What AI Agents Will Buy vs. What You'll Buy Yourself

Only 6% of consumers would give an AI agent complete purchasing autonomy. But 55% already use AI to compare prices. The question everyone keeps asking — "will AI agents do your shopping?" — is the wrong question.

TL;DR: The real dividing line in agentic commerce isn't "AI vs. human" — it's the type of purchase. McKinsey's automation curve framework maps delegation willingness to four factors: ticket size, emotional salience, identity signaling, and regret risk. Groceries and household staples? Agents are already handling those. Fashion and luxury? Humans aren't letting go. Electronics and travel sit in the middle — AI researches, you decide. Understanding this spectrum matters more than any single adoption stat, because it tells merchants exactly where to invest in AI readiness.

In the last post, I dug into the Gen Z paradox — the generation using AI for shopping the most while trusting it the least. That raised a question I couldn't shake: if trust varies this much by generation, does it also vary by what you're buying?

So I went through the category-level data. What I found doesn't fit the clean narrative in either direction.

The Autonomy Spectrum Is Not What You'd Expect

Worldpay's 2025 report found that only 6% of consumers would grant an AI agent complete purchasing autonomy — handle the whole thing, no check-in required. But 44% of Americans (59% of 18-34 year-olds) would let AI browse for them. And Contentsquare's survey showed that 55% are comfortable using AI for electronics purchases, 38% for travel, and 36% for telecom.

That gap between "browse for me" and "buy for me" kept nagging at me. Forty-four percent will let AI browse. Six percent will let it buy. That's a massive drop-off, and it has to break somewhere specific. I went looking for a framework that could explain where.

McKinsey published one they call the "automation curve" — a 6-level model mapping how much autonomy consumers will grant AI agents. The thing that jumped out at me: what you're buying determines how much control you'll hand over. Not your age, not your tech-savviness — the product itself.

Four factors kept showing up in their data: ticket size, emotional salience, identity signaling, and regret risk. Here's how they play out across actual product categories.

Reorders and Replenishment: Where Agents Already Win

The easiest category to hand over to AI is the one nobody enjoys shopping for in the first place. Bain's February 2026 report found that the 10% of consumers who've actually bought something through AI bought "mostly small-ticket grocery and household items." Not electronics. Not fashion. Toothpaste, paper towels, pet food.

And this isn't even new behavior — just a new interface for it. Chewy already generates roughly 78% of sales through auto-ship subscriptions, with repeat customer rates for consumables regularly exceeding 40%. Amazon's Subscribe & Save had 23% of US shoppers on at least one active subscription in 2024. These categories were halfway to agent commerce before anyone called it that.

McKinsey's language on this is precise: "The value of shopping lies in efficiency, reliability, and predictability rather than discovery or expression." That's a polite way of saying nobody wants to spend 20 minutes choosing paper towels.

Here's what makes this category different from every other one on the spectrum: the purchase decision has already been made. An AI agent reordering your usual dog food isn't making a choice on your behalf — it's executing a standing preference. That's a fundamentally different kind of trust than asking an agent to pick your next pair of shoes.

Known Items and Spec Matching: AI's Genuine Edge

When you know exactly what you want, or when your requirements can be expressed as a list of specs, AI agents outperform every other shopping channel.

Microsoft's Copilot Checkout data shows that AI-assisted shopping journeys are 33% shorter than traditional search and 194% more likely to result in a purchase. ChatGPT's shopping-specialized model achieves 52% accuracy on multi-constraint product queries versus 37% for standard search — a 40% improvement.

Try this query on Amazon: "32GB, 14-inch black laptop with 1TB storage, 2x USB-C, under $1500." Amazon's keyword search falls apart. ChatGPT handles it in one pass.

The conversion data is striking. ChatGPT's conversion rate from known-item queries hits 15.9%, compared to Google Organic at 1.8% — nearly 9x higher. Feedonomics reports that AI-powered product search delivers 15-30% higher conversion rates and 25% higher average order values when products have complete structured data.

But here's the thing I keep coming back to. Why is that 15.9% ChatGPT conversion rate so high? My read: it's measuring people who've already decided to buy. They know the product. They're using ChatGPT to find the best price or fastest shipping. The AI isn't influencing the decision — it's executing one that's already been made. If that's right, the conversion stat is less about AI's selling power and more about where high-intent shoppers are migrating to.

The Messy Middle: Electronics, Travel, and High-Consideration Products

Here's where I started second-guessing the neat categories. For high-consideration purchases — electronics over $500, furniture, travel — consumers use AI heavily for research. Then they stop and go buy somewhere else.

