How AI purchases work and what are main challenges for payment industry
Artificial Intelligence burst into our daily lives rapidly, and even faster into our professional ones. People now discuss daily events with chatbots, ask for weekend plans, and trust them with work tasks. Many companies have already integrated ready-made AI solutions into their workflows or are developing their own, set to implement them soon — all to boost speed and productivity.
The payment industry is no exception; in fact, it’s adopting artificial intelligence faster than many others, automating transaction monitoring, fraud protection, and analytics. But now, AI is moving beyond just analyzing past purchases — it’s starting to make the purchases themselves. This brings both significant advantages and some concerning drawbacks.
Today, we’ll explore why everyone is talking about Agentic Payments and what we can expect in practice.
What Is Agentic Commerce?
Let’s start from the beginning and clarify what this innovation is all about. Regular chatbots simply suggest options and reason through, but agentic AI goes all the way — it selects, pays, and confirms product delivery independently. You state what you need, and the agent finds the best option and completes the purchase in seconds — no intermediate steps required.
At the end of September, OpenAI brought this solution to the world with its Instant Checkout. It’s available for ChatGPT users in the U.S., initially with sellers on Etsy, and later expanding to a million Shopify merchants. The process is simple: a buyer asks the chatbot about a product, receives a carousel of options with a “Buy” button, clicks it, confirms the delivery address and card details — and the order is placed directly within ChatGPT. The key difference is that there’s no need to visit each website or platform separately to explore options, which significantly simplifies shopping and selection process and reduces shopping time.
Under the hood, this is powered by the Agentic Commerce Protocol, developed by OpenAI in collaboration with Stripe. An important note: ChatGPT acts only as an intermediary, another channel for purchases, not the seller. The seller’s business, as before, calculates taxes, displays the final amount, collects tokens, and the payment processor deducts the funds — while the merchant retains full control over their payments and customers. A fee is charged on successful purchases, but the service is free for users and does not affect the final price of the product.
Despite ChatGPT’s clear dominance, competitors are also trying to jump on this bandwagon. Perplexity launched Buy with Pro at the end of 2024, and in May, Google introduced its Agent Payments Protocol, which will process purchases through AI Mode based on Gemini, with price tracking and even virtual try-ons.
Google has thoroughly addressed the issue of purchase consent, which is the main challenge for most developers of this technology. Since chatbots can make mistakes or misinterpret requests, there is a risk of generating a large number of unauthorized purchases that users may want to dispute. The mechanism is called Mandate, literally meaning “instruction.” Intent Mandate records the request (e.g., “find me sneakers”), while Cart Mandate finally confirms the specific price and product. For deferred tasks, such as “buy concert tickets as soon as they go on sale,” the buyer pre-signs an Intent Mandate with limits and conditions. When these conditions are met, the agent automatically generates a Cart Mandate and completes the purchase.
All of this means that users will no longer interact with the seller’s website, bypassing many functional elements that are so familiar today. While UI/UX and payment paths are currently critical, in the future, they may no longer be necessary. It’s quite possible that sellers of physical goods will be able to connect their inventory directly to ChatGPT and sell exclusively there. But this raises new challenges: how can merchant stand out among competitors offering the same products? What bonuses can they offer if everyone is on same level? These questions will likely become very pressing soon.
Regulatory Concerns of Agentic Payments
Beyond user experience, agentic commerce significantly impacts fraud control systems. Automated transactions are formally authorized, but in reality, the agent doesn’t fully understand the user’s intentions. This gap risks a surge in first-party fraud.
Global payment regulations and infrastructure are designed for transactions initiated by humans. When an AI agent enters into a contract, questions arise about its validity. Under agency law, an agent must act with actual authority — explicit or implied. In disputes, the key question is: did the agent act within the given authority or not?
Additionally, agentic payments require the exchange of personal and financial data between agents and parties, immediately invoking all privacy laws. These laws, in turn, require transparency: which third parties will have access to personal data and why? Even without agents, this is already a serious stumbling block in the complex payment ecosystem. Agentic transactions will exacerbate the problem many times over.
Data minimization and protection principles are also critical. Organizations will need to impose restrictions on agents to ensure the proportional use of customer data and its disclosure only with appropriate security guarantees. Experts confirm that no existing regulation directly addresses agentic AI. Currently, the trend is to use the EU AI Act as a foundation with minor variations across countries to avoid a patchwork of dozens of versions. In the U.S., legislators at the federal and state levels are discussing possibilities, but there is no clear direction yet.
