Today, fully autonomous agentic AI agents for consumers are still emerging, but the software that powers these agents is popping up in more places.

To enable these kinds of purchases, Mastercard in April announced its Agent Pay technology, which allows verified AI shopping agents to make transactions on behalf of consumers and businesses.

An agentic commerce agent works by starting with a user request, or prompt. Unlike traditional AI systems, which give static, one-off responses, modern agents are built to interpret these requests, consider context, adapt on the fly and decide how to move forward.

The input can also start a conversation to kick-start a goal-oriented action. For instance, if you make a very broad request like “I need a new shirt,” the AI retail agent can reply asking for more specifics, such as whether you’d like a pattern or a certain type of fabric, or for what kind of occasion the shirt would be worn.

Agents then automate research and product discovery. Instead of simply searching a single website, the agent can search across multiple e-commerce platforms, access and analyze product specifications, reviews and ratings, compare prices in real time and evaluate shipping times, return policies and other logistical details.

From there, the agent doesn’t just present a list of options, but can actively reason through them based on the user’s initial parameters and its own understanding of what constitutes a good deal or best fit.

There are three key components of AI agents that help them make decisions:

• Memory: Newer AI agents can remember a user’s preferences, sizes and past purchases, making it easier to identify an ideal response to a shopping query.
• Tools: Agents have access to APIs and external databases that let them find new information and take action.
• Reasoning: An agent can break down a complex request into structured, actionable steps, instead of just offering a single, static response. It can determine which tools it needs to access, what information it needs to gather and how it can achieve its goals.

Agentic commerce can be safe with the right guardrails, permissions and technologies in place. However, this new form of commerce certainly introduces new challenges for retailers, tech companies and consumers, particularly in those cases where purchases don’t go as planned.

For example, how would a refund or chargeback work if an agent, not a person, is making a purchase? Or who would be responsible if something goes wrong — the consumer, the company that created the bot, the retailer or perhaps someone else?

Clear and easy-to-set user permissions will be vital in this new environment, as the more autonomy an AI agent has, the more important it will be to set clear limits and guidelines.

Transparency will also be critical for building trust, with users receiving a clear explanation of why an agent took certain actions. That transparency will be especially helpful if cybercriminals attempt to trick an AI agent by triggering purchases using misleading signals or mimicking legitimate offers.

Standards and frameworks to govern these agents will be needed, too, so that both users and merchants can know how transactions can be resolved when problems do arise and consumers can trust that their purchasing and payments data will be protected.

Agentic commerce is expected to quickly expand to many parts of e-commerce within the next year, with the possibility that some users will use a chatbot to make most of their future purchases.

These AI agents are expected to become more complex and specialized, with the potential for more automation for users if they want it. In addition, there could be more multi-agent systems, in which several of these AI bots work in concert to complete more complicated tasks that a single bot couldn’t do on its own.