EXPLORATION ยท IN PROGRESS

Trusted Cart: AI agents that shop for you

TL;DR

In progress ยท Research phase

An autonomy ladder for AI agents that shop and pay

Visa, Mastercard, and American Express have each launched programs that let AI agents pay on a shopperโ€™s behalf, but the controls serve banks and merchants, not yet the person whose money is on the line. Iโ€™m designing how people set an agentโ€™s limits, see what it did, and take back control. The core idea is an autonomy ladder, where each kind of purchase earns its way from Suggest to Autopilot. A clickable prototype of a household shopping agent is built, and research starts next.

Product strategy
Product strategy
Trust design
Trust design
AI prototyping
AI prototyping
MY ROLE

Self-initiated ยท End-to-end product design

TEAM

Solo, with research participants starting soon

TIMELINE

8 weeks ยท 2026

TOOLS

OVERVIEW

The rails are built.
The human experience isnโ€™t.

Between March and April 2026, Visa, Mastercard, and American Express each launched programs that let AI agents pay on a shopperโ€™s behalf. None of them gives an agent unrestricted access to an account. Each relies on scoped credentials, spending limits, and merchant rules.

Designing for

Households handing off routine shopping ยท Travelers who need help fast

THE CHALLENGE

Controls built for banks, not yet for people

Those controls serve banks and merchants well, but not yet the person whose money is on the line. Whatโ€™s still missing is a clear way for people to set those limits, see what an agent did, and take back control when something goes wrong. Thatโ€™s the problem Iโ€™m working on.

THE APPROACH

Working draft

An autonomy ladder

Each kind of purchase sits on its own rung, and an agent earns its way up through consistent good choices.

1 ยท Suggest

The agent finds options.

You pick and buy.

2 ยท Prepare

The agent fills the cart.

You review and pay.

3 ยท Ask, then buy

The agent buys after your one-tap yes.

You approve or skip.

4 ยท Autopilot

The agent buys within your limits.

You can always undo.

More oversight

More autonomy โ†’

Each step up hands the agent more of the work. A category climbs only after a run of good orders.

KEY DECISIONS SO FAR

Three decisions shaping the ladder

These are working decisions and may change as research comes in.

Earn trust per purchase

Each kind of purchase has its own rung. Trust earned on groceries doesnโ€™t carry over to flights.

Step down when unsure

When the agent is unsure, or a rule would have to bend, it steps down a rung and asks instead of guessing.

Test opposite ends

Groceries show when an agent can move up a step. Flights show when it should step back and ask.

HOW Iโ€™LL TEST IT

Two scenarios, two ends of delegation

If the ladder holds at both extremes, it should hold for most purchases in between. Iโ€™ve built a clickable prototype of the household scenario to start research. A version with a real AI model comes next. The goal is calibrated trust: people should trust the agent more when itโ€™s reliable, and less when it isnโ€™t.

Keep us stocked

A household agent that keeps groceries and supplies on hand, under $200 a month.

STAKES

Low ยท Weekly

TESTS

Standing rules ยท Budget pacing ยท Substitutions

Rebook me

An agent that rebooks a canceled flight within rules the traveler sets.

STAKES

High ยท Rare ยท Time-pressured

TESTS

Speed ยท Explaining a choice ยท Undo and refunds

The prototype

A clickable research prototype of Keep us stocked, showing the four moments itโ€™s built to test.

Try it
Watch the agent earn trust, or jump to any step
Asks before a price jump

Detergent costs $4 more than usual, so the agent waits for an OK.

1 / 4

WHAT SUCCESS LOOKS LIKE

Signs of calibrated trust

Calibrated trust means people trust the agent as much as it deserves, no more and no less. In testing, Iโ€™ll look for these signs.

01

Trust grows with a track record

People move an agent up a rung after consistent good choices, not on day one.

02

Mistakes lower trust, briefly

After an error, people tighten limits, then loosen them again as the agent recovers.

03

Control stays clear

People can say what the agent did and undo it without help.

04

Stakes change behavior

People give the grocery agent more freedom than the flight agent.

RISKS AND OPEN QUESTIONS

What I havenโ€™t solved yet

These are the hard problems the ladder has to hold up against. I donโ€™t have full answers yet, and research will shape them.

When itโ€™s confidently wrong

An agent can be certain and still wrong, like reordering the wrong size without hesitation. Every purchase has to be easy to spot and undo, even when the agent didnโ€™t flag it.

Whoโ€™s accountable

If an agent makes a bad purchase within limits a person set, who pays: the person, the merchant, or the agentโ€™s maker? Limits need to be clear enough that people know what they agreed to.

Fraud and misuse

Agents with payment access are a new target. Scoped credentials and spending limits help, but people still need a fast way to notice something odd and shut it off.

WHATโ€™S NEXT

Follow along

Research starts now, with a clickable prototype of the household scenario. Next come the framework, the key flows, and a working version with a real AI model. When the work is done, a full case study with test results will replace these notes.

Want to talk about it? Iโ€™d love to hear from you at lynngee@gmail.com.

More work

Letโ€™s work together

ยฉ 2026 by Lynn Gee