all issues, Airbnb says, not just the ones that started in its AI assistant. Customer support cost per booking fell more than 17% year over year, "driven in large part by improvements to our AI assistant."

That's one narrow job, with a baseline and a cost you can count. Compare it with Accenture's new CIO Pulse survey of 300 senior technology leaders at companies with more than $1 billion in revenue, published November 3. 64% say they're scaling enterprise AI, but only 7% are optimizing at scale. And 54% name governance and responsible-AI concerns as their top barrier to scaling AI. Accenture's own summary: "Enterprise AI deployment is broad. Value conversion is not."

My take: I like this a lot, and I'd still ask one rude question: resolved according to whom? "Resolved without a human" can mean the guest got what they needed. It can also mean they gave up, closed the chat and took it out on the review. Airbnb's release doesn't say which, and the cost per booking would look the same either way. Cheaper support isn't automatically better support.

This is where that 54% governance number earns its keep. Governance isn't a committee that slows things down. Done right, it's how a business sees overall performance, not one number in isolation that happens to look pretty. Next to any support AI's resolution rate, I'd track three things: how often the same customer comes back about the same issue within a week, satisfaction on AI-resolved cases versus human-resolved ones, and how often the bot held on to a case it should have handed to a person.

Satisfaction is the one I'd push hardest. Send every customer the same short satisfaction (CSAT) survey when the case closes, whether a bot or a person handled it, and compare the two. Then look hard at who didn't answer and who came back a few days later. If AI-resolved CSAT holds up against human-resolved CSAT while costs fall, that's real value. If it slips, guests are quietly paying for the savings, and it'll show up later in reviews and rebookings. If nobody's measuring CSAT at all, the 40% is a guess with a nice chart.

That's what good governance gives a business: cost, resolution and satisfaction read together, so one good number can't hide a bad one. Resolved isn't satisfied. Measure what happened after.

More patients say ”Yes” with Cherry Payment Plans

Had enough of, “Let me think about it?” With Cherry, patients can apply for financing solutions on their phones in about 35 seconds - and 90% of them are approved*.

The result? Practices have seen acceptance rates rising by as much as 50%!

* These figures come from Cherry’s provider case studies, so individual results vary. Details at withcherry.com/provider-reviews. Payment options through Cherry are issued by the following financing partners: withcherry.com/financing-partners

STORY 3 OF 3 · PHYSICAL AI

3. A robot maker wants robots on a timesheet

On October 29, at its Innovation Day in Shanghai, robot maker AGIBOT proposed a way to measure robots by the work they do, not by how many were built or shipped. It calls it Worken, from "work" plus "token," and describes it as "Effective Intelligent Work Hours."

The formula: active robots × autonomous working hours × output efficiency ratio. Active means robots actually doing tasks, not units sitting in inventory. Autonomous hours means time working on its own, not time powered on. The efficiency ratio compares acceptable output with a benchmark for the task, such as the work a human completes under established operating standards.

AGIBOT says its 20,000th robot rolled off the line in September, and The Robot Report notes IDC ranked it first in global humanoid shipments in the first half of 2026. So this is a big shipper telling the industry to stop counting shipments. Admirable, and a little convenient. Worken is a vendor proposal, not a standard. AGIBOT itself calls it "an evolving framework." A robot maker writing the robot scorecard is a bit like a new hire drafting their own performance review.

My take: anyone who has run a plant will recognize this. It's OEE, overall equipment effectiveness, under a new name. OEE multiplies availability by performance by quality, and plants have tracked it on fillers and packaging lines for decades. Worken maps neatly: active robots and autonomous hours are availability; the efficiency ratio rolls performance and quality together. I don't mind the rebrand. I mind the details. "Autonomous hours" is the part of the formula to focus on, because the hours a human technician spends babysitting a robot are the hours that tank the business case. And as I always insist, "acceptable output" needs a clear definition that's owned by the plant, not the vendor.

Here's what I'd leave my manufacturing operators in CPG or other sectors with, if they're considering adding robots or humanoids to their lines. What does OEE look like when part of the workforce isn't human? Do we follow the Worken model, or do we need a different one? I'd start simple: measure the robots the same way the plant already measures the line:

  • Uptime without intervention

  • Units per hour

  • First-pass quality

  • Every intervention, counted

If a vendor can only report robots delivered, you know exactly what you're buying: robots delivered.

TRY THIS WEEK · FREE PRACTICAL TIP

Check if one product page gives an AI shopper enough to buy

Set aside 15 minutes. No agent, no integration, no budget approval. Just an AI chat tool and one public product page.

Step 1: Pick a best-selling product and open its public product page. If you don't sell online, use a brand you buy from. Copy the visible text: title, description, specs, price, pack sizes, shipping and returns. Only use what's already public. For example:

Example input: "Lemon Sparkling Water 500ml. Refreshing citrus taste! Available in multiple pack sizes. Free shipping on orders over $35. See packaging for nutrition details."

Step 2: Start a new chat in any AI tool and paste this first prompt, with the page text at the end:

Read only the product page text below. Fill in a table with: product name, exact size, pack count, price, sugar or ingredients, delivery estimate and return policy. If the page doesn't clearly state something, write NEEDS CHECK. Don't guess and don't use outside knowledge. PRODUCT PAGE: [paste the page text here]

Step 3: In the same chat, send this second prompt:

Now act as a shopping agent. Your customer asked for "a 12-pack of 500ml lemon sparkling water, no added sugar, delivered by Friday, under $20." Using only your table, could you complete this purchase with confidence? If not, list exactly what stopped you.

Step 4: Check the table against the page yourself. Every NEEDS CHECK is a spot where an agent would guess, skip the product or give up. In the example, everything except the name and the 500ml size should come back NEEDS CHECK. If the AI fills in something the page never says, mark that too. It's the same kind of guess a real agent can make. And if it marks NEEDS CHECK on something the page does say, that fact is probably buried in marketing copy where a machine can miss it. Keep the customer's wish out of the first prompt: in testing, a one-prompt version copied "12-pack" and "Friday" from the request straight into the table.

What good looks like: a short list of missing or vague facts for whoever owns the product data, not a perfect score. Fixing them helps human shoppers too. A chat prompt isn't how Walmart's gateway works, but it shows fast how much of a product page is marketing and how little is data.

BUILDERS

Put this week’s ideas to work

Builders is the paid tier with the templates, prompts and workflows behind each issue, including the Agent Guest List for deciding what agents are allowed to do. $9.99/month.

The Insider Room includes Builders and more.

SHARE THE CONTEXT WINDOW

If this issue was useful, pass it on. One colleague joining counts, and referral rewards start from there.

Copy, paste and send to a colleague:

❝

I've been reading The Context Window, a weekly on practical AI for business leaders. You can join here: {{rp_refer_url}}

Your link is also in the box below, with your reward progress.

Hashi Sivananthan

We are out of tokens. ✋
- Hashi