Who Finds Out What Happened After Your Customer’s Agent Bought? Coherence at Velocity. October 10, 2026
Who finds out what happened after your customer's agent completed the purchase?
In my forthcoming book, Coherence at Velocity, I write that once customers hand evaluation and purchasing to agents, you have to state your terms, timing, and entitlements clearly enough for an outside system to read them and act. Four writers and analysts published on Thursday and Friday about what comes after that: the agent completes the purchase, and somebody records whether it worked for the customer. Find out whether that person works at your company.
Sangeet Paul Choudary argues in HBR that companies gain an advantage with agents when they observe the outcome of a transaction and learn from it. Katie Matthews reports in Marketing Week that Tesco Media now sells brands audience recommendations built on Clubcard purchase records. MarketingProfs editors compiled a week of news in which Meta, Walmart, and Stripe proposed a standard for agents, and Kantar researchers found that marketers plan to spend on AI assistants well ahead of consumer trust in them. Fareen Mehrzai at Gartner reports that 25 percent of procurement leaders have redesigned a role because of AI.
On Thursday I wrote about the person who chooses the objective your marketing AI scores against. You can only correct an objective when you see results, so today I cover who collects the results.
Find out who records the result of an agent's purchase
Choudary published Who Owns the Customer in the Age of AI Agents? on October 9. He opens with a dispute from last month. Meta launched Muse, a personal agent that fills out forms, negotiates, books travel, and makes purchases for its user, on September 8. Twelve days later, Amazon blocked Muse from shopping on its site. An Amazon spokesperson told GeekWire that third-party applications "should operate openly and respect service provider decisions about whether or not to participate."
Choudary argues that agents now handle discovery, comparison, negotiation, and purchasing, all of which companies used to control on their own sites. He lists five sources of marketplace power that may stop protecting a company's hold on its customer: engagement, network effects, supplier lock-in, proprietary data, and the company's own agent. He assigns the advantage to companies that help agents make better decisions by connecting customer intent to a transaction, observing the outcome, and using what they learn to improve the next decision.
Apply that to your own business as an inventory of three records. The first is the request: what the customer asked for and what constraints they gave. The second is the transaction. The third is the outcome: whether the customer kept the product, returned it, reordered it, or called your service team about it. When a customer buys through an agent, the agent's operator holds the request and you hold the transaction. In most companies I have worked with, operations and service hold pieces of the outcome and nobody in marketing has asked for them.
It reminds me of what Aniket Deosthali, CEO and co-founder of Envive, said in our interview on The Agile Brand podcast: "Let platforms help customers find you, but make sure you retain the ability to serve, understand and build a lasting relationship with them."
In the AI capability framework, I list memory and context as a condition every AI capability requires. An agent operator who holds outcome records has the context to make a better recommendation next time. You need the same records to improve the product, the listing, and the offer.
My take: Choudary defines the advantage in agentic commerce as holding outcome records and learning from them. Inventory yours before your next planning cycle.
Study how Tesco sells the record you lack
Matthews reported on October 9 that Tesco Media is adding influencer video to its retail media offer across the Tesco website, the app, and in-store screens. Tesco has more than 3,800 screens, 2,800 of them in Express stores. The company announced the plan at its Upfront event on October 8.
I am using the second half of her report. Tesco has partnered with the software company Kevel on a self-serve platform called Ad Manager, where a brand plans, books, measures, and optimizes a campaign in one place. A tool inside it called AI Audiences uses Clubcard data to recommend audiences, from a brand's existing buyers to predicted ones. Steve Edwards, head of agency at Tesco Media, described the tool to Matthews: "It explains why each audience fits, and shows combined reach before a penny is committed."
Tesco holds the transaction record and a good share of the outcome record for every brand on its shelves, because a Clubcard member who buys your product again next month does it at a Tesco till. Tesco will now sell you a view of your own buyers, which is a good business for Tesco. Before you buy that view, ask what Tesco reports back after the campaign and whether you can test the result yourself.
It reminds me of what Rajeev Nair, co-founder and Chief Product Officer at Lifesight, said in our interview on The Agile Brand podcast: "First-party data solves targeting, not causality. A retailer that knows a customer buys running shoes every six months can build a far more relevant audience without a third-party cookie, and that is a real, durable advantage. But knowing who to target isn't the same as knowing whether the ad worked."
My take: Matthews documents a retailer packaging its purchase records as an AI planning product. Brands that sell through retailers should treat outcome reporting as a term of the media buy and negotiate it alongside price.
Decide what you will share before someone else writes the standard
MarketingProfs editors published their AI Update for October 9, a compilation of 31 items from other outlets. Four of them describe companies negotiating over the three records.
Meta, Walmart, and Stripe backed a "personal agent protocol," an open standard led by Bret Taylor of Sierra that lets a business authenticate an agent and monitor what it does, according to CNBC reporting in the roundup. OpenAI and Anthropic have not joined, and Amazon has restricted agents. PYMNTS reported that retailers are deciding how much inventory, pricing, and customer data to expose to third-party shopping agents.
Digiday reported that seven months after OpenAI launched ads in ChatGPT, advertisers are holding budgets at test levels because of attribution gaps, count discrepancies, and delayed reporting. Some advertisers have declined to install conversion tracking or upload customer lists because of OpenAI's data-sharing terms. Clients of one agency spend $10 million a month on Google and under $100,000 on ChatGPT. Advertisers who trusted the reporting would spend more than 1 percent of their Google budget.
