A real-world agentic commerce diary

The $20 checkout that showed me the promise of AI shopping

I began with a birthday-gift idea and a budget. I ended with a completed book order, a payment lesson, a privacy cleanup — and genuine excitement about what conversational commerce could become.

Discovery was the breakthrough

A conversation replaced hours of tabs, filters, and second-guessing.

The goal was simple: find a small bundle that reflected my daughter's interests in reading, music, and learning to build with technology — all within a $50 budget.

Reading

New Kid: A Graphic Novel, chosen as an engaging and inspiring read.

Building

A micro:bit v2 board for a first hands-on coding project together.

Music

Removable piano-key note stickers to support her piano practice.

This was the clearest win. Instead of translating an idea into keywords, opening merchant tabs, comparing variants, recalculating a bundle, and keeping a mental record of tradeoffs, I could simply explain the person, the occasion, and the constraint. The conversation did the organizing. I still made the choices, but I did not have to perform all the research mechanics myself.

That convenience is real: conversational discovery compressed what could have been hours of research into a focused shortlist I could understand, question, and approve.

The first estimate did miss an important detail. The book's $14.90 item price became $20.02 after $3.99 shipping and $1.13 tax. The prepared micro:bit checkout was $28.32. Together, those two items alone would total $48.34 before the piano stickers. That lesson did not erase the discovery value; it sharpened the requirement that future recommendations show an estimated all-in cost, not just product prices.

The first transaction

A virtual card, an error message, and a human takeover

Unused Link authorization $25.00
Bookshop purchase $20.02
Temporarily affecting balance $45.02

The $25 line was not a second completed purchase. It was an uncaptured authorization hold tied to the one-time virtual card. The purchase was only $20.02, but until the unused authorization fell away, my available balance felt the effect of $45.02.

The exact reason the virtual-card payment was rejected was not established. What was clear was that my physical card completed the purchase, the virtual card recorded zero captures, and the unused $25 authorization could still reduce my available balance while it remained pending.

  1. A one-time card appeared

    Stripe Link generated virtual card details for the Bookshop checkout. I did not realize at the time that the number on screen was intentionally different from my physical card.

  2. The form rejected the payment

    The checkout displayed an "incorrect payment details" error. The automated flow did not recover cleanly and kept cycling instead of stopping for review.

  3. I took over

    Because the card number looked unfamiliar, I assumed it had been entered incorrectly. I manually replaced the virtual-card details with my physical card information.

  4. The order completed

    The book purchase went through for $20.02. The book was estimated to ship in 4–10 days.

Human control built trust

I was never asked to surrender the final decision.

The rough edge mattered, but so did the structure around it: I could inspect, interrupt, question, correct, pause, and decide what happened next.

Control

I approved purchases merchant by merchant. When the payment flow became confusing, I took over. When the balance became harder to read, I stopped the remaining orders.

Transparency

I could ask what each bank line represented, challenge explanations that sounded too certain, and separate a completed charge from an unused authorization.

Privacy

After realizing that I had entered my physical card during takeover, I asked for checkout snapshots containing it to be deleted and for retained files to be checked.

Trust did not come from pretending the flow was flawless. It came from keeping me in charge while the system explained, corrected, and paused.

The early explanation of the failure was too specific: it attributed the problem to the security-code step without evidence that established that cause. A more trustworthy account is narrower. The checkout rejected the details; the automation looped; I intervened; the order succeeded with another card. Being able to revise the story when the evidence changed was an important part of the experience.

What was verified: the physical card completed the $20.02 purchase. The one-time virtual card had zero captures and an expiry date of September 29, 2026. The available payment tool could not cancel or release its $25 authorization. The issuer controlled when the pending hold would disappear.

Privacy note: this account intentionally omits all card numbers, security codes, bank identifiers, shipping details, the merchant order number, and my daughter's name.

The positive case

The experience was genuinely good — not despite the lessons, but partly because I could work through them.

Navigating the journey felt less like operating shopping software and more like collaborating with someone who could research, explain options, prepare a checkout, and wait for my decision.

What felt new

Intent became action. I could begin with a human goal — a thoughtful birthday bundle — rather than a product identifier or merchant name.

What stayed familiar

I held the authority. The agent helped with the work, but approval, takeover, privacy decisions, and the choice to pause remained mine.

