Agentic Commerce
Human-AI Trust
Mixed-Methods Research
Only 17% of people will let an AI buy for them. Every agent built so far is loyal to the platform, not the person.
Commerce is going agentic fast: 38% already shop with AI (Adobe, 2025), yet 89% double-check what it does. Rufus, ChatGPT, and Perplexity are all structurally optimized for platform conversion, not consumer welfare. Consumer Reports is the only player without that conflict. The brief handed to our team: Conscious Commerce is an agent that buys in line with your values and stays loyal to you.
Trust isn't earned at the recommendation. It accumulates in the low-stakes moments, and Consumer Reports isn't in them.
We ran a survey with 224 participants, 50 long-form interviews, 9 expert sessions, 9 informal interviews, and ethnographic fieldwork across delegation, AI trust, user values, and the shopping journey. One pattern held throughout: people build confidence through small, low-stakes research (e.g., a Reddit thread, a quick question to an AI assistant that's free to access) then lean on whatever tool was already there when the big decision arrives. People name Consumer Reports as the source they trust most, but by the time they open it, they've usually already made up their mind elsewhere. The AI was in the tab the whole time. Whoever is present for the low-stakes moments wins the high-stakes ones.
The conditions for delegation were consistent and unforgiving. Each finding came with a hard line designers can't cross:
What we found | The guardrail it sets |
Dominant shopping values function as identity, not preference | Optimize against someone's one non-negotiable, and you've lost them: it reads as betrayal |
People hold ethical values they can't verify | The gap isn't motivation, it's verification, which an independent party is uniquely built to supply |
Delegation stops at irreversible actions | No undo, no delegation. An agent can act only if every step is reversible or backed by clear accountability |
Users set hard boundaries, not soft preferences | Cross a user-defined line once, and willingness to delegate can drop permanently |
An agent that runs autonomously, until the stakes say it shouldn't.
The research resolved into one interaction model: research mode and execution mode. In research mode, the agent has high autonomy: it explores, drafts, and surfaces options across the space. In execution mode, the human is always in control: reviewing, adjusting, and approving before any irreversible action. The handoff between them scales with stakes: the higher the cost of being wrong, the earlier control returns to the person. People aren't ready to hand off control. They're ready to share it.
For the divergent prototyping phase, I built this interaction model as a working browser plug-in: the agent works alongside you in the page, then hands control back before any irreversible action.
[ Live prototype → ]
Delegation has a hard floor: people won't hand off anything they can't undo.
The boundary for trust moves with the stakes: too much autonomy and people feel unmonitored; too little and the agent is useless. The team is now prototyping the human-in-the-loop model this research defined, testing whether the guardrail patterns hold in real use. The direction is set; what's in test is how much control people will share, and at what stakes.
