OpenAI's checkout retreat exposed where agentic commerce actually struggles
In March 2026, OpenAI quietly pulled Instant Checkout out of ChatGPT, the feature that had let shoppers buy products without leaving the chat window. Forrester analysts called it the first visible correction to agentic-commerce hype, and the numbers backed them up: of Shopify's millions of merchants, only around 30 had ever gone live with the integration [2].
The retreat wasn't just a partnership fizzling out. Forrester's own March 2026 consumer research found that completing a purchase inside an AI assistant was the least-used feature of these tools, even among people who used them regularly for other tasks [2]. The more common use cases stayed firmly pre-purchase: general questions and product research.
There was an operational problem too: most AI systems pulled product data through crawling or basic feeds, with no visibility into channel-specific stock. Agents offered sold-out items or products never meant for that channel [2]. Forrester's conclusion: agentic commerce doesn't mean the end of the retail website, and handing over checkout is premature [2].
Is the real adoption happening in the back office instead of the storefront?
While shopper-facing AI stalls, retailers have moved further and faster with agents that prepare internal decisions on assortment, pricing and promotion. A meaningful share of retail companies already run concrete agentic-commerce applications, with many more planning to, according to an EHI Retail Institute survey of 218 retail respondents conducted for IT provider Softserve; the authors note the sample isn't representative [1].
Consumers tell a different story about the customer-facing side. An ECC Köln "Social, Assisted, Agentic" study of 1,000 people, surveyed in September 2025, found most could imagine using an AI shopping agent, though far fewer say they would actually hand over a full purchase including payment, citing trust and privacy concerns [2].
Inside category management: what an agent actually prepares
A category-management agent reviews a retailer's own sales, margin and competitor pricing data, then drafts a specific recommendation, a price change, a delisting, a new listing, complete with reasoning, sources and an expected effect, before a human ever sees it. Zurich-based Zenline AI, which says its clients include Otto, Futterhaus and Conrad, builds exactly this kind of agent [1].
“Today, our agents prepare the decisions, and the category teams decide.”
Founder Arber Sejdiji draws a sharp line between two categories of retail AI: shopping agents that search and compare products for consumers, and agents that work inside the retailer's own headquarters on assortment, price and promotion decisions. He considers the second category far more urgent, because the payoff doesn't depend on waiting for consumer behavior to shift [1].
The mechanics are concrete. For a pricing decision, the agent compares a product's weekly sell-through against similar items in its category and against competitor prices every week. If a product underperforms, it proposes an adjustment within the retailer's existing margin and price rules [1].
The margin evidence: a furniture retailer's A/B test
A mid-size European furniture retailer with roughly 9,000 items gives the clearest proof point so far. Before adopting an agent, underperforming new products typically surfaced only after three to four months in weekly reports, by which point they had to be cleared with steep discounts [1].
With the agent running that comparison continuously, outliers now surface within weeks instead of months. Measured as an A/B test, a group of products managed with agent recommendations against a control group managed as before, gross margin rose and slow-moving inventory dropped noticeably [1].
Zenline says agents can identify relevant product trends three to six months earlier than before, which the company says can mean millions in revenue on a single item; pulling a weak launch earlier also frees up working capital that would otherwise sit in discounted stock [1].
| Dimension | Shopping agents (customer-facing) | Category agents (retailer HQ) |
|---|---|---|
| What they decide | Product search, comparison, purchase | Assortment, pricing, promotion |
| Who approves | Consumer, directly | Category manager, per recommendation |
| Consumer acceptance | Few would delegate a full purchase [2] | Not consumer-facing |
| Measured retail adoption | Still largely pilots | Already live at many retailers [1] |
| Near-term ROI | Unclear, checkout pulled back [2] | Margin gains reported in A/B tests [1] |
Why shoppers still won't let AI finish the purchase
Shoppers want AI for research and comparison, not for taking control of the actual decision or payment. A Kantar survey of roughly 1,000 people in Germany, conducted in December 2025, found 49% want AI only for information search, with the remainder open to more directive uses but still wary of handing over the final step [2].