Muse and Meta deserve credit for making the front end of the journey feel natural. The conversation kept context, helped shape the gift idea, narrowed the field, and carried the intent toward checkout. Stripe deserves credit for attempting to bridge that conversation to a one-time virtual payment instrument. Shopify also deserves credit for choosing participation: days earlier, it announced a deep partnership to enable Muse checkout through Shop Pay across Shopify stores.

A fair qualification: I could not see the exact division of responsibility among every company behind every screen. My compliment is about the experience and the direction of the ecosystem, not a claim that all three companies powered each step of this particular Bookshop transaction.

The most encouraging part is that a failure did not have to end the journey. I could take over manually, finish the purchase, return to the conversation, understand the bank impact, protect my privacy, and decide to wait before continuing. That continuity is rare in today's fragmented shopping flow.

For me, the headline is not "AI bought a book." It is "a conversation carried a real human intention almost all the way from idea to purchase, while leaving the human in charge."
The ecosystem is not uniform yet

Discovery and transaction readiness are two different layers.

My Gemini comparison

In my own test, Google Gemini could help surface merchant options, but I did not get the option to complete an agentic commerce transaction. The merchants it listed did not expose a compatible Universal Commerce Protocol (UCP) path in that experience. That is a report of one session, not a universal verdict. Integrations can change quickly.

Why Amazon was absent

The timing supplies a direct answer. On Sunday night, September 20 — days before this shopping session — Amazon began blocking Muse from shopping on Amazon.com. Users were shown a warning that access by an "unauthorized AI agent" violated Amazon's Conditions of Use. Amazon said Meta had not notified it, Muse did not identify itself while browsing, and the system appeared able to capture and store customer credentials. Meta has separately said Muse cannot see users' passwords or payment methods and uses secure storage.

Two strategies arrived almost at once. Amazon had asked Meta to remove Amazon from the Muse experience; Meta declined. Shopify moved in the opposite direction: CEO Tobi Lütke announced a deep partnership to enable Muse checkout through Shop Pay across Shopify stores.

This is part of a larger contest over who controls the customer relationship. Amazon has also sued Perplexity over Comet and moved to block shopping agents from Google and OpenAI. The question is no longer whether agentic commerce is arriving. It is which merchants will permit it, under what identity, consent, credential, and commercial rules.

The winning ecosystem will make participation explicit and the path from recommendation to resolution trustworthy.

Sources: GeekWire, "Amazon blocks Meta's Muse AI assistant"; PYMNTS, "Shopify Brings Shop Pay Checkout to Meta's Muse AI Agent."

The next experiments

The real test continues after "buy."

The transaction created a useful research agenda. I want to see how the system behaves across the whole payment and post-purchase lifecycle, without manufacturing a false dispute or creating unnecessary risk.

  1. Track the authorization drop-off

    Check the account on September 29, record whether the $25 pending line is gone, and distinguish virtual-card expiry from the issuer's actual release time.

  2. Map the human support path

    Identify who responds when something goes wrong: the merchant, the agent platform, Stripe, Shopify where relevant, or the card issuer — and how a person reaches each one.

  3. Observe a legitimate dispute or refund

    If a real problem ever occurs, document how evidence, provisional credit, merchant communication, and resolution work. Do not create a false dispute just to test the machinery.

  4. Compare compatible and incompatible merchants

    Run the same shopping intent through a merchant with an agentic checkout path and one without it. Measure the difference in price clarity, effort, approval, and recovery.

  5. Track the Amazon boundary

    See whether the block changes, whether Amazon results return, and whether the interface can label unavailable or manually transactable listings without implying agentic checkout.

A sign of things to come

Payments and shopping are entering an unusually exciting period. If this develops well, consumers will be able to express intent in ordinary language, compare genuinely relevant options, understand the full cost, approve a transaction with confidence, and get help when the outcome is not right.

The guardrails are not side features. Clear authorization states, evidence-based explanations, privacy controls, visible merchant identity, human escalation, and unambiguous approval are what will turn novelty into trust.

This first experience was not perfect, but it was compelling. It saved time, preserved my agency, surfaced lessons quickly, and made the future of consumer payments feel tangible. If the ecosystem gets the details right, this is a sign of things to come.