AI Shopping Agents: Why Businesses Aren’t Disrupted Yet

Activations are not transactions

Meta CEO Mark Zuckerberg said that more than 10 million people in the US and Canada had activated Muse, and that more than a quarter of them used it for shopping. That's a clear sign people are curious. But it doesn't answer the question that matters most: how many purchases were actually completed, how often users came back, and what merchants received in revenue.

This distinction is basic, but it's also easy to blur in a fast-moving market. Activation tells you reach. Shopping use tells you usage. A completed order is a transaction. And whether the channel holds up comes down to things like repeat purchases, customer lifetime value, and contribution margin.

Companies describing their own results have generally reported much smaller numbers. DoorDash said orders from Muse and other agents were in the hundreds. T-Mobile said it hadn't seen significant activity. Hilton described agent-originated bookings as insignificant, and eBay called them very small. None of those statements automatically contradict Meta's usage claims - they just measure different parts of the funnel.

The customer journey still has friction

Shopping agents have to deal with systems designed for humans. Websites often rely on visual controls, session state, login steps, or checkout behaviors that an automated assistant can't reliably interpret. And even when everything looks "online," normal business rules can trip things up - loyalty discounts, delivery windows, inventory changes, and restrictions on ordering outside business hours.

Chipotle, for example, said agents struggled with loyalty discounts and after-hours ordering. The company also said the Muse launch increased agentic orders by a single-digit percentage and that it started cleaning up its website code so agents could navigate the purchase flow more easily. That's a useful early lesson: the bottleneck might not be whether an agent can recommend something - it might be whether it can actually finish the checkout the business runs.

So what we have right now isn't a clean verdict. It's evidence that adoption, operational readiness, and successful transactions are uneven. For businesses, the real question isn't whether agents will replace every customer journey. It's which journeys they can handle reliably - and what has to change before customers start depending on them.

Ask what the metric counts

Before comparing announcements, figure out what a company means by "activity." Is it people who activated a tool, shopping conversations, referral traffic, attempted checkouts, or completed orders? Then check whether the company can identify the agent specifically, whether the data covers a representative period, and whether it factors in returns, cancellations, or repeat purchases.

A simple measurement ladder helps:

• Reach: How many customers can access the agent?
• Intent: How many use it to research or compare products?
• Referral: How many sessions arrive at a merchant from an AI tool?
• Conversion: How many sessions become completed orders or bookings?
• Economics: What are the average order value, returns, discounts, service costs, and margin?
• Retention: Do customers come back, and do they keep engaging with the merchant's brand?

Each level answers a different question. Strong reach can coexist with weak conversion. And low traffic can still be valuable if those customers convert well and cost less to serve. Without the economics, it's hard to trust any conclusion.

Attribution is not straightforward

Agent-assisted shopping can spill across multiple systems. A customer might ask Muse for recommendations, open a retailer's site, change the item, and finish checkout in an app. Depending on how the merchant tracks referrals, the order may show up as direct traffic, app activity, or an agent referral. Sometimes the reverse happens too: an agent referral can be credited even if the customer already decided what to buy before interacting with the assistant.

That makes comparisons between businesses fragile. If one retailer can clearly track agent referrals, it may report more activity than a competitor that can't separate agent-driven visits from other traffic. A low number might reflect limited customer use - but it might also reflect weak tracking, or both.

The safer response is to build a baseline and avoid forcing a dramatic story out of incomplete data. Track agent-attributed traffic and orders separately when you can, but also compare them against total conversion, margin, and customer satisfaction. And before you invest, review how the numbers are collected. In a new channel, measurement quality is part of the work - not an afterthought.

Start with the customer journey, not the trend

List the tasks your agent would actually handle. For a hotel, that could mean finding a room for specific dates and policies, then completing a booking. For a restaurant, it might mean choosing a meal, applying a loyalty reward, and picking an eligible pickup time. For an insurance marketplace, it might mean comparing offers without presenting inaccurate terms.

For each task, note where it can fail: missing or ambiguous product data, hidden fees, inaccessible controls, login barriers, shifting availability, or a promotion that only works under a narrow set of conditions. Prioritize issues that already hurt real customers. Don't build a custom agent integration until you know the underlying journey works.

A good pilot stays focused:

1. Pick one high-value task tied to a clear business outcome.
2. Test it with the agents or AI tools your customers actually use.
3. Track where the interaction breaks - wrong answers, stalled steps, abandoned checkouts.
4. Fix the content or workflow changes that help more than one channel.
5. Measure completed, profitable outcomes - not just clicks or agent activity.

Make your business legible to software

Agents rely on information they can interpret. Product names, prices, availability, service terms, delivery limits, and cancellation rules should be consistent and up to date. If the details are scattered across marketing pages, account screens, and checkout steps, an assistant can easily produce a partial answer or fail to complete the transaction.

This doesn't mean building a separate agent-only shopping site. It means reducing ambiguity in the systems that already represent your offer. Check whether key information is machine-readable, whether page behavior is predictable, and whether an automated visitor can tell an available option from an expired one. Keep human support and clear confirmation steps in place when a mistake would have real cost.

Be realistic about platform dependencies, too. Meta has partnerships with retailers including Walmart, Best Buy, and Gap, and Zuckerberg has said Meta plans to charge businesses a fee for transactions facilitated by Muse. Platform access may drive demand, but fees, customer ownership, data access, and control over the purchase journey are business terms - not implementation details.

Run the test with a defined owner, a limited budget, and a decision date. If the channel brings incremental customers at acceptable margins, expand it. If it mostly shifts existing customers into a more expensive path, adjust the terms - or stop. The point is to decide based on what you observe, not on the size of the announcement.

Fees, trust, and control

If a platform takes a fee on completed transactions, businesses need to compare that cost with the value of incremental demand. A sale that would have happened anyway isn't automatically worth paying a new intermediary fee. The analysis should consider acquisition costs, basket size, discounting, returns, customer support, and whether the merchant can build a lasting relationship with the buyer.

Control matters, too. A retailer may want accurate product representation and consistent customer service. A platform may want a predictable checkout experience and access to transaction data. Those goals can align, but leaders should work through practical questions about attribution, customer consent, refunds, customer support, and data use before scaling the partnership.

The market isn't locked in. Amazon blocked Muse, saying users would prefer its own Alexa shopping agent, then later said it expected to work with outside agents. Insurify also moved from blocking Muse to changing its approach after introducing technology meant to help agents present accurate offers. These aren't minor updates - they're proof that the business model and reliability are still actively negotiated.

Security and errors are business risks

An agent that can access an account or place an order creates a different risk profile than an agent that only answers general questions. Bank of America warned customers about sharing login credentials with AI assistants and said customers remain responsible for activity resulting from that access. The point is practical: convenience doesn't automatically settle who is accountable when an agent acts incorrectly or credentials are misused.

Businesses should separate low-risk discovery from high-impact actions. Reading public product information is not the same as changing a policy, spending money, or exposing account data. Use permission boundaries, clear customer confirmation, and audit records appropriate to the action. Don't ask customers to share credentials in a way your security and support teams can't safely handle.

A sensible policy should answer:

• Which actions can an agent take without confirmation?
• What information can it access, and for how long?
• How can a customer review, cancel, or dispute an action?
• Who handles errors, refunds, and security incidents?
• What data is shared with the platform, and for what purpose?

These controls can add friction, and that's real trade-off - not a reason to skip them. For low-cost, reversible actions, a lighter confirmation flow may be reasonable. For financial, contractual, or sensitive account actions, stronger checks are usually justified. Match the level of control to the consequence of getting it wrong.