From API-first to agent-first commerce

AI agents are becoming a new class of commerce-platform user. What does that change for B2B businesses, and what remains essential?

Part of my work at Intershop is exploring what AI changes in B2B commerce, not only in individual product features, but in the platform itself.

Working with our teams on buyer copilots, merchant agents and AI-assisted implementation has made one pattern hard to ignore: AI agents are becoming a new class of platform user.

That does not mean people, storefronts or enterprise applications disappear. It also does not make the architectural principles of the past decade obsolete. Agent-first commerce builds on API-first and composable commerce. It extends them for a world in which people, business systems and AI agents work with the same commerce capabilities.

The strategic question is therefore not whether every business needs another AI feature. It is whether the commerce platform can support agents without compromising the prices, permissions, processes and control on which complex B2B relationships depend.

1. What does agent-first commerce mean?

Agent-first does not mean agent-only. It means treating agents as first-class users of the platform, with clear capabilities, context and boundaries.

I see three important roles:

• Buyer agents help customers discover products, compare alternatives and prepare purchases.

• Merchant agents support commerce, sales and service teams with product information, customer requests and operational tasks.

• Developer agents accelerate implementation, integration and customer-specific extensions.

The buyer and merchant roles are not purely theoretical. Intershop already offers Copilot for Buyers, Copilot for Merchants and specialized agents for product content, localization and business intelligence. This gives us practical experience with how people and role-specific agents work together in B2B commerce.

These agents have different goals, but they share one requirement: they need to understand not only what the platform can technically do, but what each action means for the business.

An endpoint may allow an order to be created. That tells an agent nothing about whether the buyer is authorized, the contract price is valid, a budget has been exceeded, or human approval is required. Agent-first design makes this context explicit and the result verifiable.

2. Do API-first and composable still matter?

Absolutely. API-first made commerce capabilities accessible beyond a single storefront. Composable commerce gave companies the freedom to combine those capabilities with different experiences, services and enterprise systems.

Agent-first commerce does not replace this foundation. It adds another layer of usability:

• APIs provide access.

• Composable architecture provides flexibility.

• Agent-first design adds understanding and controlled action.

For many B2B companies, the strongest model is both precomposed and composable. They need proven commerce capabilities and a usable starting point, with packaged business capabilities (PBCs) that provide ready-to-use functionality – not just a bag of APIs and a large integration project. At the same time, they need clear boundaries where partners and developer agents can integrate, extend or replace individual capabilities.

This balance matters because the commerce platform does not operate alone. ERP, CRM, PIM, warehouse and procurement systems remain responsible for their respective domains. The commerce platform brings the relevant information together in a commercial context and provides a controlled way for people and agents to act on it.

3. Why is complex B2B the real test?

Product discovery and content generation are useful AI scenarios. The real test begins when an agent contributes to a commercially binding transaction.

Imagine an order request arriving by phone, email, PDF or spreadsheet. A merchant agent can identify products and quantities, match the request to the correct account and prepare questions where information is missing. A buyer agent may approach the same journey from the other side, starting with a business need, a budget and a set of delivery constraints.

Both need the platform to apply the correct customer context, assortment, contract price, permission and approval rule. A clear request may proceed. An unusual quantity, a new delivery address or a price deviation may need human review.

Developer agents are equally important in this scenario. Complex B2B rarely comes without integration and adaptation. A developer agent can help connect the customer’s ERP or procurement system, implement a specific approval process or create an extension at a defined boundary. The objective is faster differentiation without turning the product core into a customer-specific fork.

This is why domain knowledge matters so much. A mature commerce platform contains years of accumulated knowledge in its pricing logic, state transitions, validation, error handling, integrations and edge cases. AI can make that knowledge more accessible, but it cannot safely guess business behavior that has never been made explicit.

4. Why does more autonomy require more control?

As agents move from recommending actions to executing them, governance becomes part of the experience.

A copilot acting for a signed-in user should remain within that user’s permissions. A more autonomous agent needs its own identity and an explicitly delegated mandate. For every sensitive action, the platform should be able to answer:

• Who initiated the work?

• Which agent executed it?

• On whose behalf did it act?

• What did its mandate allow?

• Which policy was applied?

• Was human approval required?

Logging an incorrect action after it happened is not enough. Policy needs to apply before execution. Depending on the context, the platform should allow an action, deny it or request approval.

This is not governance for governance’s sake. It is what allows a business to give agents more responsibility without giving up control.

5. What should business leaders look for?

Agent-first commerce should not be evaluated by the quality of a chatbot demo. I would ask six questions:

1. Does the agent improve a measurable business outcome?

Start with customer experience, operational effort, conversion, cycle time or implementation speed. If the outcome remains vague, the technology probably will too.

2. Does every channel use the same business rules?

An agent should not invent its own prices, permissions or order logic. It should work with the same commercial foundation as storefronts, portals and enterprise applications.

If an agent offers the public list price instead of the customer’s contract price, that is not a small conversational mistake. It can become a margin problem, a contractual problem and a damaged customer relationship.

3. Does it fit the existing enterprise landscape?

The platform should work with ERP, CRM, PIM and procurement systems without creating a second source of truth.

For example, an agent should not promise immediate delivery because a search index looks current while the ERP already shows that the available stock has been allocated elsewhere.

4. Is the agent’s authority explicit and limited?

Businesses need identity, delegated authority, policy, approval and auditability before they can responsibly increase autonomy.

A buyer agent preparing a routine replenishment order may stay within an approved budget. The same agent should not submit an unusually large purchase or introduce a new delivery address without the required approval.

5. Can the business differentiate without sacrificing updateability?

Developer agents need stable APIs, documented extension points, realistic test environments and clear boundaries around the product core.

6. Can a successful proof become a production solution?

The relevant measure is not time-to-demo. It is time-to-verified-production, including security, integration, testing, operations and ongoing lifecycle cost.

Three takeaways

1. Agent-first extends API-first and composable commerce. It adds the context and controls that agents need to act reliably.

2. Complex B2B domain knowledge becomes more valuable, not less. Agents need proven commercial behavior, especially when recommendations become transactions.

3. The winning measure is trusted business impact. Faster experiments matter, but only when they lead to maintainable, governed and measurable outcomes.

For me, this is the natural next step for composable B2B commerce: open enough for new experiences and agents, but dependable enough for real business relationships.

Continue exploring

• Next in this series: How we build with coding agents at Intershop

• Explore Intershop's AI copilots & agents

• Download the Intershop B2B commerce playbook

• Explore Intershop analyst reports