The Fences That Kept Bots Out Now Keep Customers Out

Everyone is excited about personal AI agents that book the flights and manage the calendar. Almost no one is discussing the structural precondition that makes any of it work: the systems those agents talk to — your product catalog, your transactional infrastructure, your systems of record — have to be readable and writable by agents. Not the chatbot bolted on top. The company underneath.

We spent twenty years building the wrong fences

Captchas, gated APIs, JavaScript-heavy interfaces — the modern stack was built to keep bots out, because bots polluted human experiences. That instinct is now backwards. If a growing share of buying intent arrives through an agent, the most valuable traffic you get is a bot acting for a human. And an agent doesn't browse a ranked list — there is no "above the fold." It evaluates structured data against explicit constraints. If your delivery window, return policy, or product schema is unclear, the agent skips your offer and no human ever sees it.

Wrapping an API in an MCP server is not the answer

The cheap move — stick a connector on the existing API and declare victory — covers a few percent of the real cases. The hard part is that most of what a customer actually asks about ("the ball used in this year's tournament," "proven to scale to 10,000 seats") lives in marketing copy and tribal knowledge, not in your data. Making that legible to agents is a multi-quarter project, and the companies leaning in hardest still haven't finished it.

The misconceptions that will cost you

• "We'll optimize for agent discovery like SEO." There's no ranked page to win; clean schemas beat ad budgets.
• "Schemas only work for simple products." The opposite — complexity is exactly what agents help customers navigate.
• "We'll wait and see." The category passes you before you're ready.

Build for agents first, and the humans follow — the clean data that serves an agent makes better human experiences, too.