McKinsey Says $1 Trillion In Sales Will Go Through AI Agents. Most Businesses Are Invisible.
TL;DR
AI agents only work if companies become agent-readable and agent-writable at the core — Nate’s main point is that OpenClaw-style personal agents are just the surface; the real bottleneck is whether product catalogs, checkout flows, shipping promises, and internal systems can actually be read and acted on by agents.
The old anti-bot internet is now blocking your future customers — the same CAPTCHAs, gated APIs, locked-down apps, and JavaScript-heavy interfaces built over the last 15 years to keep bots out are now preventing the very AI systems people want to use to shop, compare, and transact on their behalf.
This is already a trillion-dollar shift, not a sci-fi demo — McKinsey projects up to $1 trillion in U.S. B2C retail revenue could be orchestrated by AI agents by 2030, Google has launched a Universal Commerce Protocol, and Shopify’s Tobi Lütke is calling agentic shopping “the transformation of a lifetime.”
Slapping an MCP server on messy systems is not enough — Nate uses Stripe as the cautionary example: basic actions like refunds may work through MCP, but deeper systems like Sigma produce CSV-scale data that can blow up context windows and require secure intermediary databases, not just a wrapper.
If your delivery, returns, and product data are unclear, agents may skip you entirely — citing nShift, he argues that if shipping windows, return policies, schemas, or other structured details aren’t legible, the agent may simply never show your offer to the human, making your business effectively invisible.
The hardest part is turning tribal knowledge and marketing copy into structured data — whether it’s “the basketball used in March Madness,” “coffee that supports a school in Ethiopia,” or B2B claims like “scales to 10,000 customers,” the meaning humans care about often lives outside databases today, and that’s the real data-cleanup project companies can’t postpone.
The Breakdown
OpenClaw exposed the real bottleneck
Nate starts by saying everyone is obsessed with OpenClaw, personal AI, and flashy demos, but almost nobody is talking about the structural precondition underneath it all: companies themselves must be readable and writable by agents. His framing is blunt — it’s not about a chatbot feature, it’s about whether discovery, evaluation, purchase, and usage flows are built agent-first at the transactional layer.
The internet’s anti-bot defenses are now working against us
He flips the old product playbook on its head: the fences we spent 15 to 20 years building to keep bots out are now keeping our most valuable customers out. He gives WhatsApp as a perfect example, noting Peter Steinberger had to hack around Meta’s restrictions, and compares the cultural shift to being told as a kid not to get into cars with strangers — and now happily taking Uber or even Waymo.
Big incumbents are resisting, but consumer demand is louder
Google has reportedly moved to shut down some OpenClaw bot usage, and Apple recently restricted vibe-coding apps like Replit in the App Store, which Nate casts as incumbents trying to preserve closed ecosystems. He compares this moment to Napster: the original product got crushed, but the consumer demand it surfaced survived and eventually reshaped the industry through iTunes.
The money is real: agentic commerce is already being wired up
This is where he brings in the hard numbers: McKinsey says AI agents could orchestrate up to $1 trillion in U.S. B2C retail by 2030. He pairs that with Google’s Universal Commerce Protocol and Shopify’s push to bring more than a million merchants into agent-mediated shopping, then makes it concrete with a running-shoes query like “under $120, size 10, ships before Thursday, flexible returns.”
If your data is fuzzy, you disappear before a human even sees you
Nate’s warning here is sharp: if delivery windows, shipping costs, returns, or product schemas are unclear, agents may skip the offer entirely. That means the best product in the world can become effectively nonexistent if it’s not legible to machines, because more and more customer attention is shifting into chat interfaces where structured clarity beats branding polish.
Why “just add MCP” is the wrong mental model
He argues that execs love cheap shortcuts — wrap the API, spin up an MCP server, declare victory — but that misses the hard part. Using Stripe, he explains that basic agent actions are one thing, but deeper analytics like Sigma create huge data outputs that don’t fit neatly into a context window, so the real challenge is secure architecture, intermediary databases, permissions, and thoughtful data flow design.
SAP shows how deep the enterprise gap really is
On the opposite end, Nate says SAP’s incentives have always been to keep data inside the system, not make it portable for agents. Yes, SAP announced an MCP server for Commerce Cloud, but he says the gap between one AI-enabled slice and making real-world SAP installs broadly agent-readable is “the Grand Canyon,” and predicts this will become a major, undercovered vendor pressure story in 2026.
Four bad assumptions — and the ugly work companies can’t avoid
He closes by attacking four misconceptions: agent discovery won’t work like SEO, structured schemas are not just for simple products, trust in agent transactions is gradual rather than binary, and “wait and see” is basically a death sentence. The deepest point is that companies have hidden most product meaning in tribal knowledge and marketing copy — things like March Madness provenance, Ethiopian coffee sourcing, or proof a SaaS tool supports 10,000 customers — and now have to extract that into durable data if they want agents to surface them at all.