Post
JA EN

Don't Build AI Agents—Build Services That AI Agents Choose: A Practical Guide to Agentic Commerce

Posted
By
15 min read
Don't Build AI Agents—Build Services That AI Agents Choose: A Practical Guide to Agentic Commerce
  • Target audience: E-commerce operators, SaaS providers, marketers, and engineers interested in business strategy
  • Prerequisites: Basic understanding of AI trends and web technologies
  • Reading time: 15 minutes

Overview

“We think it’s time to stop rebuilding agents and start building skills instead.” In late 2025, Anthropic engineers Barry Zhang and Mahesh Murag made this statement at the AI Engineer conference1. The company that builds AI agents is telling you not to build agents.

This message addresses a dilemma many businesses face. “Should we build our own AI agent?” is a question many companies struggle with. But there’s a more important question: Is your service ready to be “chosen” by AI agents?

Today, ChatGPT has shopping features, Google’s Gemini app sports a “Buy for me” button, and Perplexity has enabled in-app purchasing through “Buy with Pro.” Morgan Stanley estimates that AI agent-driven e-commerce spending could reach $190 billion to $385 billion in the U.S. alone by 20302. The “primary user” of e-commerce sites is shifting from humans to AI agents.

Just as traditional SEO aimed at being chosen by Google’s search algorithm, the new competitive advantage lies in being chosen, understood, and transacted with by AI agents. This article examines why you shouldn’t build agents, and how to build services that agents choose—backed by the latest case studies and data.

“Don’t Build Agents”—Three Reasons Why

Reason 1: The Companies That Build Agents Are Saying So

Anthropic’s Zhang and Murag make the case clearly. AI agents possess intelligence and general capabilities, but they lack the domain expertise needed for real work1. Hiring a math genius is less useful for filing taxes than hiring an experienced tax accountant. Likewise, building a general-purpose agent from scratch is far less effective than preparing “skills”—structured procedural knowledge—that agents can leverage.

Fortune 100 companies are already focusing on building “skills” that teach agents about organizational best practices and internal software. The agent itself is built by Anthropic or OpenAI. What businesses need to prepare is the knowledge and interfaces that agents need to perform domain-specific tasks.

Reason 2: 40% of Agent Projects Will Be Canceled

In June 2025, Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 20273. Based on polling of 3,412 respondents, the forecast cites escalating costs, unclear business value, and inadequate risk controls as key drivers.

Even more concerning is the “agent washing” problem. Of the thousands of agent vendors on the market, only about 130 have genuine agentic capabilities. Most are simply rebranding existing chatbots and RPA tools, according to Gartner3. With the risk of building agents in-house this high, the rational strategy is to invest in building the “service side”—making your business ready for the agents that platform companies like OpenAI and Google are developing.

Reason 3: OpenAI’s Pivot Proved the Merchant Side Matters

The most compelling evidence comes from the agent-building giant itself—OpenAI.

OpenAI initially pushed for “end-to-end checkout within ChatGPT.” But the complexity of merchant onboarding, ensuring accurate product data, handling multi-item carts, and integrating loyalty programs proved to be insurmountable barriers4. As a result, OpenAI shifted to a “discover in ChatGPT, purchase on the merchant’s site” model.

The implication is clear. No matter how advanced AI agents become, the final transaction depends on merchant-side infrastructure. Even the agent builders themselves admit they can’t succeed without merchants being prepared.

Where to Focus—Become a Supplier

Box CEO Aaron Levie has said that “agents could be the biggest users of the internet” and that “all software is going to have to be built for agents”5. Shopify CEO Tobi Lütke declared that the company is “making every Shopify store agent-ready by default,” making clear that merchants don’t need to build agent integrations themselves6.

Of course, for large platform companies and tech firms whose competitive advantage lies in AI infrastructure, building agents is their core business. But for e-commerce sites, SaaS providers, service businesses—the vast majority of companies—the right move is not to build agents but to become agent “suppliers.”

A restaurant owner shouldn’t build their own Uber Eats. They should optimize their listing on Uber Eats to make ordering as easy as possible. The AI agent era follows the same logic. Becoming the provider of services, data, and products that AI agents use—that’s the strategy for most businesses.

flowchart LR
    subgraph Traditional["Traditional E-commerce"]
        H["Human Consumer"] --> W["Website"]
        W --> P["Purchase"]
    end
    subgraph New["Agentic Commerce"]
        U["Human Consumer"] --> A["AI Agent<br>(ChatGPT, Gemini,<br>Perplexity, etc.)"]
        A --> M["MCP / UCP / ACP<br>Protocol Layer"]
        M --> S["Your Service"]
        S --> O["Purchase Completed"]
    end
    Traditional -.->|Paradigm Shift| New

The Age of AI Shopping—What’s Happening Now

What does it actually look like to be “chosen by AI agents”? From late 2025 through 2026, AI platforms have been rapidly expanding their commerce capabilities.