The hesitation has specific causes: respondents in related research point to hallucinations, opaque data use and distrust of giving platforms too much influence over what they buy [2]. UserTesting's "2026 State of AI in Retail Experiences" covered 89 retail executives and 1,107 consumers across the US, Germany, France and the UK.
That study found that few consumers currently rely on an AI assistant like ChatGPT for product search, well behind search engines and retailer sites and apps [2].
That same study exposed a mismatch in where retailers are spending: executives prioritized AI-powered customer support as a top 2026 investment, while product search and discovery trailed. Consumers wanted the opposite, favoring smarter search and better recommendations over AI customer support [2].
Who approves what in agentic commerce today?
Sejdiji frames today's structure, in which a person signs off before anything changes in the live system, as a deliberate staging, not a permanent ceiling, with autonomy expanding as trust and model reliability grow [1].
Liability questions remain open at the policy level, too. Europe's payment regulatory framework, from PSD2 through PSD3 to DORA, was built around human authorization, and it isn't yet clear how autonomous AI-triggered payments fit into that structure [2].
- 1Today: agents prepare recommendations with reasoning, sources and expected effect; category managers review and approve every change before it goes live [1].
- 2Next: routine, low-risk tasks, like cleaning product data, run automatically, while higher-stakes decisions such as price changes move inside fixed margin and price guardrails [1].
- 3Later: broad decisions like delisting a product stay with a human approver for as long as the retailer wants that control; Zenline's own estimate for a fully autonomous retail organization in Germany is five to seven years, with Amazon, Temu and Shein already closer to that model today [1].
How ready is retail product data for AI agents?
An agent's recommendations are only as good as the data behind them, and most retailers aren't fully there yet. EHI found that a large share of surveyed retailers do not yet rate their own data foundation as good [1]. Publicis Sapient's "Guide to Next 2026" study found that most retailers still describe their product data as not fully structured and machine-readable [2].
The gap between confidence and reality shows up in that same Publicis Sapient study of more than 150 consumer-goods executives across five countries:
Publicis Sapient's own analysts dispute the high self-assessment outright, arguing that agent-to-agent commerce barely exists in practice yet [2].
Vendors have adapted to that reality by starting without internal data at all. Zenline begins by reading a retailer's public-facing catalog and benchmarking it against competitors before any internal data transfer happens, then layers in sales, margin and stock data through weekly uploads, no ERP integration required, to sharpen recommendations over time [1].
For context on how much runway this still has: German online retail reached 13.4% of total retail in 2024, while US e-commerce accounted for 17.1% of total retail sales in the second quarter of 2026, according to the US Census Bureau [3]. After roughly three decades of e-commerce, most retail revenue still moves through channels that AI agents barely touch yet.
- Most respondents say they're prepared for a future where AI agents interact directly with customers [2].
- A much smaller share describe themselves as highly confident in their underlying data [2].
- Few check regularly how AI assistants describe their products [2].
What commerce leaders should do now
Commerce and IT teams don't need to wait for shopper-facing AI to mature before capturing value. The category management opportunity is available today, provided the underlying sales, margin and product data are clean enough for an agent to reason over reliably [1].
Zenline recommends piloting on a narrow slice of the catalog, 15% to 20% of assortment is typically enough to demonstrate effectiveness, over four to eight weeks depending on how elastic the category is, beauty products move faster than electronics [1]. Measure the pilot against criteria the retailer defines up front, like avoided delistings or recovered margin, before expanding scope [1].
On the shopper-facing side, the lower-risk move is fixing fundamentals: a robots.txt that allows AI assistants to access the catalog, clean structured attributes, and well-maintained relationships between products, so that when consumer trust in agentic checkout does shift, the retailer isn't starting from zero [1].
Sources
- [1]etailment, KI-Agenten im Category Management: 'Dort wartet die Marge schon heute — etailment.de
- [2]etailment, KI-Agenten im Handel: Wie Wunsch und Realität auseinanderdriften — etailment.de
- [3]U.S. Census Bureau, Quarterly Retail E-Commerce Sales, 2nd Quarter 2026 — census.gov
