AI Grocery Shopping

Forty-six minutes. That’s the average American grocery trip — drive, park, navigate, wait, load. Every week. For decades. And now Albertsons says its AI shopping assistant does the same job in four.

I want to be specific about what “the same job” means, because the hype has been — unsurprisingly — a mess. The AI builds you a cart. It suggests meals. It cross-checks your pantry. Then you confirm before anything happens. That distinction matters, adn we’ll get to why in a minute.

So here’s where we actually are: Kroger has Google Gemini integrated into its shopping flow. Walmart is running an OpenAI partnership. Uber rolled out an AI personal shopper in February 2026. Albertsons went live across all its banners — Safeway, Vons, Jewel-Osco — with agentic cart-building in late 2025. The technology is functional. In production. Not vaporware.

And yet: 15% of US consumers have actually used AI for grocery shopping in the past year. That’s from Dunnhumby’s Consumer Trends Tracker, February 2026. Representative survey of US adults; exact n not publicly disclosed. Meanwhile, 69% say they’re interested, per PYMNTS Intelligence, January 2026. n = 2,299 US adults, Nov–Dec 2025.

That gap — 15% actual, 69% interested — is the whole story. Everything else is footnotes.

What “Agentic AI Grocery Shopping” Actually Means — Five Levels, One That Matters Right Now

The public debate is a disaster because it keeps conflating five different things. McKinsey published a framework in early 2026 that finally draws the map clearly. Here it is with current retail positions added.

Level 2 is where the industry is racing. Level 3 is where the hype lives. Confusing them is why the same tech gets called “here today” and “years away” in the same news cycle — and both camps are right, they’re just talking about different rows in that table.

One production number has arrived. Albertsons CTO Anuj Dhanda disclosed in a February 2026 CIO Dive report that the Ask AI tool drove a 10% basket size increase for users. Self-reported; covers predecessor tool, not the December 2025 agentic launch. No independent audit. Treat as directional. Albertsons has commercial incentive to frame adoption positively. It’s the first internal production number any major grocer has released publicly. Directional — not verified — but notable.

Baker’s framing is useful. But what it understates is how sharp the transition feels from inside a grocery app. The Kroger/Gemini integration — announced at NRF Big Show in January 2026 — doesn’t just suggest items. One instruction triggers delivery-time selection, loyalty-point optimization, and checkout sequencing simultaneously. That’s not a better search bar. That’s a different kind of tool.

You follow? The tech works. So why is usage stuck at 15%?

The $2.6B Warning Shot Everyone Ignored

Before the success stories, the failure. Because it contains the most important lesson in grocery automation that almost nobody is saying plainly.

In November 2025, Kroger announced it was closing three robotic Customer Fulfillment Centers — Wisconsin, Maryland, Florida — and writing off $2.6 billion in impairment charges. Plus a $350 million exit fee to UK partner Ocado. The partnership had launched in 2018 with plans for dozens of robotic US warehouses. Seven years later: digital business unprofitable, centers running below viable order volumes.

Here’s what’s underreported: Kroger built the facilities outside cities, on cheaper land, which worsened the economics automation was supposed to solve. A robotic warehouse picks faster than any human, sure. But it still has to deliver. With low population density and long delivery distances, per-order costs spiraled. Meanwhile, Instacart and DoorDash were expanding 30-minute delivery in the cities Kroger had stepped back from. Analyst Neil Saunders of GlobalData put it directly: the location decision was a key structural flaw — “hard places to make this model work.”

Agentic AI, operating as software on existing delivery infrastructure, has no equivalent capital risk. $2.6 billion can’t evaporate from a software rollout the same way it evaporates from fulfillment centers built in the wrong zip codes. But the strategic risk is analogous: deploy AI against the wrong problem, and you get abandonment, not adoption. And abandonment, unlike a write-down, is invisible until it’s a competitive rout.

The Trust Gap: 15% vs. 69%, and Why It’s Design-Solvable (But Not Equitably Solvable)

The 15% actual-usage figure vs. 69% stated interest — that gap has a specific anatomy. Dunnhumby isolates it: 38% of non-users “don’t see the need,” and 37% “prefer to make their own decisions.” US consumers are, Dunnhumby notes, “significantly less trusting of AI recommendations than the Americas average.”

