Some of the traffic your security team blocked last night belonged to your customers.
In section 9.4 of Coherence at Velocity I describe what changes when customers delegate evaluation and action to agents. The enterprise stops communicating only with people, and systems acting on their behalf start to interpret it. Offers have to stay consistent across channels, eligibility has to be explicit, and terms, timing, and entitlements have to be machine-readable enough to survive comparison. A fragmented enterprise gives those systems a reason to drop the brand before a person ever sees the creative. Authors at McKinsey and Forrester published new evidence on October 27 and 28 that the delegation has started, that it touches both consumer and business purchases, and that most companies still treat the resulting traffic as a security problem.
On Tuesday I argued that every AI program needs a capacity ledger before go-live. This edition turns outward, to the agents that arrive at your storefront and your sales process on a customer's behalf, and to the decisions a marketing leader has to own before those agents decide whether the brand belongs in the shortlist.
Treat the agent on your site as a customer with a different interface
Elizabeth Silliman, Jack Trotter, Kelsey Robinson, Desirae Oppong, and Shreya Jaggi of McKinsey report that 53 percent of consumers across 11 markets used AI for shopping in the past three months, and that the share of US AI users who completed a purchase through an AI tool rose from 10 percent in December 2025 to 17 percent in June 2026. Emily Pfeiffer and Sandy Carielli of Forrester write that retailers have to stop treating all automated traffic as hostile, because some of it belongs to customers' shopping agents and answer engines, and that 80 percent of online adults in the US, UK, and Canada who have used AI to shop use it to compare products. Barry Vasudevan of Forrester predicts that agentic buying will play a role in 75 percent of B2B purchases in 2027 and that buyers will double their human interactions to validate what their agents recommend. Peter Kim of Forrester reports that 78 percent of US B2C marketing executives say their leadership now requires stronger evidence of value for AI spend than it did 12 months ago.
Read together, these authors describe a customer who shows up twice. The agent arrives first to read structured data, compare, and shortlist, and the person arrives second to confirm what the agent found. Most marketing organizations have built everything for the second visit and left the first one to the bot-management vendor.
Measure where your brand lands in the AI consideration set
Silliman, Trotter, Robinson, Oppong, and Jaggi published Who wins when AI does the shopping? New evidence on agentic commerce on October 28, drawing on a June 2026 survey of about 21,000 consumers in 11 countries. Fifty-three percent used AI for shopping in the past three months, and the US sits at 48 percent, with Gen Z and millennials in the low 60s. Purchases follow discovery with a lag: 7 to 18 percent of AI users have completed a purchase through AI in a given category, and 21 to 30 percent in the top categories say they would. The authors found that 56 percent of AI shoppers used a stand-alone chat or search tool and 27 percent used a retailer's own assistant, and that stand-alone tools lead at every stage except purchase, where retailer assistants edge ahead at 15 percent to 14 percent. Nearly a third of AI users named AI as the single biggest influence on their last purchase. The authors give CMOs six priorities: win the AI consideration set, measure continuously, make product data AI-ready, protect customer and loyalty data while opening product data, invest in an owned-site assistant, and put a senior leader in charge across functions.
Two of those priorities carry most of the weight. In Marketing Operations 3.0 I define reach in an agent-mediated market as inclusion in the model's consideration set, and I propose Share of Model as the operating metric: citation frequency for category prompts, response accuracy against your current pricing and specifications, and sentiment alignment. McKinsey's "measure continuously" and "win the consideration set" describe the same program. A team that tracks Share of Model monthly will see the product-data problem before the sales team sees the lost deal.
It reminds me of what Dan Russotto, General Manager at furniture.com, said when I interviewed him on The Agile Brand podcast: "We get basic product feeds from each of our retailers, but that's all it is. It's basic."
Russotto runs an aggregator that lives on other companies' product data, and the feeds he receives carry the minimum a human merchandiser needed. McKinsey's "AI-ready product data" means attributes, availability, eligibility, and policy terms structured well enough for an agent to compare them without guessing.
Commentary: Pick your five highest-margin categories and run the prompts a shopper would ask ChatGPT, Gemini, and Alexa for each one. Record whether your brand appears, whether the price and specifications match your site, and who on your team owns the fix.
