Greg Isenberg put out a solo episode of The Startup Ideas Podcast called "The 9 Stages of AI Agent Adoption." His opening line is a Seinfeld bit. What's the deal with this war on AI agents?
He's not really joking.
Here's the setup, in Greg's telling. Amazon started blocking Meta's new shopping agent, Muse. Delta says it works with zero outside AI agents for booking flights, and calls them "unauthorized automated activity." Yelp will let agents in... if they pay.
So the biggest tech companies spent two years telling us agents would book our trips and buy our groceries. And now the places where those errands actually happen are shutting the door.
Then I did the thing he tells you to do at the end. I ran the test on my own business.
It did not go great. More on that in a minute.
The customer with no eyes
Greg's best idea is one line. An agent is a customer with a credit card and no eyes.
It can't see the homepage video you paid a fortune for. It doesn't notice your banner, your pop-up, or your countdown timer. It reads your prices, your shipping, your return policy. That's about it.
Everything you designed for human eyes? Mostly invisible.
He adds two more ideas that stuck with me.
First, the ladder of trust. You give an agent a goal, some rules and a limit. "Find me a replacement filter. Check with me before you spend over $50." Then it climbs three rungs:
Research. Find options, compare them.
Prepare. Fill the cart, hold the seats, get a quote.
Pay. Actually spend your money.
Almost nobody minds rung one. That's just search. The whole fight is over rung three. And Greg thinks a lot of businesses are blocking all three when the real risk only lives in one.
Second, agents are loyal to whatever worked last time. You and I stick with brands because of how they make us feel. An agent sticks with the site that got the order right. If it got stuck on your checkout, it goes somewhere else. And you never find out it happened.
Why the big guys are blocking
Greg is fair here. Some blocking is real safety. If an agent orders 50 of something instead of five, the customer calls the store, not the AI company. If an airline can't prove you approved a booking, that's fraud.
But then he says follow the money. Amazon makes a fortune on sponsored products at the top of your search. An agent skips them. Airlines make a fortune on seat upgrades, bags and the credit card offer at checkout. An agent books the cheapest nonstop and clicks none of it.
The agent isn't stealing the sale. It's stealing everything that used to happen on the way to the sale. And that's often the part with the best margins.
Then he lays out nine stages for how this war plays out. The names are my shorthand. The ideas are his.
Stage 1: The Lockout
What it means: Big companies block outside agents from their sites.
Why: Agents skip the ads and the upsells, which is where a lot of the money is. Greg says this is where we are today.Stage 2: The Arms Race
What it means: Agents start acting more like humans to sneak past the blocks.
Why: Every block gets a workaround. Every workaround gets a new block. Round and round.Stage 3: The Split
What it means: Companies pick a side based on how they make money.
Why: If you make money on sales, you let agents in, because they bring orders. If you make money on data, you charge agents to get in (that's Yelp). The giants build their own agents (that's Amazon).Stage 4: The Plaid Moment
What it means: A trusted middle layer shows up that proves who an agent is and handles payment. It becomes the standard way in.
Why: Banks once fought apps that pulled your account data. Then Plaid built a safe connection between the two, and banks went from blocking to signing deals. Greg says Visa, Mastercard and Cloudflare are already racing to be the Plaid for agents.Stage 5: The Switch
What it means: Businesses that welcome agents start winning customers from the ones that block them.
Why: If one airline lets your agent book in 10 seconds and another makes you do it yourself, your agent remembers. Next time, it goes back to the easy one.Stage 6: The Missing Stage
What it means: Nobody knows. Greg never says it out loud. Go listen. He jumps from five to seven.
Why it's funny: Later he says some companies are already "at stage six." So somebody's there. We just don't know where "there" is.Stage 7: The Race to Please Agents
What it means: Everyone competes for agents on purpose.
Why: Better prices for agents. Faster checkouts for agents. Product pages written for agents. SEO, but for agents. Greg asks a fun question here: what's a billboard for an AI agent?Stage 8: The Ads Come Back
What it means: Companies pay to be the first thing your agent recommends.
