• Deutsch
  • English
  • ChatGPT Wallet is designed to let agents based on Artificial Intelligence (AI) select purchases and pay after you approve them. At the same time, Amazon, Meta, and specialized providers are working to delegate product research, phone calls, and marketing tasks to such systems. The first users are mainly paying subscribers in the United States, but the broader shift is from chatbots that offer advice to digital contractors that take action.

    What do AI agents take over?

    An AI agent does more than answer a question: it carries out several connected steps on a person’s behalf. When shopping, it may research products, compare options, assemble an order, and provide the merchant with the information required to fulfill it. The ChatGPT Wallet announced by OpenAI adds payment to that sequence.

    A practical everyday example is finding a product that meets a defined set of requirements. You tell the agent what you need, then review the items, total, and selected card before confirming the purchase. According to the reports, the initial Wallet release is limited to a small number of adult ChatGPT Pro subscribers in the US with an eligible US-issued card.

    Meta is expanding agents in a different direction. Its Muse agent can call US businesses for paying subscribers, according to a report on Muse’s calling feature. Examples in the source include dealing with cable companies, motor vehicle offices, and restaurants, as well as waiting on hold and working through complicated voice menus.

    Workplace routines are another target. Nectar’s Creator Agent is meant to find suitable content creators for brands, review previous partnerships, send personalized messages, and track campaign performance. According to TechCrunch’s report on Nectar, the system searches more than 300 million creator profiles; that provider claim has not been independently verified.

    How does ChatGPT Wallet work?

    You add an eligible payment card once and then ask a dot, OpenAI’s name for the relevant agent, to help with a purchase. Before checkout, the system displays the selected items, total cost, and card. OpenAI says each purchase requires a biometric security check such as facial recognition or a fingerprint.

    The agent and merchant are not supposed to receive your actual card number. A payment token, meaning a substitute number with its own security credentials, is used instead. The merchant still receives the information needed to complete the order, including your shipping address and contact details.

    OpenAI says existing card rewards and customary protections for valid disputes will remain available. A Wallet report from The Verge also describes the longer-term plan: users may eventually preapprove purchases by setting spending limits, product categories, or time periods. The agent would then no longer need separate approval for every transaction covered by those rules.

    The sources do not describe the launch scope in exactly the same way. OpenAI’s short announcement mentions ChatGPT Work, Codex, and dots, while the accompanying reports focus on a limited release through dots for ChatGPT Pro subscribers. Card eligibility is also worded differently, although the accounts agree that qualifying US-issued cards are required.

    This is not OpenAI’s first attempt at AI shopping. The Decoder describes Wallet as a second attempt after an earlier direct checkout feature inside chat was dropped because it lacked flexibility. The sources provide no separate price for Wallet, but access initially requires a paid ChatGPT Pro subscription.

    Where do control and privacy fit?

    Approval before each purchase is currently the main control. You can see what the agent plans to buy and confirm the transaction with a biometric check. This reduces the chance of unnoticed purchases, but it cannot prevent an agent from choosing the wrong product or misunderstanding your instructions.

    Tokenization protects the card number, but it does not make agent-assisted shopping free of personal data. A delivery requires at least contact and shipping information, while the agent processes your preferences and the surrounding purchase context. The sources do not explain how long those details are retained, whether they are used for other purposes, or what settings would apply outside the US.

    Merchants must also cooperate. Some websites restrict or block AI agents, and reports say Amazon has already blocked Muse. OpenAI acknowledges that websites must avoid accidentally excluding legitimate agents, while payment and risk systems need to recognize purchases made by them.

    Acceptance cannot be assumed on the phone either. According to the Muse report, some businesses hang up as soon as they detect AI. Meta temporarily tested human phone operators, while people reportedly handled one in four calls in Google’s earlier Duplex system. Increasingly human-sounding synthetic voices also make it harder to know whether a person or software is speaking.

    Pros and Cons of AI agents that take action

    Pros:

    • Less routine work – Agents can handle product research, hold queues, and repetitive contact tasks.
    • Visible purchase control – ChatGPT Wallet shows the items, amount, and card before currently requiring individual approval.
    • Protected card details – OpenAI says payment tokens prevent the agent and merchant from receiving the actual card number.
    • Greater workplace scale – Marketing teams can review large pools of profiles and coordinate outreach without managing every message manually.

    Cons:

    • Incorrect selections – A technically valid purchase may still involve the wrong product, model, or offer.
    • Additional data trails – Preferences, addresses, contacts, and transaction context become part of an agent-assisted process.
    • Limited acceptance – Merchants may block shopping agents, while businesses can reject automated calls.
    • Unclear costs and availability – Several features require paid subscriptions, but exact prices and international launch dates are missing.

    What does this mean in practice?

    For a beginner, the most sensible first step is a narrowly defined assignment. You might let an agent find products that meet explicit criteria, while personally checking the description, seller, total price, and shipping details before approving payment. For calling features, a simple information request is a safer starting point than a conversation with significant financial or contractual consequences.

    Advanced users can structure instructions more precisely by specifying a maximum budget, permitted categories, excluded merchants, and a limited time window. These resemble the parameters OpenAI envisions for future preapproved purchases. Until that delegation is available and adequately tested, approving each individual transaction remains the more transparent option.

    Amazon is pursuing a different position in the shopping chain. Its planned World Product Graph would collect information about products sold across the internet, including goods Amazon does not sell. That could encourage both people and AI agents to use Amazon as the starting point for research.

    Whether the project will actually launch remains uncertain, according to the source. An internal document reportedly acknowledges that many answers from Amazon’s Rufus assistant are wrong or incomplete and that running searches through AI is expensive. More product data is supposed to improve answers and allow the use of cheaper models, but those benefits have not yet been demonstrated.

    Nectar shows how far delegation may extend in marketing. Its agent is designed not only to rank profiles but also to send messages in a brand’s voice, create custom landing pages, and help manage contracts, product gifts, and shipping details. The reported expectation that US creator advertising spending will reach $44 billion in 2026 comes from the Interactive Advertising Bureau and remains a forecast rather than an established outcome.

    For Switzerland, immediate availability is limited: ChatGPT Wallet and the described Muse calls are launching only in the US. The sources provide no Swiss launch date, no details about supported national languages, and no account of how data would be handled under Swiss or European requirements. The technical protection of card numbers therefore does not establish that the entire process meets Swiss privacy expectations.

    AI agents are moving from advisory chatbots toward systems that spend money, call companies, and manage business relationships. Their practical value is strongest for clearly bounded routine tasks, while individual approvals and strict limits remain the more cautious arrangement. The unresolved risk is not only one dramatic failure, but a sequence of small and plausible actions whose data flows and consequences are difficult to see in full.

    Sources

    AI-FunghiAI-Funghi

    Image: Mikhail Nilov via Pexels

    © 2024 - 2026 ai-funghi.com | All Rights Reserved | Impressum | Datenschutz