In 2023, President Biden issued an executive order on safe AI. Trump revoked it in 2025 and issued his own, eliminating anything considered a barrier to AI development in the U.S. Even in this uncertainty, experts advise taking steps to ensure compliance. The AI Act has already been adopted and will come into full force within three years. In a rapidly changing regulatory landscape, a conservative approach — anticipating likely standards such as human oversight, data protection, and fairness — will help stay ahead and avoid costly disputes. However, quality regulation is still far off, and the technology is already here, meaning we’ll likely have to learn from mistakes made by real people.
Security of Agentic Payments
Beyond convenience and speed, security is crucial in payments. Monoup experts have repeatedly emphasized this in our articles, and AI agents are no exception — in fact, quite the opposite. They handle vast amounts of data and perform automated actions, increasing the risk of leaks or unauthorized access to this data. When a breach occurs, users lose trust and question the security of the entire system.
To function effectively, agents need deep access to consumer data — financial profiles, transaction history, preferences, payment information — to make accurate decisions and act autonomously. Beyond the risks associated with international and national privacy laws, this concentration of sensitive information increases vulnerability to misuse by malicious actors. If millions of people’s card data are stolen from various sources now, what could happen if access is gained to just one AI?
Industry players have already begun developing protection protocols. OpenAI and Stripe created the Agentic Commerce Protocol with a focus on security. The protocol allows agents to transmit payment data to merchants, reducing friction during checkout and helping distinguish legitimate agents from bots.
Visa states that its Trusted Agent Protocol focuses on three aspects: signaling the agent’s intent, recognizing the consumer behind the agent, and transmitting payment data. The goal is to provide merchants with clarity without overhauling systems. Both companies claim their protocols are designed to support other standards, not compete with them. Visa is aligning its Trusted Agent Protocol with OpenAI’s Agentic Commerce Protocol and Coinbase’s x402 standard.
The foundation of Visa and Mastercard’s systems is Web Bot Auth technology from Cloudflare, developed with Microsoft, Shopify, Checkout.com, Worldpay, and Adyen. Cloudflare created it to support agentic commerce at scale, allowing agents to operate across millions of merchants globally. Amex also plans to adopt the standard.
It’s important to note that agentic payments carry specific security risks. Fraud, including replay attacks or fake mandates, can occur if protocols are poorly implemented; compliance gaps arise because regional rules assume human-managed payments. All of this creates a trap for agentic payments and numerous ways to compromise the technology.
The challenge isn’t just creating capable purchasing agents but ensuring that the entire lifecycle is clear to users and regulators. Organizations must design agents as transaction participants, not just conversational interfaces. Restrictions are needed — transaction limits, revocation APIs, multi-signature approvals. Control is also crucial — logs, dashboards, and alerts for agent spending.
Systems must be built with tracking mechanisms from the start. This means recording not only the agent’s actions but also the prompts, decisions, internal state changes, intermediate reasoning, and source data that led to such behavior. These systems are necessary for audits, root cause analysis, regulatory compliance, and reviews. Organizations must regularly assess performance to ensure agents remain aligned with their purpose. This means there will be no anonymous dialogue with AI, as every situation must be clearly verifiable and traceable.
Additionally, even well-designed agents can fail or be compromised. There needs to be a plan in case an agent malfunctions or behaves unexpectedly. To ensure the success of agentic AI pilot projects, organizations must assess their current level of skills, knowledge, and resources — including AI security engineering, security testing, threat modeling, and skills for governance, compliance, and risk management.
For many companies, this will require not only new technologies but also new business models. Marketplaces, loyalty platforms, and intermediaries that thrive on human engagement may face different questions: Should they build their own agents? Welcome or block agent traffic? How to monetize when algorithms, not people, make purchasing decisions?
We’ll likely find out soon, but for now, the future of the customer journey is uncertain. Some might say it’s still far off — after all, who would trust ChatGPT with their card details or financial habits? But in practice, AI already has access to vast amounts of even more personal data, and the number of people willing to use it for purchases is only growing. ChatGPT has 800 million weekly users, and alongside it, other agents are also developing.