PlatformKey FeaturesStatus (March 2026)
ChatGPTShopping research, product discoveryTarget, Sephora, Nordstrom integrated4
Google Gemini“Buy for me” button, AI Mode searchRolling out to select U.S. retailers via UCP7
Perplexity“Buy with Pro,” PayPal instant checkoutFree shipping for Pro members8

On these platforms, consumers simply ask “Compare running shoes under $100 for women in their 30s with 4+ star reviews,” and the AI agent searches multiple e-commerce sites, compares products, and makes recommendations. If your product can’t be “found” by the AI agent, you’ll never reach that consumer.

Market estimates underscore the scale of this shift.

  • Morgan Stanley: AI agents will account for 10–20% of U.S. e-commerce spending by 2030 ($190–385 billion)2
  • McKinsey: Up to $1 trillion in U.S. B2C retail by 2030, $3–5 trillion globally9
  • NRF 2026: In a live poll during a Stripe session, 75% of participating retailers said they had implemented or were planning agentic commerce10

A Practical Guide to Being Chosen by AI

Being “chosen” by AI agents requires building up four layers incrementally. You don’t need to do everything at once. For each layer, we show what you can start immediately and what requires medium- to long-term investment.

flowchart TB
    L1["Get Discovered<br>AEO + llms.txt<br>── Start Now ──"]
    L2["Get Understood<br>Structured Data + APIs<br>── Now to Short-term ──"]
    L3["Get Connected<br>MCP<br>── Mid-term (3-6 months) ──"]
    L4["Get Transacted<br>UCP + ACP<br>── Long-term (6-12 months) ──"]
    L1 --> L2 --> L3 --> L4

    style L1 stroke:#4A90D9,stroke-width:2px
    style L2 stroke:#7B68EE,stroke-width:2px
    style L3 stroke:#E6783E,stroke-width:2px
    style L4 stroke:#2ECC71,stroke-width:2px

Get Discovered—AEO (Answer Engine Optimization) and llms.txt

The first step to being chosen by AI agents is being found in the first place.

According to Gartner, traditional search engine query volume will decline by 25% by the end of 2026, shifting to AI chatbots and voice assistants11. When Google AI Overviews appear, organic CTR drops by up to 58%, while sites cited within AI-generated answers see CTR increases of up to 35%11. Sites that aren’t “chosen” lose more traffic than ever, while those that are chosen actually benefit—a widening gap.

AEO (Answer Engine Optimization) is the optimization approach for this new environment12. While traditional SEO aimed at ranking high in search results, AEO aims at being selected as a trusted source when AI generates answers. AEO doesn’t replace SEO—it extends it. Technical SEO fundamentals (fast loading, mobile responsiveness, crawlability) remain prerequisites for AEO as well.

llms.txt is a proposed standard for presenting site information in an LLM-friendly format13. Just as robots.txt tells crawlers about site structure, llms.txt provides structured core information to LLMs. Major tech companies including Anthropic, Cloudflare, and Stripe have adopted it, with over 844,000 sites implementing it as of 2025.

However, the effectiveness of llms.txt is currently limited. None of the major AI platforms have officially stated that they read llms.txt files. In 8 out of 10 surveyed sites, no significant traffic changes were observed after implementation13.

What You Can Do Now

  • Provide structured answers: Organize content in formats AI can easily extract (FAQs, comparison tables, clear heading structures)
  • Strengthen authority signals: Clearly attribute authors, cite sources, improve E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness)
  • Deploy llms.txt: Implementation cost is minimal. Even with limited current effectiveness, it’s a reasonable early investment in future standardization

Get Understood—Making Structured Data and APIs AI-Friendly

Even if AI agents find you, they won’t recommend your products if they can’t “understand” the product information. Humans can read web pages visually, but AI agents consume structured data directly.

Schema.org markup was already important for traditional SEO, but its importance fundamentally changes with agentic commerce. For AI agents to automatically determine “Is this item in stock?”, “What are the shipping terms?”, or “How does the price compare to similar products?”, machine-readable data is essential.