Write a policy for each kind of bot and give it an owner
Pfeiffer, a principal analyst, and Carielli, a VP and principal analyst, published Don't Block All The Bots. They Might Be Shopping For Your Customers. on October 28. They argue that bot and agent trust management has moved from a CISO topic to a commerce decision. They describe three categories of automated visitor a retailer may want to serve: marketplace agents such as Amazon's Buy for Me, answer engines such as ChatGPT and Gemini, and consumer-owned shopping agents such as Instinct and Meta's Muse. Of the 22 reference customers Forrester interviewed for its bot and agent trust management buyer's guide, 17 used the software for account fraud and 13 had started applying it to AI agent trust. The authors recommend separate policies for malicious automation, known partners, answer engines, commerce agents, and customer-authorized agents, each with a named business owner, and they tell retailers to test proofs of concept against legitimate agent scenarios so the vendor's engine does not block the customers it was meant to protect.
The named business owner is the part most retailers will skip. I have sat in enough bot-management reviews to know that the default owner is whoever manages the CDN contract, and that person has no reason to care about a shopping agent's conversion rate.
In my AI Capability Framework, identity and permissions form one of the foundation layers, and that layer has to work in both directions. The enterprise authenticates its own agents and it authenticates the agents that arrive from outside. Alyson Clarke, David Hoffman, and Chang Liu of Forrester published a companion piece on October 27, Banks Can't Block Their Way Into The Agentic Future, after five banks blocked third-party agents outright. They recommend graduated trust: authenticate both the customer and the agent, assign each agent a unique identity with least-privilege permissions, and phase access in by risk, with research and monitoring as proving grounds and payments gated behind explicit customer approval. The same sequence applies to a retailer, a software vendor, or an insurer.
It reminds me of what Robin Ross, Digital, Analytics and Loyalty Executive at Activate Insight, said when I interviewed her on The Agile Brand podcast: "We've worked for so long to protect our site from bots. And so bots are bad, but all of a sudden there's this good bot."
Take a hypothetical outdoor-gear retailer with 1 million monthly site sessions and a $120 average order. Its bot-management platform classifies about 5 percent of sessions, roughly 50,000 a month, as automated and blocks them under a rule written in 2023 for scrapers. Nobody has reviewed the rule since.
In the first version, the retailer leaves the rule in place. The 50,000 blocked sessions include an unknown share of answer engines and shopping agents, and the retailer's products appear in those agents' comparisons only through third-party data, with last season's prices. The marketing team sees no cost, because the lost orders never enter any report.
In the second version, the ecommerce director takes ownership of agent policy. The team writes five policies on Forrester's lines, lets answer engines and known commerce agents read structured product data, and requires authentication before any agent adds to cart. The product content team spends $150,000 structuring attributes, availability, and return terms for the top 2,000 SKUs, the security team spends $90,000 tuning and testing the bot platform against legitimate agent scenarios, and the director's time comes to about $60,000, for a total of $300,000. Agent-referred sessions now convert at 2 percent, which produces about 1,000 orders a month, or $1.44 million a year. Half of those customers would have arrived another way, so the team counts $720,000 as incremental. Return on Agent comes to ($720,000 minus $300,000) divided by $300,000, or about 140 percent, and the second-version retailer also knows, for the first time, how many of its visitors are machines.
Commentary: Ask your security lead for last month's blocked-traffic report and have someone in ecommerce classify a sample by Forrester's five categories. If the answer is "we don't know which bots we blocked," that is your first policy.
Prepare your sales team for buyers who verify what their agents found
Vasudevan, a VP and principal analyst, published Predictions 2027: AI Takes Over Business Buying, But Humans Retain Control on October 28. He starts from Forrester's finding that 82 percent of business buyers said their most recent purchase included some AI functionality, whether they wanted it or not. He predicts that agentic buying will play a role in 75 percent of B2B purchases, for bounded tasks such as spend analysis, contract-renewal analysis, supplier discovery, and RFP creation, while fully autonomous purchasing stays confined to high-volume, low-risk categories. He also predicts that buyers will double their human interactions to validate what their agents recommend, seeking out product experts, references, and analysts, and that risk reviews will delay or block one-third of enterprise AI software purchases. More than half of AI purchases already involve four or more departments, against 26 percent of non-AI purchases, and buying groups have grown from 8 stakeholders to 13.
Vasudevan tells providers to make product, pricing, security, privacy, compliance, and performance information accurate, structured, and accessible to buyer agents, and to equip subject-matter experts to explain where their AI delivers value and where it stops. Put that beside McKinsey's consumer findings and the pattern holds in both markets: the agent shortlists, the person verifies. In Marketing Operations 3.0 I describe the Business to Agent to Consumer model, in which a customer agent performs the first round of comparison before a human reviews the shortlist. In the B2B version, the human review doubles in intensity, because the buyer wants to trust the agent's work before signing a contract with 12 other stakeholders.
