Teamday
← Back

Greg Isenberg: The 9 Stages of the War on AI Agents

Greg Isenberg
agentsautomationbusinessstartupfuture-of-work

Why Greg Isenberg Says the War on AI Agents Decides How We Buy

Greg Isenberg, host of the Startup Ideas Podcast, former advisor to TikTok and Reddit, and co-founder of LCA, argues that the most important fight on the internet right now is between personal AI agents and the businesses they try to buy from. Amazon has blocked Meta's Muse agent from its store, Delta says it works with zero outside agents for shopping or booking, and Yelp only lets agents in through its paid data licensing program. In his view, whoever wins decides how people will buy things for the next decade.

Three camps: Isenberg sorts companies into those that block agents (Amazon, Delta, Yelp), those that block them by accident (Zillow welcomes agents acting for real customers, but its anti-scraping tools sometimes stop them, and Walmart, a Muse partner, called failures on its site an accident), and those that welcome them. On the welcoming side, Meta is working with Walmart, Stripe and others on an open standard for agents and businesses, Shopify lets merchants sell through AI channels, and Visa and Mastercard are building ways to recognize trusted agents. Many blocks are old bot defenses that can't tell a scraper from a personal assistant.

A new mental model: "an agent is a customer with a credit card and no eyes." It never sees the homepage video or the countdown timer. It reads product details, prices, shipping costs and return policies. Agents also work on a ladder of trust: research, prepare (fill the cart, hold seats, get a quote), and pay. Most companies are fine with research because it is basically search. The whole war, he says, is happening on the third rung. Isenberg thinks many businesses err by blocking all three rungs when the real risk sits in one. And loyalty changes: "an agent is loyal to whatever worked last time." A checkout that stalls loses the customer without the business ever knowing.

Follow the money: Some reasons to block are fair. Someone has to fix it when an agent orders 50 items instead of five, an airline that can't prove a booking was authorized has a fraud problem, and plenty of "agent" traffic is really scraping. But agents also skip Amazon's sponsored products and airline upsells such as seat upgrades, bags and card offers, which are often the highest-margin parts of the sale. The deeper fear is Meta or OpenAI sitting between the store and the customer, which is why the companies blocking outside agents are building their own, such as Amazon's Rufus and Delta's AI concierge.

Each player's bet: Amazon is a walled garden. Meta wants every door open because it owns the agents walking through. Walmart, as number two, partnered with Muse to take share. OpenAI puts checkout inside ChatGPT for Etsy and Shopify stores. Google builds agents and payments into Gemini while protecting a click-based ads business. Shopify is the arms dealer, Visa and Mastercard want the money to move through them whoever wins, and Cloudflare wants to be the bouncer.

The nine stages: Isenberg uses the Mint and Plaid story, where banks went from blocking to signing deals to charging for access, as the template. (1) Companies block agents because they skip ads and upsells, which is where we are today. (2) Agents learn to act like humans and an arms race follows. (3) The market splits: sales-driven companies let agents in, data businesses charge, and the giants build their own agents. (4) A trusted middle layer for agent identity and payments becomes the standard way in, "the Plaid moment." (5) Businesses that welcome agents win customers from those that block. He doesn't describe stage six in the episode. (7) Everyone competes for agents with better prices, faster checkout and "SEO but for agents." (8) Ads come back as companies pay to be the first thing your agent recommends. (9) Selling to agents becomes normal, and "a business that can't be bought by an agent is going to look like a business without a website" did in 2005. He adds that the stages will blend, and different companies are already at different ones.

5 Moves Isenberg Recommends for an Agent-Ready Business

  • Run an agent test - Pull 10 real buying questions from your support inbox, give each to two or three agents, and watch where they get stuck. Then do the same on a competitor's site.
  • Check for accidental blocking - Bot protection set up years ago may be turning agents away. Make that decision on purpose.
  • Make the offer readable - Put total price, delivery date, compatibility and return terms in plain text on the page, not in a PDF or on the last checkout screen.
  • Make the advantage readable - If you cost more but deliver tomorrow, the agent has to see the delivery date. Agents work for people who still value speed and reliability.
  • Set rules per rung - Let agents research freely, let them build a cart or request a quote, and add an extra check before money moves on big orders.

Where the Startup Opportunities Are in Agentic Commerce

Isenberg's list of businesses this shift creates: agent readiness for one niche industry (start by fixing it by hand 10 times), a Plaid for one vertical such as contractor booking or restaurant supplies, insurance for agent mistakes, analytics that tell agent visits from human ones, and being the specialist that agents reward for doing one thing well. He would avoid any business whose edge is a trick to get mentioned by AI, because tricks stop working.

For organizations deploying agents, the same logic runs in both directions. Your agents will be the customers with no eyes on other people's sites, and other people's agents will judge yours by what they can read and complete. His closing test is the useful one: "Are you protecting the customer or are you protecting a habit you'd like the customer to keep?"