Kroger and Other Retailers Train Shoppers to Use AI Agents for Commerce

Full analysis

Kroger, DoorDash, and Giant Eagle say they are training shoppers to use AI agents, software that shops for you, as a routine part of grocery buying. Strip the behavioral-science language and what's happening is a feature launch with a monetization motive underneath. The question for anyone building commerce or retail media is whether this is a new place to put ads, a real shift in how products get discovered, or a demo that never clears the trust bar.

This is easy to undo on the consumer side. Nobody is locked in. A shopper who gets one wrong substitution deletes the habit faster than it formed. What's actually being decided is whether agents become a real buying surface that retail media can sell against, or stay a novelty bolted onto apps people already use. Nothing sets a deadline here. There's no shutdown, no contract, no price change forcing the issue. That absence matters: it means the only thing pushing this forward is the retailers' own monetization pressure, and the only thing that settles it is retention data nobody has published.

The Skeptic. "Training shoppers" is a flattering way to say "we shipped a feature and hope it sticks." Grocery habits move at a glacial pace. Curbside pickup took roughly a decade to normalize, and that fixed a concrete problem: you didn't have to walk the aisles. What problem does an agent fix for someone who already has a saved cart and a loyalty app that reorders in two taps? The saved cart is the competitor here, and it's free, instant, and already trusted. The real engine is retail media revenue. Agents are a fresh ad surface, and the convenience story is the wrapper. Show me week-over-week re-engagement before I call this a trend.

The Safety Lens. The authorization gap is the part nobody has priced. When a Kroger agent buys on your behalf and swaps in a product that triggers an allergy, who owns that? Current deployments barely specify what an agent can buy, how much it can spend, or what it's allowed to substitute. There's no rulebook for autonomous purchasing agents. The FTC's unfair-practices doctrine and state consumer protection law are the closest things, and neither was written for this. There's also a quieter issue: these agents train on your full purchase history, building a behavioral profile that sits well outside what most people think "my grocery app" knows about them.

The Researcher. The interesting signal is that retailers aren't waiting for the tech to be ready. They're running live habit-formation experiments on their own customers. The open question is whether agents widen or shrink what you consider buying. Decades of recommender-system work point one way: algorithmic middlemen narrow choice. If an agent optimizes for convenience and your past behavior, discovery collapses into a reinforcement loop, you buy what you already bought, and brands trying to break in lose their shot at the shelf. There are no published numbers yet on how often these agents convert correctly, how far preferences drift, or how often they just get it wrong.

The Builder. The plumbing is harder than the announcement lets on. Kroger's systems span loyalty data, live inventory, substitution logic, and fulfillment routing. Stitching an agent across all of that without inventing SKUs that don't exist or botching substitutions is a multi-quarter job, not a toggle. The thing that breaks first is substitution: an agent confidently swapping a product the shopper never approved. Trust goes once and doesn't come back. DoorDash has the tightest feedback loop, since delivery confirms in near-real-time, but the loosest catalog. Giant Eagle's tech stack is probably the ceiling on its ambition, not its appetite.

The Compute Pragmatist. This is light on raw compute by frontier standards but tight on latency where it counts. "Oat milk is out, what does this shopper actually want" is a sub-second call against a preference model and live inventory, made during fulfillment. That's not a job for a big general-purpose model. It's a small fine-tuned model or a retrieval lookup running close to the warehouse. The real infrastructure choice is whether Kroger and DoorDash build their own preference models or route through someone else's API. Routing through an outside API ships your customers' purchase behavior to a third-party model provider. That's a data governance problem, and nobody has addressed it out loud.

Where they disagree. The Skeptic and the Researcher are looking at the same fact and reading it opposite. The Skeptic says there's no evidence of a behavioral shift, so this is a launch pretending to be a trend. The Researcher says the shift is real because retailers are actively conditioning for it, and the thing worth measuring is which direction discovery moves. Both can't be the headline. The second split is Builder versus everyone selling the vision: the Builder says substitution errors kill trust inside a quarter, which means the product's reliability caps the ambition long before the retail media team gets its new ad surface. And the Safety Lens sees an authorization and liability hole that grows exactly as fast as adoption, so success and exposure arrive together.

What this hinges on. Two things. First, does anyone shop through an agent twice? Retention, not launch announcements, is the whole ballgame, and no retailer has shown it. Second, does an agent reliably buy the right thing? Substitution accuracy is the trust gate, and a single bad swap undoes weeks of coaxing. If you're building in or around retail media, the thing to check this quarter is whether your agent partner can show re-engagement data, not install counts, and whether the authorization scope (spend limits, substitution rules) is specified tightly enough to survive an angry customer. Everything else is downstream of those two.

The council leans skeptical. Not because agentic commerce can't work, but because the evidence on offer is a quote about intent, and intent is cheap. The saved cart is a brutal incumbent. The monetization pressure driving this is a problem the retailers own, not one shoppers feel.

Prediction: Neither Kroger nor DoorDash will publish week-over-week retention or repeat-use data for their AI shopping agents by the end of Q1 2027 earnings season in May 2027; the public pitch will stay on launches, partnerships, and total users rather than on how many shoppers come back and buy through an agent a second time.

Confidence: Medium. The silence protects a number that is almost certainly weak.

Why: The only evidence in this story is a stated intent to "train shoppers," with no retention figure attached, and grocery habit change is historically slow, which means early repeat-use is probably low. Retailers disclose the metrics that flatter them and bury the ones that don't, so if week-over-week re-engagement were strong, they'd lead with it, the way they lead with loyalty-app numbers. A vanity metric like total users or orders placed lets them claim traction while the harder question, do people come back, stays unanswered. The opposite outcome, a retailer volunteering soft retention data, runs against every incentive a public company has when the number isn't good yet.

Revisit by 2027-05-31: We're right if neither Kroger nor DoorDash has disclosed a week-over-week or repeat-use retention figure for its AI shopping agent by the close of Q1 2027 earnings season. We're wrong if either one publishes such a retention figure (not total users, orders, or installs) in an earnings call, investor deck, or press release before then.

That silence would itself be the answer. A product people re-use is a product whose owner can't stop talking about re-use.