And then there’s this: Bain’s research finds consumers trust retailers’ own on-site AI agents three times more than third-party agents like ChatGPT or Perplexity. That’s the window every major grocer is racing to jump through right now. It’s also why Amazon sued Perplexity in late 2025 — framing it as a “degraded shopping experience,” which is technically true and also not the real reason. The real reason is that an agent buying you the objectively best product at the lowest price bypasses sponsored listings, upsells, and discovery friction — the architecture that makes Amazon’s margins work. The conflict of interest is structural, and it means the consumer most likely to benefit from a genuinely objective AI shopping agent is, right now, the consumer least likely to get one.

PYMNTS’ analysis concludes something important: the barriers are largely design-solvable. Trust rises when autonomy is interruptible — approve before it acts, reverse after, invoke human review at any point. Consumers aren’t rejecting AI grocery shopping. They’re insisting it be accountable. Different thing.

But. And this is where the piece gets uncomfortable: “design-solvable” is not the same as “equitably solvable.”

Recommendation engines trained on majority purchasing patterns reflect the preferences of higher-income, urban, digitally-literate shoppers. They can systematically misweight cultural food staples with lower online review counts. Substitution logic can misfire on ingredients that don’t have robust US e-commerce histories. The 15% adoption floor isn’t evenly distributed — consumers with older devices, limited broadband, or less consistent digital infrastructure are disproportionately in the 85% not yet using these tools. And disproportionately in the populations that would benefit most from price-comparison automation.

Any grocery AI system not explicitly tested for substitution equity across income brackets and cultural food categories is, by default, optimized for the customers who least need the help. That’s not a reason to slow deployment. It’s the most important design constraint that no major retailer has publicly committed to solving. Sound familiar?

The 2028 Inflection Point — and Who’s Actually Positioned for It

Put the McKinsey automation framework, the Dunnhumby trust gap, and Kroger’s write-down together. What’s visible isn’t a technology inflection point. It’s a quality-of-trust inflection point.

The capability is ready. Albertsons, Kroger, and Walmart have Level 2 in production. Kroger’s Sage platform — actively used by 150,000 associates, now extending to customers via Gemini — is live. What’s missing isn’t capability. It’s the interruptibility infrastructure: consumers feeling that the agent works for them, not for the retailer’s margin architecture or a third-party platform’s data harvest.

Bain’s warning is direct: retailers have a short window to build on-site AI capability before the trust advantage over third-party agents closes. Once a consumer trusts Perplexity to do their grocery shopping as much as they trust the Albertsons app, the retailer loses the price premium, the loyalty data, the substitution margin, and the discovery surface — everything that makes grocery retail profitable at thin margins.

BCG’s December 2025 scenario describes retailers reduced to “mere background utilities in increasingly agent-controlled digital marketplaces.” Plausible risk scenario, not a confident prediction. BCG framing is explicitly scenario-based. Based on the current third-party agent adoption trajectory, it’s realistic for retailers that treat agentic AI as a feature launch rather than a trust-building program.

What to Do Right Now

The Honest Forecast

Grocery shopping isn’t going anywhere. The ritual, the pleasure of a good farmers’ market, the discovery of something new — none of that’s being automated. What’s ending is the 46-minute obligation for the same 30 items you buy every week. That’s a cognitive tax agentic AI can eliminate now, for anyone willing to try the supervised version that already exists.

The harder question — who benefits, and who gets optimized around — is the one nobody’s committed to answering publicly.

Kroger spent seven years and $2.6 billion learning that automating the wrong bottleneck produces write-downs. The industry is now discovering whether the right bottleneck — consumer trust — is actually automatable before someone else does it for them.

Whether that someone is Perplexity, ChatGPT, or a browser agent that doesn’t exist yet: the window is measured in months, not years. Bain said so. The usage numbers confirm it. And the retailers still running trust-infrastructure as a Q4 budget item are going to feel that mismatch.

So yeah.