What You Can Do Now Through Short-term

  • Enrich Schema.org markup: Make product information machine-readable with Product, Offer, Review, FAQ schemas
  • Reflect inventory, pricing, and shipping in real time: Outdated information erodes AI agent trust
  • Build product catalog APIs: Expose product data via REST/GraphQL
  • Register with product data feeds: Google Merchant Center, Perplexity Merchant Program8

Get Connected—MCP (Model Context Protocol)

The next step beyond structured data is direct connection with AI agents. MCP (Model Context Protocol) is an open standard for AI tool integration created by Anthropic in late 2024. Just as the USB port standardized device connections, MCP standardizes connections between AI agents and external services14.

E-commerce platforms are leading the way. Shopify offers four types of MCP servers as of March 2026, with live integrations with ChatGPT, Perplexity, and Microsoft Copilot15.

MCP ServerCapabilitiesUse Case
Storefront MCPProduct search, cart management, checkoutAI shopping
Customer Account MCPOrder tracking, returns managementCustomer service
Dev MCPDocumentation, API schemasDeveloper tools
Checkout MCPPayment processing (preview)Payment integration

In March 2026, Shopify activated “Agentic Storefronts” for all eligible merchants, making products from 5.6 million stores discoverable through ChatGPT, Microsoft Copilot, Google AI Mode, and the Gemini app6. Merchants just configure it once in their admin panel—no custom integrations required.

commercetools also launched “Commerce MCP,” making product catalogs, carts, pricing, promotions, and orders accessible to AI agents16. It supports integration with major frameworks including OpenAI Agent SDK, Vercel AI SDK, and LangChain.

MCP’s value extends beyond e-commerce. Real estate property search, travel booking, B2B SaaS feature access, financial services account inquiries—any service can become an AI agent’s “trading partner” by providing an MCP server.

What You Can Do in the Mid-term (3–6 months)

  • If you’re on Shopify or similar platforms: Enable Storefront MCP from your admin panel (no development required)
  • If you run your own e-commerce: Build an MCP server exposing product search, inventory queries, and cart operations
  • Test with AI agents: Verify that Claude, ChatGPT, Gemini, etc. can correctly access your data
  • Design access controls: Define the scope of data exposure, authentication, and rate limits

Get Transacted—UCP and ACP (Payment Protocol Standardization)

Even when AI agents can “discover,” “understand,” and “connect” with your products, the final hurdle is the “transaction.” Two major protocols have emerged to standardize the complex processes of payment, shipping, and returns.

UCP (Universal Commerce Protocol) is an open-source standard announced by Google and Shopify at NRF (National Retail Federation) in January 20267. Over 20 partners including Etsy, Wayfair, Target, and Walmart have endorsed it, standardizing the entire commerce journey from product discovery to checkout and order management through a single abstraction layer.

ACP (Agentic Commerce Protocol) is an open standard co-developed by Stripe and OpenAI10. Merchants already using Stripe can enable ACP with as little as one line of code change. It’s also available through Wix, WooCommerce, BigCommerce, and Squarespace, keeping the barrier to entry low.

What You Can Do Long-term (6–12 months)

  • If you use Stripe: Enable ACP integration (minimal code changes)
  • Via Google UCP: Support AI agent checkout through AI Mode search and the Gemini app
  • Enable full agent-driven flows: Complete the entire journey—product discovery → cart → checkout → order management—via AI agents
  • Analyze and optimize: Measure conversion rates through AI agents and compare with human-driven transactions

Risks to Evaluate Soberly

The fear of “falling behind if you don’t adopt everything” is counterproductive. There are real risks in this space that deserve sober evaluation.

Protocol fragmentation. Multiple protocols—MCP, UCP, ACP—coexist today. They’re described as complementary rather than competing16, but supporting multiple standards may be necessary until consolidation occurs.

Uncertain effectiveness. As llms.txt demonstrates, the impact of new technical standards can be limited in practice. Evaluate costs against expected benefits soberly, and start with low-cost, low-risk measures.

AI agents as new “gatekeepers.” Just as Google became the gatekeeper of search traffic, AI agents risk becoming the new gatekeepers. The possibility that certain merchants could be favored by AI agent algorithms or partnerships cannot be dismissed, and resistance among retailers exists17.

Security. Exposing product data and customer information to AI agents creates new risks. Payment data handling in particular (such as ACP’s Shared Payment Tokens) demands careful design.

Conclusion

Anthropic’s engineers said “Don’t build agents, build skills.” Gartner predicted 40% of agent projects will be canceled. OpenAI abandoned end-to-end checkout in ChatGPT, proving that merchant-side infrastructure is what matters.

The message is consistent. For most businesses, the right answer is not to “build agents” but to “build services that agents choose.”

