Last weekend, Amazon told Meta’s new AI assistant Muse that it wasn’t welcome.
People using Muse to shop on Amazon began seeing an error: the agent’s continued access violated Amazon’s conditions of use. Amazon said the issue was partly about transparency, privacy and security. Meta hadn’t told Amazon that Muse would be browsing its store, and the agent wasn’t identifying itself as an agent.
It would be easy to file this under Big Tech turf war. Amazon versus Meta. Another fight over who gets to control the interface.
But there’s a more interesting question hiding inside it.
What happens to the customer journey when the customer is no longer the only thing moving through it?
For decades, market research has treated the human being as the meaningful unit of analysis. We study what people notice, what they remember, what they compare, what persuades them and, eventually, what they choose. Marketing is built around much the same assumption. Get into the consideration set. Win attention. Build preference. Make the sale.
AI shopping tools complicate that sequence because they can intervene before a person sees the options.
That doesn’t mean consumers are handing over their credit cards and going to bed. Far from it.
A Gartner survey of 322 US consumers in January found that willingness to let AI make purchase decisions topped out at just 11%, even in relatively low-stakes categories. But 31% were willing to let AI narrow choices for household supplies, and 28% for personal electronics.
That gap is revealing.
People may not want AI to choose for them. They may be quite happy for AI to choose what they choose from.
And those are very different things.
Visa’s research across the US, Australia and New Zealand points in the same direction. Consumers showed substantial interest in delegating parts of discovery and evaluation – searching, browsing retailers, comparing options and reading reviews — while remaining sensitive to control, privacy and purchase accuracy.
This is where the customer journey starts to look different.
Imagine someone asks an AI assistant for a laptop suitable for frequent travel, video calls and light data analysis. The assistant searches, filters, compares and returns four options.
The human still makes the final decision.
But perhaps hundreds of products were reduced to four before the human became meaningfully involved.
What happened to everything else?
Traditional research might ask why the consumer selected laptop A rather than laptop B. But there is now another question upstream: why did the agent show A and B in the first place?
That filtering layer could become increasingly important for brands. OpenAI, for example, says people are increasingly beginning their shopping in ChatGPT to explore, compare and decide what to buy, and has expanded its Agentic Commerce Protocol to support product discovery.
So the nature of visibility starts to change.
Digital marketing has spent years competing within algorithmically mediated environments: search rankings, recommendation systems, feeds and marketplaces. But those systems have largely been designed to rank and surface options for people to browse.
AI agents could take that mediation a step further. Instead of simply influencing what appears first, they can help assemble the consideration set itself – searching, comparing and excluding options before the consumer sees them.
For brands, that could mean competing not only for human attention, but for machine consideration.
That could depend on things marketers aren’t used to treating as persuasion: structured product information, availability, price clarity, reliable reviews, return conditions, compatibility, provenance – or whether an agent can access the environment at all.
Amazon blocking Muse makes that last point unusually concrete. Before an AI agent can recommend or buy something, somebody has to let it through the door.
This creates an interesting problem for research.
If AI systems increasingly mediate discovery, asking consumers only about the options they remember seeing could tell us less about how their consideration set was actually formed. We may need to understand the interaction between human preferences and machine filtering: what instructions people give agents, what attributes agents prioritize, which sources they trust, what they exclude, and how much consumers inspect the reasoning behind the shortlist.
There’s plenty of reason to be cautious. Agentic commerce is surrounded by the usual technology-cycle enthusiasm, and current evidence is mixed. Gartner found that 54% of consumers who had used AI while shopping for a recent purchase said they had to double-check all the information it provided. Fully autonomous purchasing remains uncomfortable for many consumers.
Maybe that persists.
But autonomy isn’t actually required for the deeper change to begin.
An agent doesn’t need permission to buy your laptop to influence which laptop you buy. It only needs permission to decide which ones are worth showing you.