Agentic Commerce Fails on Product Data, Not on the Model

Agentic commerce fails on product data, not on the model. An agent that shops, compares or recommends a product can only act on the attributes it can read, and those attributes have to be verified and current. When a required field is missing or stale, the safe move is to stop and escalate, not to guess. That is the argument Tarun Chandrasekhar, President and Chief Product Officer at Syndigo, makes in You're Not Deploying an Agent. You're Deploying a Liability., published on 10 November 2026. He cites Gartner's prediction that more than 40% of agentic AI projects will be canceled by the end of 2027, and says the causes Gartner lists (rising costs, unclear business value and weak risk controls) point to a data and accountability problem rather than a model problem.

The distinction he draws is practical. Generative AI drafts, and a person can review the text before anything is published. Agentic systems query other systems, take a decision and trigger work downstream. By the time someone notices an error, the decision may already have been executed. Adding a review step after the fact is not the same as limiting what the agent is allowed to decide. In commerce, Chandrasekhar writes, that means an agent should act only on product attributes that are verified and current, and should escalate when a required field is missing or stale. In Syndigo's own assessments, fewer than a quarter of brands have product data that meets the bar for an agent to recommend their products with confidence, and most of those brands assumed they were fine. He also warns against baking changing facts such as specifications, allergens and assortments into a model's weights, because those values go stale; current facts have to be retrieved at the moment the agent acts.

Two jobs that share the same catalog

Noemy Miyamoto's Pimberly guide to agentic PIM, published on 12 October and updated on 3 November 2026, separates two jobs that often share the same word. Agentic PIM is work inside the catalog: agents notice gaps, propose or apply a fix within permissions, and leave an audit trail. Agentic commerce is work on the other side of that catalog: agents search, compare and, in some cases, buy on a shopper's or business buyer's behalf, reading whatever product data you expose. MCP is not an agent in either case. It is a way for an agent to connect to a system's data and actions, whether that system is a PIM or a storefront.

The through-line is the record itself. Shopping agents only work as well as the data they can read, and agentic PIM is one way to keep that data in shape before it leaves the catalog. Our articles on DwyerOmega's production use of Akeneo MCP and on AI guardrails and decision lineage cover the permissions, review and evidence on the catalog side. This piece looks at what happens when an outside agent has to trust the published result.

Confidence is not the same as readiness

That gap between assumption and evidence is now a published benchmark, not only a vendor observation. inriver's 2026 Product Data Maturity Index, presented in the PIMTalk on the Product Data Paradox, surveyed 405 industrial manufacturers and wholesale distributors in the United States and Europe. Of the 117 organizations that said they were fully ready for agentic commerce, none met the operational bar when tested. The index scores operations on evidence such as launch automation, SKU publish-readiness, integration failures and content measurement, rather than on self-reported confidence. Just 13.6% had end-to-end automation with an audit trail, and only about a third had 85% or more of active SKUs publish-ready at a given moment.

The finding matches Chandrasekhar's point that agents do not fill gaps the way a person reading a product page will. An incomplete allergen field, an inconsistent attribute or a specification written for a human visitor is not a small content issue once an agent reasons over it at scale. Our guide to product data quality and governance sets out the ownership and completeness work this depends on, and Trends in PIM, Part 2 on AI governance covers the rights behind any agent that is then allowed to act.

The work sits in the product data flow

The ShoppingTomorrow PIM and AI expert group reaches the same conclusion from the operating chain rather than from a readiness score. In Agentic commerce starts with product data. Not with AI., published on 8 October 2026, Arnout Schutte of ConnectingTheDots, which hosts the group with Truvio and with Squadra as chair, argues that an agent can only work with product information it can find, understand and trust. If information is missing, contradictory or impossible to verify, the product becomes harder to compare and may drop out of the recommendation. The bluepaper's main conclusion is to start with the data and processes an agent has to trust: product information that is complete, accurate, current, consistent and traceable, from supplier onboarding through enrichment and publication to webshops, marketplaces and AI channels.

The cases in that bluepaper show the work as a flow, not as a model choice. Azerty, one of the organizations described, processes around 20 million records a day from about 100 data sources. Product information for more than 400,000 products is handled centrally and distributed to webshops, ERP, a data lake and marketplaces, and a new product can be live with the correct specifications within ten minutes. Adding another assistant is not the goal in that account. Making the product data flow is. Schutte's line is the practical test for a shopping agent as well as for an internal one: if an AI agent cannot see it, it does not exist.

What a shopping agent actually reads

For most shoppers, that visibility still means research rather than a completed purchase. Pimberly's Black Friday 2026 guide to agentic commerce, published on 6 November 2026, states the current pattern plainly: most AI tools are not buying yet. They read product pages and feeds, then tell shoppers what they found. A BCG survey cited there found that only 4% of consumers would let an agent finish a purchase on its own. What has to be right, today, is the data those tools read. Price, stock, identifiers, shipping and returns have to match across the product page, the feed and the structured markup. A sale price on the page that disagrees with the feed, a GTIN that cannot be verified, or a stock status that has not caught up with the warehouse is exactly the kind of ambiguity an agent will not repair with judgment.

The protocols now forming around checkout do not change that input. In WISEPIM's merchant guide to the Universal Commerce Protocol, UCP is Google's open standard, announced in January 2026, for catalog, cart, checkout, identity and orders. ACP, from OpenAI and Stripe, covers checkout inside ChatGPT. AP2 handles payment authorization, and MCP can carry catalog or checkout capabilities as a transport. UCP and ACP compete at checkout; MCP and AP2 do not. Both checkout protocols still need a catalog a machine can read without guessing: stable product and variant IDs, plain-text descriptions, barcodes with a type, prices, availability and explicit options such as size and color. Our guide to AI-ready product data for shopping agents covers the identifiers, feeds and structured values this depends on.

The first agent is not the hard part

The same record then has to stand up when more than one agent works on it. Casey Paxton's Akeneo article Your First AI Agent Is the Easy Part: What Happens After the Pilot?, published on 26 October 2026, uses Snowleader as the successful first case. The outdoor retailer manages around 450,000 SKUs across more than 400 brands, with as many as 17,000 new products in a seasonal launch. It generates descriptions from structured product data, with prompts set by product family and brand tone and with the content team reviewing the result. Snowleader reports that this cut the time spent writing descriptions by around 20%. That is a real gain, and it is still a single-task pilot.

The next questions arrive quickly: an agent that onboards supplier files, one that fills missing attributes, a quality agent that validates a result before a content agent writes, a localization agent that prepares approved copy for each market. Once several agents act on the same product record, they need one governed source, a clear sequence, and a way to see which source was used, what was recommended, who approved it and what changed. Without that, one agent can update a specification while another writes copy from an older value, or a translation can go out before the upstream change is approved. Chandrasekhar's two questions still apply at that point: what "correct" looks like for the data the agent will reason over, and who is accountable when it is not. The organizations still running agents in production, he writes, will be the ones that put verified data and clear accountability in place before they let an agent act.