Marketing owns the first visit and sales owns the second, and the two have to agree on the same facts. A pricing page that an agent reads one way and a rep explains another way costs the deal at the verification step.
Commentary: Audit your pricing, security, and compliance pages for structure an agent can parse, then hand the same pages to your sales enablement lead. Whatever the agent finds there is what the rep will have to defend.
Fund agent readiness as growth, with the evidence your board now demands
Kim, a VP and principal analyst, published Marketers, Stop Random Acts Of AI Investing on October 27. He lists the reasons marketers funded AI over the past few years: executives reading newsletter headlines, hearsay from peers, conference case studies with no lift or incrementality data, open-ended promises about what agents can do, the credibility value of having an AI case study to present, and self-preservation. A member of Forrester's Board of Clients told him that executives insisted everyone use AI without a plan or a stated reason. Kim's one statistic comes from Forrester's Q3 2026 CMO Pulse Survey: 78 percent of US B2C marketing executives say leadership requires stronger evidence of value than it did a year ago. He expects blank checks to disappear in the 2027 budget cycle and the emphasis to shift toward growth and revenue.
Kim's list reads like the intake form for the capacity ledger I described on Tuesday, and his conclusion points the same direction as the BCG research in HBR on Monday: the companies earning a return pick a growth path and fund the capability it needs. Agent readiness qualifies. The outdoor-gear retailer's $300,000 buys structured product data, tested agent policies, and an owner, and the evidence Kim's executives want comes out of the same work: Share of Model by category, agent-referred conversion, and incremental orders.
Agile Brand Principle 1, the customer at the center, now has to account for a customer who sends a proxy. The proxy reads the operating facts and ignores the headline.
Commentary: Before you submit a 2027 AI line item, write its growth path and its evidence plan on the same page. If the evidence plan reads "usage," send it back.
Take these actions before the holiday traffic peaks
Classify last month's blocked traffic. Sample it against Forrester's five categories: malicious, known partner, answer engine, commerce agent, customer-authorized agent.
Name an owner for agent policy. An ecommerce or digital leader, with the CISO as a partner.
Run a Share of Model check. Five categories, three assistants, monthly, with accuracy against your live prices.
Structure the operating facts. Attributes, availability, eligibility, return terms, and for B2B, pricing, security, and compliance pages.
Gate actions, open reading. Let recognized agents read; require authentication before they add to cart or submit an RFP.
Brief sales on the first visit. Reps should know what the buyer's agent already found before the call.
Make the brand legible before a machine reads it
Every one of the 25-plus books I have written went to the publisher with a metadata sheet before the manuscript: categories, keywords, comparable titles, a 200-character description. Nobody buys a book from the metadata sheet, and a book without one never appears in the retailer's search results, so the manuscript never gets its chance.
Chapter 6 of Coherence at Velocity makes the case that brand has become behavior: decision policies, eligibility logic, exception handling, and the consistency of an answer across channels. Section 9.5 extends that argument to agents, where operating performance and persuasion become hard to separate, because a system acting for the customer has less reason to recommend an enterprise that reads as slow, inconsistent, or contradictory. McKinsey's authors measured the delegation at 53 percent of consumers, Forrester's authors measured the verification habit at double the human interactions, and both findings rest on the same dependency: the agent works from the data the enterprise structured, and the person's trust depends on what that agent could verify.
Answer the opening observation this week. Find out which of your customers' agents your systems turned away last night, give one person the authority to decide which ones get in, and make sure the product data they read matches the promise your brand makes.
Featured this cycle:
Who wins when AI does the shopping? New evidence on agentic commerce. Elizabeth Silliman, Jack Trotter, Kelsey Robinson, Desirae Oppong, and Shreya Jaggi, McKinsey, October 28, 2026
Don't Block All The Bots. They Might Be Shopping For Your Customers. Emily Pfeiffer and Sandy Carielli, Forrester, October 28, 2026
Predictions 2027: AI Takes Over Business Buying, But Humans Retain Control. Barry Vasudevan, Forrester, October 28, 2026
Marketers, Stop Random Acts Of AI Investing. Peter Kim, Forrester, October 27, 2026
Also cited: Banks Can't Block Their Way Into The Agentic Future. Alyson Clarke, David Hoffman, and Chang Liu, Forrester, October 27, 2026