McKinsey Says $1 Trillion In Sales Will Go Through AI Agents. Most Businesses Are Invisible.

AI agents are on track to drive a massive share of commerce—McKinsey projects up to $1 trillion in orchestrated revenue by 2030 in the US retail market—but the real bottleneck isn’t agent intelligence. It’s whether companies’ systems are “agent readable and agent writable,” meaning agents can reliably discover, evaluate, and transact using the underlying transactional and data infrastructure.

For years, businesses built “anti-bot” defenses to protect human experiences and keep automated traffic out. Those same barriers now risk blocking the very customer interactions agents will soon dominate. The shift is already visible: vendors that once tried to lock down access are being forced to reconsider as agent-driven shopping and orchestration becomes mainstream. The core claim is blunt—personal AI platforms and enterprise agent frameworks only work at scale if the entire ecosystem of company systems can be consumed and acted on by agents, not just through chat interfaces, but through the transactional layer that powers product discovery, pricing, shipping promises, returns, and checkout.

That ecosystem change is hard because it requires deep internal restructuring of data stacks. Adding an MCP server or wrapping an existing API is treated as “low-hanging fruit,” not a full solution. Agents need structured, trustworthy schemas and secure access to the right slices of data. Without clean, consistent data across the stack, agents either skip offers or produce unreliable results—leading to lost sales even when the underlying product is strong. The transcript draws a parallel to earlier personalization efforts at Prime Video: without clean data, customer experiences collapse. With agents, the stakes rise because the agent must operate at scale and with precision.

Several examples illustrate the difficulty. Stripe is cited as an early adopter that shipped an MCP server, but the challenge appears when deeper analytics (like Sigma queries returning large CSVs) can’t simply be poured into an agent context window. The solution requires intermediary storage and secure, queryable database structures—plus careful security and authentication controls at Stripe’s scale. On the other side, SAP is described as announcing an MCP server for commerce cloud while the broader SAP portfolio remains far from “agent readable by default,” implying multi-quarter work for most installations.

The transcript also argues that common executive misconceptions will slow progress. Winning won’t come from optimizing for agent “discovery” the way search engines do; agents evaluate structured data against explicit constraints rather than browsing ranked lists. Complexity isn’t a reason to avoid schemas—agents benefit from structured access because it helps customers optimize decisions they can’t evaluate manually. Trust is framed as a spectrum: agents start with narrow delegated tasks and earn broader permissions over time.

Finally, the most important strategic point is that businesses must encode higher-order intent into data. Humans can ask for vague, emotionally loaded product attributes (“the basketball like the one used in March Madness,” or coffee tied to a specific farm and school). If those meanings live only in marketing copy or tribal knowledge, agents can’t verify them and transactions fail. The prescription: benchmark competitors by attempting agent-mediated transactions, then invest in making data architecture agent-first. The payoff is twofold—agents can transact, and humans benefit from the same clean data through better, more personalized experiences.