Why: The whole reason companies block agents is that agents skip ads. But ads never die. They just move, from print to TV to search. Greg says ChatGPT is already testing ads. He thinks almost nobody is talking about this one.Stage 9: The New Normal
What it means: Selling to agents becomes as normal as having a website.
Why: In a few years, a business an agent can't buy from will look like a business with no website looked in 2005.
He's also clear these won't happen in a neat order. Some companies are already far down the list while others are still at stage one.
Here's where I disagree a little
I run OutSnapped, an experiential photo and video company. We've been in the event photography industry since 2004 … So I asked myself the obvious question.
If a planner's AI agent showed up tomorrow to book a photo booth, would I let it in?
Yes. Easily.
But Greg says agents read your prices and specs. In my world, that's exactly the wrong thing to shop on.
Most people who book photo booths confuse the hardware for the experience. When a client comes to us, I don't want to talk about cameras. I want to know what they want the night to yield. The goals. The KPIs. And not just "how many emails did we collect."
What do they want their guests to feel? To remember? To walk away with? What's the experience while they're standing in the line? What's the rest of the event like, so we know whether to fit in or stand out?
An agent comparing hardware and hourly rates will pick the cheapest box with a camera in it. It's a perfect shopper for the wrong question.
I'm human … aka lazy. I want the agents.
Here's the funny part. I'm still all in.
I'm human, so I'm lazy. If our clients could book through their agent talking to our agent, I'd sign up today. It would help our margins a lot.
Most clients don't realize this, but almost all of our work happens before the event. The ideas. The builds. The revisions. On the night itself, we show up and execute. That's when we do the literal heavy lifting. Our vintage black-and-white photo booths need a box truck and a two person team to deliver them.
If agents can handle the back-and-forth up front, that's where the time comes back.
In theory, our agent could know everything we've ever told every client. It could answer questions on the fly, based on how I've done it before.
But here's the catch. My favorite part of my job is talking to clients and teasing out their hopes and dreams. That's the part an agent would replace.
So the human becomes the upsell
There's always a battle between quantity and quality.
My gut says it shakes out like this. Clients who want fast and simple book through an agent, at the price we charge today. Clients who want the hopes-and-dreams conversation get a human. And that becomes a new, higher tier on top.
Greg says agents skip the upsell. I think they might create one. The human touch is the upsell.
But I won't know until we try it.
Nobody tells you about the months
This is the part the hype skips, .
The agent-to-agent version is going to have a lot of growing pains on both sides. It won't save anyone time until the kinks get worked out. That means months of building skills, watching what happens, and fixing things.
People think these systems are set-it-and-forget-it. Turn it on, watch the money roll in. That's what the internet hype sells.
It's not true. It takes months of daily tweaks to get a system to where you're happy with it and feel safe letting it run on its own. I know, because I've been doing those daily tweaks on my own AI systems for months.
So I ran the test
Greg's homework is simple. Take a real buying question. Give it to a few agents. Watch where they get stuck. Then do the same on a competitor's site.
So I did. I built it from a real request in our inbox. A company after-party in Manhattan, midnight to 2am, about 150 guests. Classic enclosed vintage booth. Printed photo strips. Their logo wrapped on the outside. An attendant the whole time.
I gave it to Claude, ChatGPT and Perplexity.
Here's the scoreboard.
Task: Did the agent find OutSnapped on its own?
Passed? No.
Result: 0 for 3.
Task: When I asked about OutSnapped by name, did it rank us near the top?
Passed? Yes.
Result: 3 for 3. One ranked us #1.
Task: Did it quote our real price?
Passed? No.
Result: 0 for 3. Two of them made one up. Different numbers.
Task: Could it book us through our website?
Passed? No.
Result: 0 for 3.
Read that again. The agents liked us. They just couldn't find us, price us, or book us.
Below the line, I'll show you exactly what each agent said, the seven things on our site that tripped them up, and the copy-paste prompts so you can run the same test on your business this afternoon.
If you’ve been getting value out of Artificially Intimidating, this is the kind of work paid subscribers make possible. I’d love to have you.
