A few short years, the idea of computers autonomously pursuing goals and taking action on behalf of companies seemed a bit far-fetched. Then again, so did the notion of machines writing poetry, creating original artwork, or producing a deepfake of Neil deGrasse Tyson so realistic, it convinced people he thought the Earth was flat. But here we are, having moved quickly into the realm of AI everything, first with predictive AI, then generative AI, and now agentic AI, a technology that’s as promising as it is complex.
In commerce, agentic AI promises huge benefits to consumers and businesses. It’s a new phase of online experience, one where external AI agents (independent from humans) can interact freely with your website and your site visitors. This will be as transformative as any previous groundbreaking new technology, from the Canadian Pacific Railway to the earliest days of eCommerce itself.
In the 1880s, Canada was faced with the choice to invest early in the Canadian Pacific Railway to connect East and West, or wait and risk losing traffic to U.S. rail lines. They moved first, securing access and influence despite high costs and delayed returns.
The same dynamic applies now. Build agent-preferred commerce infrastructure before others do, and your site becomes the reliable default for external AI agents. If you wait, competitors are more likely to become those agents’ go-to destination, capturing visibility and automated demand ahead of you.
What Is Agentic AI and Why It Matters Now
Agentic AI is best defined by what it requires, versus naming a single technology. To be agentic, an AI system can:
- Perceive information on websites like your product catalog, pricing, and availability in structured formats.
- Infer and reason about the information by, for example, comparing options, evaluating specifications, and assessing fit for their requirements.
- Act based on assigned tasks by initiating quotes, configuring products, and placing orders.
- Learn and improve from each interaction which allows agents to adapt their approach based on successful interactions and outcomes.
According to PwC, agentic AI is currently being adopted by businesses for targeted use cases in customer service, compliance, software development, and drug discovery. In fact, agents are increasing productivity and boosting software launch times by 50%. In commerce, Agentic AI is great for streamlining and scaling simple repeatable tasks like predicting inventory shortages, forecasting and monitoring stock needs, negotiating supplier contracts, and automating quality control processes.
Many organizations are adopting AI rapidly, building their own agents to interact with your site. But these agents will fail if your data and design aren’t structured to support them.
Preparing your websites for external AI agents starts with recognizing that first-generation AI agents often don’t work perfectly. By creating a “digital doorway” for agents, you allow them to access and use site information effectively.
Making Your Website Discoverable and Usable by AI Agents
The core idea of an agentic “digital doorway” is that you’re creating an environment that lets external AI agents access your structured and unstructured data. When an agent arrives at your site, it perceives whatever you present—things like product information, services, images, and metadata. It then figures out what it can do in line with its assigned task. You can help the agent be successful at whatever task it arrived to perform by properly structuring:
- Product data: category taxonomy, specs/attributes, variants, pricing, images, documents, compatibility info.
- Availability and commerce data: inventory, lead times, order status, quotes, invoices.
- Service and support data: policies, returns, warranties, FAQs, knowledge base.
- Action model data: structured definitions of what agents can do on your platform. Specifies required inputs, permissions, and expected outputs. Also maps business logic to structured workflows.
Because so many first-generation agents won’t know what you sell or what your site is optimized to let them do, they need a clear entry point that spells it out for them. This is the purpose of the “doorway.” It’s an agent directory that includes a product data model that structures categories, specifications, and the attributes that differentiate SKUs.
The doorway also includes an action model that reveals the tasks your system supports—inputs, outputs, completion criteria, permissions, and fallbacks. By investing in data clarity, standardization, and transparent action definitions up front, you’re creating an infrastructure helps external agents correctly perceive and act on your data. According to a recent MIT study, 95% of AI pilot programs fail. If you want to succeed, you need to design your systems in a way that supports and facilitates external agent interaction.
The Building Blocks: Product Data Models and Action Models
To make your site legible to external agents, you need two models that work together. The product data model describes what you sell in a way machines can understand, and the action model explains what tasks your system supports and how to perform them reliably.
Product Data Model
The product data model is the foundation for how you describe what you sell. For example, take a grocery store. You begin at the highest level with broad sections like dairy or produce, then move into sub-sections like fruits vs. vegetables, then into more specific types like berries, citrus, and apples.
