Agentic commerce: the reinvention of e-commerce shopping

We are rapidly moving towards a fully connected retail world, where AI can anticipate customer needs, negotiate purchases, and execute transactions fully aligned with customer intent, according to a recent McKinsey study (October 2025).

We are witnessing a new way of thinking about e-commerce, not merely an evolution, but a true paradigm shift, in which the entire customer shopping experience will be reinvented.

In this new paradigm, the boundaries between technology, platforms, and services are dissolved, merging into a single experience within a seamless flow of highly personalised interactions. Customer journeys become faster, reducing or even eliminating friction altogether.

We are moving from a product search and decision-making process entirely carried out by humans to an agentic-driven process, where the customer simply states their objective and the agent executes the entire process. AI handles discovery, filtering, and final decision-making, selecting the product best suited to the customer’s goal. The shopping experience is increasingly delegated to agents that communicate and negotiate with one another.

This new reality will extend across all industries: retail, banking, insurance, telecommunications, services, transport, and more. E-commerce platform providers and payment system providers will also play a crucial role in this new purchasing process, with integrated communication between agents and their systems. Some providers are already making these capabilities available on their platforms.

Agentic Commerce: Communication and Execution Between Agents

Companies such as Anthropic, OpenAI, and Google have already introduced communication protocols between agents, retailers, and payment providers, enabling agents to access external tools and data while communicating and negotiating with one another. But how does this work in practice?

Through protocols such as Model Context Protocol (MCP), Agentic Commerce Protocol (ACP), Agent-to-Agent (A2A) Protocol, and Agent Payments Protocol (AP2), it becomes possible to create fully autonomous agents that automate the purchasing process end-to-end.

Put simply, the Model Context Protocol (MCP) acts as an adapter connecting AI to external systems or data, allowing AI agents to initiate the purchasing process by accessing online store catalogues, reading product information, checking availability, prices, and delivery dates.

An extremely important aspect of this process is hyper-personalisation. The research agent automatically analyses and interprets the user profile based on preferences, characteristics, demographics, historical behaviour, and real-time information. For example, if a user searches for a cream suitable for sensitive skin, the agent incorporates this information into both current and future searches and establishes product selection criteria accordingly.

By mentioning “sensitive skin”, we are referring not only to explicit data, but also to implicit data such as browsing history and product interaction analysis (time spent analysing product pages, clicks, scrolling behaviour, comparisons, products viewed but not purchased, and more). Finally, there is transactional history, which is more factual and concrete, including average purchase frequency, average basket value, favourite categories, and preferred brands.

Furthermore, using this skincare example, the agent may also incorporate highly relevant information for future product suggestions. If the user has previously shown sensitivity to a product containing a high percentage of ingredient X, the agent will take biological compatibility into account during future searches. It may also adapt product recommendations according to seasonality and changing skincare needs throughout the year.

This happens in real time and dramatically elevates the customer experience in terms of personalisation.

In practice, the agent creates a list of permitted products and a list of prohibited products regarding ingredients and skincare compounds, continuously updating these lists over time based on customer feedback, product ratings, written or spoken feedback in chats or other channels, and similar interactions.

There is also a strong, though often overlooked, behavioural component, as agents learn whether the user’s decision-making style is more rational (price versus benefit), emotional (brand affinity), impulsive, or analytical.

External data sources (where authorised), such as global product reviews, dermatological data, and market trends, can also be combined with the customer profile to refine recommendation models further.

After the research phase, the Agentic Commerce Protocol (ACP) enables agents to negotiate automatically with one another, compare prices across different stores, and ultimately complete the purchase. This is the transaction stage itself. Different access levels, permissions, and human validation requirements may exist depending on the user’s preferences and the implementation choices made by the online retailer.

Some implementations of these protocols already exist in certain countries, although still at a very early stage, raising concerns related to GDPR, the EU AI Act, compliance, and information security.

These AI agent-based commerce protocols are built upon recommendation models that have been used for years in machine learning, based on customer behaviour patterns and recommendation systems. They now also use embeddings that help understand preferences (for example, natural skincare or clean skincare) with the ability to infer and contextualise information automatically.

To ensure context is continuously updated and retained over time, a memory layer stores all information and remains in constant learning, refinement, and optimisation.

For customers, this process will only succeed if they trust what the agent presents (confidence in the agent’s competence), perceive transparency (clear explanations behind product recommendations), and can correct the agent to refine future searches whenever necessary. This is essential for user adoption; otherwise, the customer experience may become negative and discourage future use.

OpenAI launched Instant Checkout in October 2025, promising in-ChatGPT purchasing capabilities. However, the company later announced a reformulation, using ChatGPT primarily for initial product research and analysis while leaving the final transaction to the brand’s online store.

Customer Digital Twin Powered by Generative AI

In the era of agentic commerce, personas evolve into Customer Digital Twins (CDTs): dynamic digital replicas of customers powered by real-time data that enable the storage, analysis, personalisation, prediction, control, and optimisation of behaviours.

The digital twin evolves over time alongside the real customer as contexts and needs change. Through digital twins, companies can hyper-personalise experiences, anticipate customer needs, and simulate risk-free scenarios, both in terms of product offerings and predicting customer churn so proactive action can be taken.

CDTs enable businesses to optimise pricing strategies, campaigns, and offers far more accurately, with real-time adjustments to customer propositions. Furthermore, with Generative AI, CDTs can operate through conversational interfaces, enhancing personalisation even further. The CDT becomes the central asset of the business.

According to the McKinsey article “Enhancing the customer journey with gen AI-powered digital twins”, not only can customer insights be improved and optimised, but also customer sentiment towards the brand and, consequently, customer loyalty. McKinsey estimates that companies using CDTs achieve revenue growth of around 10%.

Boston Consulting Group (BCG) also highlights that agents will manage complete customer journeys rather than isolated interactions.

The CDT acts as the customer’s brain, while Generative AI becomes the executor, enabling automated decision-making, full customer journey management, and autonomous action execution without human intervention. We move from merely understanding customers to simulating and acting on their behalf.

For this system to function effectively, all customer information sources must be integrated: CDPs, e-commerce platforms, CRMs, and external data sources now work together as data layers rather than isolated core systems.

The full implementation and adoption of autonomous shopping systems still depend on several factors, with barriers remaining around data privacy and ethics, information security, systems integration, and customer adoption and usage. Nevertheless, it is clear that we are entering a turning point: a revolution in online shopping as we have always known it.

Ana Candeias – Data & AI Director

Publicado in Marketeer