April 6, 2026

The proof now exists. Upvalo reports that Walmart's autonomous checkout pilot lifted conversion by 18% in its first month and dropped cart abandonment from 70% to 52% in pilot markets. The agent confirms order intent in natural language, suggests related items inside a customer-defined budget, and routes payment and fulfillment without a redirect to a checkout page. The economic logic is unromantic. A shopper who says "household essentials under $300" gets a basket built from prior purchases in seconds, with one approval gate. That is not a chatbot improvement. It is the elimination of the cart funnel for habitual categories, and Walmart can run it because it owns the catalog data, the payment rails, and the customer history that make the agent's first guess accurate.

Discovery is no longer the only signal worth tracking. McKinsey's February 2026 consumer survey of 4,008 shoppers, summarized in Retailgentic's roundup of agentic commerce data, finds that 23% of AI-using shoppers are now completing checkout through agents, not just browsing. Adobe's 2025 holiday data, as reported in a LinkedIn analysis by Anthony DeMotte and Mike Kramer, shows AI-driven US e-commerce traffic grew 758% year-over-year in November and 670% on Cyber Monday alone. The same analysis cites Deloitte's projection that 81% of retail executives expect generative AI to weaken brand loyalty by 2027 because algorithms optimize for price, availability, ratings, and returns terms, not affinity. Sara Wingo's Retailgentic piece adds that 91% of merchant leaders expect AI to influence at least 20% of their business by 2027, 57% are implementing agents this year, and 47% are committing more than $1M in 2026 specifically.

Bain research, also cited by DeMotte and Kramer, sharpens the strategic question. Consumers trust retailers' own on-site agents three times more than third-party assistants today, but Bain frames that gap as a closing window. About half of retail executives in the same dataset expect the multi-step shopping journey to collapse into a single AI-mediated interaction by 2027. The retailers using their trust lead to learn now will be the ones whose proprietary models get tuned on real first-party behavior before ChatGPT, Gemini, and Copilot reach parity. Trust does not transfer between cohorts. It compounds inside the cohort that uses it first.

Amazon is converting that compounding into ecosystem lock-in. The same Upvalo analysis reports that Amazon's buying agents now account for 40% of voice-based shopping. Amazon owns the device, the assistant, the payment instrument, the inventory, and the fulfillment network, which means a voice reorder rarely surfaces a competing offer. Competing retailers cannot match that vertical integration. What they can do is make their catalogs and policies legible to the agents shoppers are already using elsewhere. nShift's research on agent legibility frames the requirement plainly. Humans tolerate ambiguous delivery windows and fuzzy returns terms. AI agents do not. Offers with unclear shipping costs, missing return windows, or contradictory inventory data get skipped in favor of cleaner alternatives. Catalogs are now data products, not marketing assets.

Harder ROI is showing up in B2B, with one important constraint. 8allocate's documentation of industrial agentic deployments describes Danfoss handling more than 80% of transactional decisions on email-based orders through an agentic system on Google Cloud. The agent extracts customer ID, part numbers, quantities, and ship-to data from unstructured email, validates against ERP, and creates orders within configured guardrails. Exceptions route to humans. Siemens and PepsiCo demonstrated a related approach at CES 2026, using a Digital Twin Composer where agents simulate supply chain reroutes against physics-level models before executing them in production systems. These deployments show very high automation rates in domains where order data, contracts, and logistics are already structured. They also show why generic autonomous procurement is not happening at scale. The ceiling on B2B agentic commerce is not the model. It is the surrounding governance and integration.

Forrester's analysis of zero-click checkout in answer engines names the equilibrium that is emerging. Agents own intent capture, search, and configuration. Merchants retain checkout, fulfillment, liability, and compliance. The OpenAI pivot to merchant-controlled checkouts is consistent with this division of labor, not a temporary retreat. Merchants who matter in that equilibrium are the ones whose backends already speak protocol. UCP and ACP are the interfaces, but the readiness gap is data, not adapters.

Capital is following the analysis. Wingo notes in the same Retailgentic roundup that retailers are pausing replatforming projects and major on-site rebuilds to fund off-site discovery investments, especially AI-powered product discovery through external agents. Growth HQ reports that up to 25% of referral traffic for leading retailers is already being routed by algorithms rather than human browsing. Incremental UX work on a merchant's own product detail pages now yields diminishing returns relative to making goods, prices, and policies machine-selectable across answer engines. That trade was unthinkable 18 months ago when conversion-rate optimization was the dominant retail discipline. It is now a defensible reallocation.

What changes in 2026 is that the CFO now has a number to ask about. Walmart's 18% conversion lift, Amazon's 40% voice share, and Danfoss' 80% automation rate are no longer projections. They are line items, and competitors who cannot produce comparable numbers will be asked to explain why. The next budget cycle will not reward retailers who spent 2026 evaluating agent vendors while their peers were measuring agent revenue. Walmart, Amazon, and Danfoss did not run a pilot to learn whether agents work. They ran one to learn how big the gap gets when only one side is measuring it.