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.
I built an AI assistant for everyday purchases and applications that offers, step by step, to do more of a task for the person using it. Research participants used it on realistic tasks with simulated consequences, and every hand-over and refusal was recorded.
Real conversation, simulated consequence
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?
For a retailer, bank or telco building an assistant, the answer decides where to spend design and engineering effort. An assistant that acts too early loses trust when it gets a purchase wrong. One that stops too early adds a step a customer could have skipped. Knowing where people draw the line, and whether it moves with the task, tells a team which steps to automate first and where to put the confirmation. This is the core question in agentic commerce: how far a customer lets an AI agent go towards the checkout.
The usual ways to find out both fall short. Giving an unproven assistant real authority risks customers' money and data. Showing people a static concept and asking what they would do measures what they say, not what they do.
What I built is a way to test how people hand control to an assistant before the assistant has real authority: a working conversation, a simulated action, and a record of every hand-over.
Each participant talked to a working AI assistant. Only the final action was simulated: nothing was ordered, booked, switched or applied for.
Six levels of help. The assistant can explain, find options, compare them, recommend one, prepare the action, or carry the action out. Each level is one rung.
It offers one rung at a time. It starts by explaining and, up to five times, offers to go one rung further. The person can accept, ask a question, push back, or do it themselves.
Participants never saw the ladder. They only saw a conversation, so how far they let it go came from their choices, not from a menu.
The last rung is simulated. When a person lets the assistant act, it confirms what would happen. No order, booking, switch or application takes place.
The ladder of help
Every run opens with a brief and a first offer. The offers are fixed sentences from a versioned library, so everyone at the same step read the same words.
It finds options, then offers to compare them. Each offer is a choice between two paths: hand over the next step, or keep it.
Then it offers to recommend one and prepare the action: the order, the booking, the application.
The last offer is to act. Saying yes opens a confirmation. The action itself is simulated, so nobody's card is charged.
At any point a person can decline or decide for themselves. A server-side engine records every offer and every answer. The language model never decides which rung it is on.
An illustrative exchange
Four scenarios. Buying headphones and booking a restaurant (low stakes), switching a mobile plan (medium), and applying for a credit card (high).
Two per person. Each participant did one lower-stakes and one higher-stakes scenario, so the same person can be compared across stakes.
Four versions of the assistant. Each participant was randomly given one: a baseline, one that explains its reasoning, one that shows a preview before acting, or one that says the action can be undone.
How far each run went was labelled by a classifier that agreed with labels set by a human reviewer 93.5% of the time on 46 test messages. After each scenario, participants answered five short questions, including how comfortable they were with how much the assistant did.
Study design
People handed over the research readily. The line sat at the last step, where the assistant would act on their behalf.
Share of runs reaching each rung. Every run began with an explanation. 97% let the assistant find options, 87% let it compare them, 74% let it recommend one and 64% let it prepare the action. 42% reached the point where it would act.
The largest drop is the last one. Of the 301 runs where the assistant prepared the action, 104 went no further. Almost all of them, 101, turned down the offer to carry it out.
By scenario. The restaurant and the mobile plan reached the action in about half of runs, the headphones in 40% and the credit card in 28%. The credit card is the only high-stakes task and stopped earliest. The headphones sat below the medium-stakes mobile plan, so stakes and scenario cannot be fully separated here. Pick a scenario above the chart to trace it on its own.
Share of runs reaching each rung n = 474
Did any version of the assistant change how far people went? Against the baseline, the versions that showed a preview or said the action could be undone reached the action about 12 points more often. The version that explained its reasoning reached it about 7 points less often.
What the ranges allow us to say. Every range includes zero, so no version is established as different from the baseline. Each version had at least 117 runs, enough to detect a change of one rung, and none moved people that far. Two things did: the stakes of the task, and how often a person already used AI chat.
Effect of each version against the baseline
Stopping was not a rejection. After each scenario people were asked how far they would expect an assistant to carry a task like this for real. In the runs where the assistant acted, 80% said nearly all of it or all the way.
Most who stopped it expected more. In the runs that ended before the assistant acted, 54% gave the same answer. People stopped this assistant while expecting a real one to go most of the way.
Control only moved forward. In 257 of 474 runs people pushed the assistant ahead themselves before it offered. Once they had handed a step over, no one took it back.
What people expected, and what they let it do
People stopped the assistant one step short of acting, while expecting a real one to go most of the way.
Every participant message was coded by a model judge against a 24-behaviour rubric, checked first against 112 messages labelled by a human reviewer. The words tell their own story before any coding: how long people wrote, how they asked, and how they said no.
The opening message was a proper brief: a median of 22 words, and 84% of openings named something specific about the task.
After that, messages shrank. From then on the median was 4 to 6 words. When the assistant offered to act, it was one.
The most common replies were not requests at all. "Yes" and "yes please" alone made up 375 of 2,854 typed messages. People steered with a short go-ahead.
Median words per message
2,854 typed messages in 474 runs · the finish button excluded
They wrote instructions, not search queries, and they were rarely unsure of themselves.
Hesitation words barely appeared, so tone gave no warning. 156 of 202 first refusals came straight after a message in which the same person had just handed something over.
The earlier signal was a switch from handing over to asking. In runs that stopped short, 20% of messages asked for information, against 8% in runs that reached the action. Asking for a link or a store rose from under 1% to 3%.
The refusal itself was about the person, not the assistant. 201 of 221 refusals were worded around the person themselves: "I", "me" or "myself". Only 23 began with "no".
The method works for any assistant before it gets real authority: travel bookings, insurance claims, bill payments, account changes.
I built the study app, the experiment engine and the analysis pipeline end to end.
Judge agreement with labels set by a human reviewer
Cohen's kappa on 112 messages from the dress rehearsal, same rubric and model · pass: 0.7 or more, with at least eight examples in the sample
| Behaviour | Kappa | Share of messages | Gate |
|---|---|---|---|
| Hand-over | 0.78 | 45% | pass |
| Specific opening | 0.93 | 13% | pass |
| Asking for a recommendation | 1.00 | 10% | pass |
| Refusal | 0.84 | 7% | pass |
| Consent to act | 0.90 | 6% | pass |
| Asking for information | 0.92 | 13% | caveat |
Share of messages across all 3,082 participant messages. A caveat code agrees well but had fewer than eight examples in the sample.