Independent experiment · Agentic commerce · 2026
Where people stop an AI assistant that offers to act for them.
People let it find, compare and recommend. At the step where it would act, they turned it down one time in three, and four in ten runs ended with the assistant acting.
42%
Of 474 runs let the assistant act
101 of 308
Offers to act turned down
28%
Of credit card runs reached the action
54%
Of runs that stopped short still expected a real assistant to go most of the way
The question
AI products are moving from answering questions to taking actions: placing orders, making bookings, submitting applications. That raises a product question: how much of a task will people let an assistant do on its own?
The method
I built a working AI assistant for everyday purchases and applications. It can explain, find options, compare them, recommend one, prepare the action, or carry it out, and it offers one level at a time. Only the final action was simulated: nothing was ordered, booked, switched or applied for.
239 research participants completed 474 runs, two tasks each, one lower and one higher stakes. Each was randomly given one of four versions: a baseline, one that explains its reasoning, one that shows a preview before acting, or one that says the action can be undone. A classifier labelled how far each run went, agreeing with a human reviewer 93.5% of the time on 46 test messages.
Results
| Level of help | Share of runs |
|---|---|
| Find options | 97% |
| Compare them | 87% |
| Recommend one | 74% |
| Prepare the action | 64% |
| Act | 42% |
What it found
- The largest drop is the last one. People handed over finding, comparing and recommending readily, then turned down 101 of 308 offers to act.
- A preview, an undo promise and an explanation changed nothing measurable. Each version had at least 117 runs against the 80 needed to detect a one-rung change, so finding no difference is itself a result.
- Two things did move people: the stakes of the task and how often a person already used AI chat. The credit card, the only high-stakes task, reached the action in 28% of runs.
- Stopping was not rejection. In runs that ended before the assistant acted, 54% still expected a real assistant to go most of the way or all the way.
My role
Solo project: research design, assistant design and build, study platform, analysis and write-up.
Participants were recruited and paid through Askable.