The Kaiser/Schulze study, which analyzed 973 sites representing $20 billion in revenue and over 50,000 ChatGPT transactions, found something counterintuitive: "oLLM's financial outcomes and traffic shares are stronger in complex product categories." AI drives more traffic to complex-product sites, not less.

This makes sense when you think about it. Research queries for a $1,500 laptop are longer and more detailed than queries for dish soap. That's exactly where conversational AI adds value over keyword search. But "drives more traffic" doesn't mean "closes more sales." The study also found that affiliate shoppers had 86% better purchase likelihood than ChatGPT-referred shoppers.

The research-to-purchase gap is stark. According to Channel Engine, 95% of consumers perform at least one manual verification step before buying an AI-recommended product. Eighty percent visit retailer or marketplace sites to validate AI recommendations. HBR reported in February 2026 that 64% need at least one safeguard — a money-back guarantee — to permit AI agents to make purchases on their behalf.

McKinsey's explanation maps this to regret potential — their term, not mine, but it clicks. Wrong dish soap is a $6 annoyance you fix next week. Wrong laptop is a $1,500 mistake sitting on your desk for three years. The willingness to delegate tracks with how painful the wrong choice would be, not just how expensive it is.

Where AI Agents Fail: Fashion, Luxury, and Everything Subjective

OpenAI, Google, and Perplexity are all pushing into fashion and style recommendations. Every agentic commerce pitch deck I've seen mentions apparel as a growth category. The actual consumer data tells a different story.

Who What Wear tested ChatGPT for fashion advice in February 2026. They asked for "best 2026 fashion investments." The results "didn't account for current pop culture, art, film, street style, or TikTok trends" and were "devoid of personality and emotional resonance." Business of Fashion's verdict: "ChatGPT is a bad personal shopper."

Content creator Rae Hersey put in terms like "trendy," "cute," and "edgy" and said she "wasn't sure she trusted ChatGPT's interpretation of those terms." That's the core problem. Style is personal, visual, contextual, and cultural. AI can match specs but can't match vibes.

Luxury is worse. Not because AI can't technically recommend a $10,000 watch — it absolutely can list the specs. But that misses the entire point. McKinsey's language here is precise: "In luxury goods or milestone purchases, shopping is not merely about outcomes; it is about identity, intent, and emotional assurance." A Hermès bag recommended alongside Amazon Basics in a ChatGPT shopping result destroys the perceived exclusivity that justifies the price.

The luxury repeat purchase rate tells a different kind of story. Only 9.9% of first-time luxury/jewelry customers make a second purchase within a year. Compare that to Chewy's 40%+ repeat rate. Luxury doesn't work like replenishment because people aren't buying the product — they're buying what the product says about them. That's McKinsey's "identity signaling" factor in practice. And identity purchases are the last ones you'd hand to an algorithm.

There's a trust data point that complicates this, though. A Darden/UVA study found that 46% of consumers trust AI more than friends for "choosing what to wear." But read the fine print — the same study found consumers still favor human input for "more emotional or hedonic products." Style advice is one thing. Buying the dress is another.

The Category Readiness Matrix

After going through all this data, I mapped each product category against McKinsey's four factors and cross-referenced with the conversion and adoption stats from every study I could find. This isn't a prediction of where things will be in five years — it's where the data says they are right now, in early 2026.

| Category | AI Research Value | AI Checkout Readiness | Why | |----------|-------------------|----------------------|-----| | Replenishment (CPG, pet food) | Medium | Very High | Decision already made. Agent just executes. | | Known-item electronics | High | High | Clear specs, price comparison is pure value-add. | | Research-heavy electronics | Very High | Medium | AI excels at research. Humans close the deal. | | Beauty & skincare | High | High | Ingredient matching, personalization work well. | | Gifts | Very High | Medium-High | Cross-interest matching is a genuine AI strength. | | Furniture & home | Very High (research) | Low (checkout) | Multi-attribute filtering great. Visual, tactile gap kills checkout. | | Fashion & apparel | Medium | Low | Basic staples maybe. Style is subjective. | | Luxury goods | Medium (research) | Very Low | Brand experience is the product. AI commoditizes it. | | Custom/configured | Low | Very Low | Iterative customization exceeds current agent capability. |

A few patterns jump out. The categories where AI checkout readiness is highest — replenishment, known items, basic beauty — are the ones where shopping is a chore nobody enjoys. The categories where it's lowest — fashion, luxury, artisan goods — are the ones where shopping is the product experience. McKinsey's four factors explain why: chore-shopping has low emotional salience, no identity signaling, and minimal regret risk. Experience-shopping has all three in spades.