You will eventually reach the terminal or “leaf” nodes where actual products are placed. At that level, you define an attribution schema that lists the features and specifications whose values let you tell any two SKUs apart. You also add metadata about those attributes so systems know how to handle them, whether a field is numeric, enumerated, text, or boolean.
Many companies have already done much of this when they implemented a PIM or MDM, which is why strengthening the product data model is often the fastest way to improve site search, merchandising, and personalization.
Action Model
Your action model clarifies what your site is designed to let agents do. It should focus on common B2B tasks like requesting a quote, checking availability, configuring a product, or placing an order. For each task, describe the inputs needed, the outputs produced, the signal that tells you it is complete, and define the permissions or restrictions that apply.
The action model also defines what should happen when a task fails (e.g., routing to a human, prompting a chatbot, etc). This helps external agents reliably interact with your system and it also benefits human users. Once the actions are explicit and concrete, they become a strong guide for shaping navigation and workflows in account and order management.
Introducing the Agent Directory
An agent directory connects the product data model with the action model—what you sell with what can be done on your website. It functions like the directory at the entrance of a mall or large department store. When a new AI agent arrives, the directory provides a clear map of what’s is available and the context needed to navigate. By defining this, you replace trial-and-error parsing or scraping with an upfront guide that orients agents.
This is what helps agents interpret your catalog accurately, including categories and the attributes that distinguish products. It also points them to the action model so they understand what your site is optimized to let them do.
Discoverability increases when agents can find categories, attributes, and actions in one place rather than inferring structure from scattered pages. Transparency improves when capabilities and limits are stated up front, including error handling and escalation paths. Standardization follows when categories, attributes, and actions are encoded consistently in machine-readable forms, which reduces failure rates and makes external interaction more reliable over time.
Practical Steps to Build an Agent-Preferred Architecture
You can help external agents succeed by strengthening the foundations you already have—and making those foundations explicit and machine-readable. Here’s the high-level roadmap (and you can download this checklistorwatch our masterclass to go deeper):
- Start with your product data model. Make sure your taxonomy, attributes, and metadata are accurate and complete—especially in your highest-value categories.
- Prioritize what matters. Focus first on the categories and journeys that drive the most revenue or traffic.
- Map key user actions. Identify the tasks visitors rely on most (availability checks, quotes, configuration, ordering, etc.).
- Define clear action models. Spell out the inputs, outputs, permissions, completion signals, and fallback paths.
- Centralize it in an agent directory. Give external agents one place to discover your categories, attributes, and supported actions.
- Iterate as agents learn. Monitor outcomes and refine definitions to improve success rates over time.
The Hidden Cost of Inaction
All of this may seem a bit overwhelming and, while it can be tempting to defer the focus on agentic AI to next year, there is a high cost to waiting. The market for agentic is incredibly dynamic, but that won’t be the case for long. It’s likely that if you don’t start to build an experience that makes your website more attractive to agents, your competitors will—costing you a lot in the long run.
The agentic wave is coming and, much like the Canadian Pacific Railway project, the choice to embrace it is now. Consider that the U.S. had already completed its first transcontinental railroad in 1869. If the Canadian Pacific had not moved forward with that investment, it might have led to a shift in opinion, perhaps even a stronger partnership with the U.S. instead of Canada.
From Human-First to Agent-Preferred Commerce
Consider that over 40% of agentic AI projects will be canceled by 2027, according to Gartner. But if you make an investment in the infrastructure and the design of your systems, you have a real opportunity to benefit from the external agents that make their way to your website.
Agent-preferred commerce helps you gain visibility by making your catalog and capabilities easy for agents to find and understand at a single entry point. It also drives efficiency. A strong product data model improves site search and personalization for humans while reducing mis-selections and returns for agents.
Explicit action models cut friction in common B2B tasks and provide fallbacks to a human when needed. The result is faster cycles, fewer errors, and better outcomes across customer and supplier workflows.
All of this amounts to a huge competitive advantage. In a market where most first-generation agents will fail on poorly structured sites, companies that standardize early become the default choice for external agents. As agents learn which domains are dependable, they will return to those properties more often. The early builders will capture more automated demand, lower operating costs through cleaner data and clearer actions, and widen the gap over competitors who wait.