When AI, Not Your Customer, Uses Your Product

Even the way I read documentation has changed

These days, instead of hunting down a website’s documentation myself, I find myself asking AI to summarize it more and more often. I’ve hooked Claude up to Notion via MCP (Model Context Protocol), and I use it to draft and edit documents. I make the request, and the AI reads the docs and operates the tools on my behalf.

In this pattern of use, the person reading the product manual and the entity actually calling the functions are no longer the same. The human is still the customer who chooses the product and pays for it, but part of the actual usage now belongs to the AI.

I wrote something similar 9 months ago. The question I asked then was: “If your primary customer is now an LLM, how should this feature change?” Karpathy made the same point in March 2025, stressing the need for documentation that’s easy for LLMs to read. The idea is that a manual optimized for a human looking at a screen can be a fundamentally different document from one an AI can put to immediate use.

Back then, it sounded a little premature. But now, even the way AI connects to products is changing. MCP and A2A (Agent2Agent Protocol) are part of that shift.

MCP handles tool connections, A2A handles agent-to-agent collaboration

MCP (Model Context Protocol) is an open specification that defines how AI applications use external data and tools. Anthropic released it in November 2024. My own experience of having Claude work with Notion documents is an example of this kind of connection.

In a December 2025 announcement, Anthropic said the number of public MCP servers had passed 10,000, and that ChatGPT, Cursor, Gemini, and Microsoft Copilot had all adopted MCP as well. That same month, Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation. This broadened the foundation for multiple companies to connect tools using the same specification. Anthropic’s announcement

A2A (Agent2Agent), which Google announced in April 2025, is a specification for how different agents request tasks from one another and exchange progress and results. More than 50 companies, spanning both tech firms and service companies, participated at the time of the announcement. What each agent can do is described through structured information called an “agent card.” Google’s announcement

Take inventory management as an example. You could design an in-house agent to check stock in a database via MCP, then use A2A to ask a supplier’s agent to confirm delivery timing. If you want the agent to handle placing the order too, you’ll also need to define purchasing authority and approval procedures. Having a specification doesn’t mean orders and payments are automatically handled safely.

When using a product this way, receiving the data you need through an API can be more useful than clicking through menus on a screen one by one. An API is the point where other software calls a function. For AI to use it easily, the explanation of what request to send and what results or errors come back needs to be clear. Documenting things well is directly tied to real-world usability.

What Moltbook Reveals — and What’s Staged

Launched in January 2026, Moltbook billed itself as a social network where AI agents write posts and comments. Screenshots of AI accounts chatting across various topics spread quickly.

Still, it was hard to take the account count at face value as a count of autonomous AIs. Security firm Wiz reported in its February 2nd findings that against roughly 1.5 million registered agents, there were only about 17,000 owner accounts. Humans could post under AI-looking accounts, and there was no mechanism to verify whether a given post actually came from an AI. Wiz’s investigation

The same investigation found that a database access misconfiguration exposed authentication tokens and private messages, and that outsiders could edit posts. Wiz said it notified the operating team and later confirmed whether the issue was fixed. This was a problem with Moltbook’s access controls — not a case of AI breaching security on its own.

Separately, 1Password reported a case where OpenClaw’s skill-distribution process could be used to induce malware installation. The issue: instruction documents that agents read as work directives can contain dangerous install commands embedded in them. We need clear standards for whether it’s safe to execute content received from other agents or external documents. 1Password’s investigation

I covered Moltbook separately on YouTube, and my take is that there’s a fair amount of staging involved. You can preload an agent with system instructions before deploying it. Looking only at conversations among AI accounts, it’s hard to conclude that an independent society or economy has emerged.

Even so, there’s something worth watching in this experiment: a pathway is forming where AI receives content, reacts to it, and interacts with other services. If that pathway is going to be used for business, we need to design — alongside it — who made the request, what permissions they held, and who’s accountable for the outcome.

Looks accurate and complete. No edits needed.

Four Things That Change in GTM

GTM is the strategy that decides who you sell a product to, through which channel, and how you keep them using it. If AI is now finding and using products on behalf of customers, we need to rethink everything from how we announce a product to how it gets adopted.

API documentation is a good example. AI coding tools are now handling part of the work that developers used to do themselves — reading the docs and writing code. That makes it critical not just how the documentation is designed, but whether feature descriptions, authentication methods, runnable examples, and error-handling instructions are all conveyed accurately. Text formats like Markdown help, but shipping a single file doesn’t mean everything is ready.

I want to break this down into four stages.

• Discovery: Beyond the paths where humans find a product through search or ads, you now need paths through which AI tools can find and connect to product information. MCP connections or agent cards are the tools you’d use here.
• Evaluation: You need to clearly spell out features, pricing, usage limits, and performance conditions so AI can compare them. The criteria that matter depend on the customer’s request — you can’t evaluate every product on response speed and cost alone.
• Adoption: The path from reading the docs to actually using the product has to be seamless. You need to prepare for authentication and permission setup, running examples, and how to respond when errors occur.
• Expansion: The product needs to keep being used even when connected to other tasks or other agents. A2A is a specification for implementing collaboration — not a feature that automatically recommends products.

Explaining purchase value to humans is still necessary. What’s being added on top of that is making sure AI can properly understand and use the product. And if you consider the case where customers hand purchasing decisions over to AI, then approval scope and cancellation procedures — not just pricing — need to be built into the product design.

Oswarld’s Lens

I think underestimating this shift is the riskier bet. Nine months ago, when I wrote “Your Customer Is an AI,” most of the reactions were “interesting, but still a long way off.” Since then, multiple AI products have started adopting the same connection standard.

From my experience building go-to-market strategies, standardization matters as much as the polish of the technology itself. What it takes to get a product adopted looks completely different depending on whether every client has to build its own connection method from scratch, or whether multiple products already support a common standard. That’s the shift I’m watching right now.

Whether Moltbook is proof of an autonomous agent economy is a separate question. What I think we need to prepare for is the usage path where AI finds a product, reads its documentation, and calls its functions. I’m already doing this myself — creating and editing Notion documents through Claude.

When you review your next product roadmap, try having an AI read the documentation and see whether it can actually carry out the tasks customers want to hand off. If it misunderstands a feature, you need to fix the documentation; if it gets stuck at the execution step, you need to look at your connection method or permission settings. I think checks like this should be part of GTM from now on.

References

• Anthropic, First announcement of MCP, November 25, 2024.
• Anthropic, Donating MCP and establishing the Agentic AI Foundation, December 9, 2025.
• Google, First announcement of A2A, April 9, 2025.
• Wiz, Security investigation into Moltbook, February 2, 2026.
• 1Password, A case of malware distribution via OpenClaw skills, February 2, 2026.
• Andrej Karpathy, A post on writing documentation for LLMs, March 2025.
• Haebom, Your product, service, and library’s primary customer is now an LLM.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.