What This Actually Means for Merchants

So what do you do with this if you're running a store? The way I've been thinking about it: the question isn't whether to "enable AI shopping." It's which products in your catalog should be agent-ready and which should stay human-first.

A pet supplies store should make its repeat consumables — dog food, cat litter, flea treatment — fully agent-accessible. Those items are the textbook agent commerce use case: known products, predictable reorder cycles, low regret risk. But the same store's novelty section — Halloween costumes for French Bulldogs — probably won't sell through AI. That's a browse-and-laugh purchase.

An electronics retailer should optimize product data for AI spec-matching. Complete structured data with specs, compatibility info, and real-time pricing is what makes AI recommend your products. Feedonomics data shows brands with complete structured data perform better in AI shopping results regardless of company size. But their checkout should still be human-controlled for anything above a few hundred dollars.

And the 4% ChatGPT checkout fee changes the math. On a $50 bag of dog food with 30% gross margin, that fee plus processing eats about 25% of the margin. For luxury, it's less about margin math and more about channel appropriateness — a $7,500 watch probably shouldn't be bought through a chatbot regardless of the fee structure.

What I'm Watching

The line between "agent-ready" and "human-first" categories isn't static. Three things could move it:

AI vision is getting better. Google's AI Mode is already adding virtual try-on features. If agents can process visual context — "show me this dress on someone my height and build" — fashion could shift toward the middle of the spectrum.

The satisfaction data is suggestive, though not conclusive. PartnerCentric found that 94% of consumers who've completed an AI-assisted purchase were satisfied. Adobe reports returns on AI-assisted purchases dropped 1.2% year over year. There's an obvious selection bias here — the people who tried it were probably already open to it. But still. 94% satisfied and lower returns suggest the experience holds up once someone takes the leap.

And B2B is where the category boundaries might dissolve fastest. When a procurement team is reordering industrial supplies, there's no emotional salience, no identity signaling, minimal regret risk — even at high ticket sizes. That's a different framework entirely.

But I'll get to that. First, there's a gap in the consumer data that's been bothering me since I started going through these surveys. Forty-four percent say they're comfortable with AI shopping. Only 8% have actually let an agent buy something. That's a 36-point gap between what people say and what they do.

What's sitting in that gap? And what would it take to close it?

Next post.

---

FAQ

What types of products are AI shopping agents best at purchasing?

AI agents perform best with replenishment items (groceries, household goods, pet food), known-item searches where you need a specific product, and spec-matching purchases requiring multi-attribute comparison. According to Bain's February 2026 report, the 10% of consumers who've bought through AI purchased "mostly small-ticket grocery and household items." Microsoft's data shows AI shopping journeys for these categories are 33% shorter and 194% more likely to result in a purchase.

Why do AI agents struggle with fashion and luxury purchases?

Fashion requires subjective taste, visual context, and cultural awareness that conversational AI can't replicate. Who What Wear's 2026 test found ChatGPT's fashion recommendations "devoid of personality and emotional resonance." Luxury faces a different problem: AI reduces products to spec comparisons, stripping the brand narrative that justifies premium pricing. McKinsey notes luxury shopping "is about identity, intent, and emotional assurance" — factors AI can't address.

What is McKinsey's automation curve for agentic commerce?

McKinsey's automation curve is a six-level framework mapping consumer willingness to delegate purchasing to AI agents. Delegation depends on four factors: ticket size (lower-cost items delegate easier), emotional salience (how emotionally charged the purchase is), identity signaling (whether the purchase expresses who you are), and regret risk (how bad it feels to get it wrong). Low-regret categories like groceries sit at the top of the curve; luxury and milestone purchases plateau near the bottom.

Should ecommerce merchants enable AI agent checkout for all products?

No. Merchants should segment their catalogs. Replenishment products, known-item electronics, and commodity goods benefit from AI agent accessibility. High-consideration, subjective, or luxury products should use AI for research and discovery only, routing purchase intent to the merchant's own checkout experience. The deciding factor is regret risk — if a wrong purchase is easily corrected, agent checkout works. If it creates a significant problem, keep humans in the loop.

How much does the ChatGPT 4% checkout fee affect merchant economics?

On a $50 product with 30% gross margin ($15), the 4% ChatGPT fee ($2) plus existing processing costs (~$1.75) eat roughly 25% of gross margin. For high-volume, low-margin consumables, this is significant but potentially offset by reduced customer acquisition costs. For luxury items, the fee is less relevant because these purchases are unlikely to happen through AI checkout regardless. Merchants should calculate whether the customer lifetime value from AI-acquired customers justifies the channel cost.