The path follows four layers: get “discovered” by AI through AEO and llms.txt, get “understood” through structured data and APIs, get “connected” through MCP, and get “transacted” through UCP/ACP. Start with low-cost, low-risk measures like structured data improvements and AEO, then expand investment incrementally as the market matures.

What’s certain is that AI agents are becoming a significant commerce channel. It’s never too early to start preparing.

You may also find these related articles useful:

References

References are listed in order corresponding to the citation numbers in the text.

Additional References (not cited by number in text)

  1. Don’t Build Agents, Build Skills Instead – Barry Zhang & Mahesh Murag, Anthropic - AI Engineer Conference (2025). [Reliability: Medium-High] Conference talk by Anthropic engineers—the AI agent development company itself advocating “don’t build agents” ↩︎ ↩︎2

  2. Agentic Commerce Impact Could Reach $385 Billion by 2030 - Morgan Stanley (2026). [Reliability: Medium-High] Market forecast by a major financial institution; projections carry inherent uncertainty ↩︎ ↩︎2

  3. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 - Gartner (2025). [Reliability: High] Official prediction from a major research firm, based on polling of 3,412 webinar participants ↩︎ ↩︎2

  4. Introducing shopping research in ChatGPT / OpenAI revamps shopping experience in ChatGPT - OpenAI (2025-2026) / CNBC (2026). [Reliability: Medium-High] Official announcement and major media coverage ↩︎ ↩︎2

  5. Box CEO Aaron Levie’s advice to developers? Build software that AI agents want to use - AOL / Fortune (2026). [Reliability: Medium-High] CEO interview in major media ↩︎

  6. Introducing Shopify Agentic Storefronts: Sell your products everywhere AI conversations happen - Shopify (2026). [Reliability: Medium-High] Official press release ↩︎ ↩︎2

  7. Under the Hood: Universal Commerce Protocol (UCP) / New tech and tools for retailers to succeed in an agentic shopping era - Google Developers Blog (2026). [Reliability: Medium-High] Official technical blog ↩︎ ↩︎2

  8. Shop like a Pro / Perplexity Shopping: How to Optimize Your Store for AI (2026) - Perplexity (2024) / Shopify Blog (2026). [Reliability: Medium-High] Official blogs ↩︎ ↩︎2

  9. Agentic Commerce Trends and Statistics for 2026 - MetaRouter Blog (2026). Citing McKinsey estimates. [Reliability: Medium] Secondary source. McKinsey original: The agentic commerce opportunity ↩︎

  10. Developing an open standard for agentic commerce / Agentic Commerce Suite - Stripe (2025-2026). [Reliability: Medium-High] Official blog and press release, includes NRF 2026 statistics ↩︎ ↩︎2

  11. Google AI Overview SEO Impact: 2026 Data & Statistics / AI Search and SEO Statistics 2026 - Stackmatix / Digital Applied (2026). Aggregating statistics from Gartner, Ahrefs, etc. [Reliability: Medium] Aggregation of multiple primary sources. Gartner original: Gartner Predicts Search Engine Volume Will Drop 25 Percent by 2026 ↩︎ ↩︎2

  12. FAQ on GEO and AEO: Where AI search and SEO overlap in 2026 / AEO vs SEO: The Future of AI Search Optimization - eMarketer (2026) / AthenaHQ (2026). [Reliability: Medium-High] eMarketer is an industry-standard research firm ↩︎

  13. Should Websites Implement llms.txt in 2026? / LLMs.txt: Why Brands Rely On It and Why It Doesn’t Work - LinkBuildingHQ / SE Ranking (2026). [Reliability: Medium] SEO industry media; adoption data based on BuiltWith tracking ↩︎ ↩︎2

  14. The Model Context Protocol (MCP): The Agentic AI “USB-C Port” for Ecom - Badger Blue (2026). [Reliability: Medium] Industry blog covering MCP for e-commerce ↩︎

  15. About Storefront MCP / Shopify MCP Server Guide: Building for Agentic Commerce 2026 - Shopify Dev Docs / WeArePresta (2026). [Reliability: Medium-High] Official documentation and technical analysis ↩︎

  16. Commerce MCP: AI for your Enterprise Commerce / Understanding MCP, ACP & UCP in Agentic Commerce - commercetools (2026). [Reliability: Medium-High] Official site and blog ↩︎ ↩︎2

  17. AI shopping tools gain traction, but retailer pushback could cloud 2026 progress - eMarketer (2026). [Reliability: Medium-High] Industry-standard research firm ↩︎

AI・Technology
AI agentic commerce MCP llms.txt AEO ecommerce business strategy
This post is licensed under CC BY 4.0 by the author.