Kantar researchers supplied the demand figures in Media Reactions 2026, a survey of more than 800 senior marketers and 23,000 consumers. A net 75 percent of marketers plan to increase investment in AI assistants in 2027. Thirty-two percent of consumers use AI assistants to research brands and products, and 23 percent trust the recommendations. Gonca Bubani, global director of media at Kantar, said: "There's a lot of money moving towards AI on the assumption people will use it the way the industry expects."
Last month I wrote that you should instrument the handoff to an agent. Add the terms of the exchange to that work. Write down what you will publish to agents, which covers product attributes, price, inventory, and policies, and see the wiki entry on product feed optimization for AI for the mechanics. Then write down what you require in return: a flag that identifies the order as agent-originated, the customer's stated request, and a way to tie the order to what happened afterward. In the capability framework this belongs to identity and permissions, and the companies drafting the protocol are deciding those permissions now.
Commentary: The editors compiled this issue from secondary reporting with help from ChatGPT, which they disclose, so check the CNBC and Digiday originals before you quote a figure in a board deck.
Assign the outcome record to a named role
Mehrzai, a Senior Director Analyst in Gartner's supply chain practice, published survey results on October 8. Among 213 senior procurement leaders, 63 percent expect AI to significantly improve procurement performance over the next three years, and 25 percent have redesigned jobs or roles because of an AI implementation. She said: "Successful AI implementation at the functional level is fundamentally a role redesign challenge."
Gartner surveyed procurement leaders. I include the finding for two reasons. If you sell to businesses, these are your buyers, and they expect their own AI tools to improve how they evaluate you. And marketing has the same role problem with outcome records. Operations logs returns by order number, service logs complaints by ticket, and finance sees reorders. No job description in most marketing departments includes joining those three to the source of the order.
Christopher Stanton, Justin Shriber, and Simon Rudat make a related point about sales teams in HBR on October 8. They write that productivity gains depend on whether reps and managers convert the time AI saves into better customer understanding, deal strategy, and judgment. Outcome records are where a marketing team gets customer understanding once an agent handles the conversation.
My take: Mehrzai measured a 38-point distance between leaders who expect AI gains and leaders who have changed a job to get them. Marketing leaders can start by adding outcome records to one job description.
Price the outcome gap in a business you recognize
This example is hypothetical. A $400 million cookware brand takes 1.5 million orders a year on its own site at an average of $140. Shopping agents now place 8 percent of those orders, or 120,000 a year. Each agent order arrives through the checkout interface with a name, an address, and a product code. The ecommerce team cannot tell an agent order from any other.
Operations logs every return by order number with a reason code. The site-wide return rate is 9 percent, and nobody has had cause to look closer.
A marketing operations analyst adds an origin flag to incoming orders and joins one quarter of returns to it. Customers return 14 percent of agent-placed orders. The top reason code on those returns is "not compatible with induction cooktop."
The analyst traces it in an afternoon. The product feed marks an entire cookware line as induction compatible. Two of the seven pans in the line have an aluminum base and do not work on induction. Each product page carries the correct callout, so people shopping the site choose the right pan, and an agent matching a customer's request for induction cookware reads the feed attribute and buys one of the two.
Five extra points of returns on 120,000 orders comes to 6,000 returns a year. At $140 each the team refunds $840,000, and at $28 per return for shipping and restocking it spends another $168,000. The total is about $1.0 million a year from one attribute entered at the wrong level.
The team corrects the attribute for the two pans, keeps the origin flag on every order, and gives the analyst a standing weekly join of returns, complaints, and reorders by origin, reviewed with merchandising. The team still lacks the customer's original request, so the CMO adds it to the agenda for the next conversation with each agent operator.
Take these four steps before year-end planning
Inventory the three records for one product line. For the request, the transaction, and the outcome, write down which company holds each one and which team inside your company can reach it.
Flag the origin of every order. Separate agent-placed orders, retail media orders, and human sessions, then join returns, complaints, and reorders to that flag every week.
Write your position on data exchange. List what you publish to agents and retailers and what you require back, and bring that list to your next media or platform negotiation.
Name one owner for outcome records. Give that person the authority to request data from operations and service, and put the weekly join in the job description.
I took piano lessons for 12 years. I practiced alone most days, and once a week a teacher listened to the result and told me what I had been getting wrong. I improved because someone heard the outcome every week and reported it back to me.
In Coherence at Velocity, the last of the five governance gates is the post-release review, where a team measures whether a change delivered the intended outcome and carries the lesson into the next cycle. Run that gate on the orders agents place with you. Ask who records what happened after your customer's agent completed the purchase. If the answer is another company, put that record on the agenda for your next contract discussion with them.
Featured this cycle:
Who Owns the Customer in the Age of AI Agents? Sangeet Paul Choudary, Harvard Business Review, October 9, 2026
Tesco brings influencer content into store as it 'simplifies' retail media experience. Katie Matthews, Marketing Week, October 9, 2026
AI Update, October 09, 2026: AI News and Views From the Past Week. MarketingProfs, October 9, 2026
Gartner Survey Shows Chief Procurement Officers Stalling on Needed AI Role Redesign. Fareen Mehrzai, Gartner, October 8, 2026
Also referenced: What Top Performers Do Differently with AI, and Why They See the Biggest Benefits, Christopher Stanton, Justin Shriber, and Simon Rudat, Harvard Business Review, October 8, 2026; Net 75% of marketers plan to increase investment in AI visibility, Kantar, October 2, 2026; and Amazon blocks Meta's Muse AI assistant in new standoff over agentic shopping, GeekWire, September 20, 2026.