The use of AI agents by consumers is growing, but not in the ways retailers and brands may think.
A new 57-page report from Koddi, the enterprise commerce media platform, tracked the rise of agentic AI in commerce across the U.S., U.K. and Germany found that the market is already moving, but consumers are only moving with it in tightly defined ways.
The report’s authors said people are increasingly open to AI acting as a co-pilot to help them search, filter, compare and recommend. Seventy-two percent of consumers polled said they want AI to be a co-pilot and not a full autopilot. The researchers describe a consistent pattern of “conditional acceptance” in their behavior. This is not a rejection of AI but an insistence on boundaries that protect consumers from mistakes they cannot easily undo.
Across categories, that boundary is shaped by perceived risk. Consumers appear most comfortable in retail, travel, groceries and tech, which are sectors where decisions feel familiar, reversible and less likely to produce lasting damage. The research found that they are least comfortable in banking, investments, healthcare and medication, where errors carry bigger consequences and the path back to correction is harder.
In practical terms, the study suggests a preference for agentic AI to simplify drudge work and logistics, such as narrowing choices or suggesting better options, rather than taking full control of emotionally loaded or high-complexity decisions.
The research also found that trust is the organizing principle behind acceptance, and it is “explicitly procedural rather than emotional.” Users of AI agents do not extend confidence to AI by default. Instead, they demand system guardrails such as secure data handling, human checks for high-value decisions and the ability to undo, cancel, review and change choices.
On the commerce media side of the equation, the report’s authors said AI agent users show a strong aversion to opaque commercial influence, with “paid placement” style nudges undermining trust more than many consumers initially expect. Even where monetization is not rejected outright, it must be clearly disclosed, because the biggest trust reducers are actions without approval and unclear use of personal data.
Trust is a key issue not only for consumers, but for merchants who want to ensure a positive online shopping experience for customers—especially when it comes to financials and personal data. Earlier this month, Sourcing Journal reported on two solution providers, Cloudflare Inc. and Onton, who rolled out technology that addresses these issues.
For retailers and brands, the implications are immediate and operational: winning will not come from louder claims about autonomy, but from designing permissioned delegation into the shopping journey.
The study points to a shift in what “value” means in commerce media, moving attention away from placements and impressions toward access, influence and measurable outcomes.
As AI advances from embedded assistance to more active decisioning, brands will need to rethink how they earn inclusion in the decision process itself, whether that inclusion is about relevance, price, reliability or clearly labeled sponsorship, because visibility alone is losing its edge.
The research also suggests where competitive advantage is likely to form over the next 12–18 months, as the market shifts from surfaces to systems. Advertisers are already investing at scale to test and scale agentic capabilities, but they have not fully settled on whether the best opportunities lie in onsite environments, in large language model experiences or in combined architectures.
The most successful players will treat governance as product design, the report’s authors said, which will done by building “reversible pathways” and keeping human oversight for consequential choices as well as disclosing when commercial incentives are at work.
In short, consumers want value before vision, and brands that “operationalize transparency and control” are the ones most likely to turn agentic momentum into more durable demand.