How AI Agents Drain Profits from Great Businesses

Get smarter on investing, business, and personal finance in 5 minutes.

This week's newsletter is an adaptation from my recent YouTube video on AI Agents.

This week’s Five Minute Money is going to be a bit different than some of my previous issues.

I'm just going to share with you all sorts of different things I'm thinking about how AI can change a lot of businesses, make some worse off, and be enablers for others, with specific examples and a few frameworks to think it through.

To start off, what happened was that Meta recently released something called Muse AI.

This is Meta’s version of a personal AI agent for consumers.

You may ask, “Couldn't Claude do a lot of this stuff before if you just gave it permission?”

Yes, but the thing that changed was how soon this could come, and how this would look for a lot more average consumers.

Muse is very similar to ChatGPT, except it also has its own virtual computer, so it can open up its own web browser on the server side and click into pages with its own mouse.

The UI is also very simple, whereas ChatGPT and Claude's agentic workflows are more complex, and most people don't even use them.

It became scary when people saw how this is like the transformation of the internet, some have called it the end of the open web.

Websites have seen traffic drop as much as 20%, 30%, 40%, 50%, because people are getting the search results right there in AI mode, or going to ChatGPT or Claude instead of Google.

With Muse you have this first mass consumer example of AI agents surfing the web on your behalf, looking up restaurants, checking reservations through connected apps, or navigating a restaurant's own reservation system directly.

Meta has given everyone their own virtual computer and a lot of free usage on it, subsidizing enough usage to get to the point it becomes very mainstream, at which point they can monetize it, potentially through a commission on purchases done through Muse.

This gets to the first idea of the point of monetization.

The Point of Monetization.

I came up with this idea thinking about dating apps, notoriously horrible businesses, and it was a paradox why that was so.

It's a digital service, no capex, should have extremely high gross margins, and once you get network effects going, why is it so hard to monetize them, especially given the value provided is literally finding your spouse?

Imagine I told you, I can help you find your husband or wife, what would you pay, $20?

No chance, you'd pay thousands of dollars, especially if you're a little desperate, which is sometimes the cohort of people on dating apps.

Look up the stocks of Bumble and Match, they've struggled, issues growing revenue, retention problems.

If you successfully match someone, they leave the app, so getting users to leave is a sign of success, a weird quirk in the business model.

The real problem is you charge people before you prove any value, they can't charge you saying, give us $10,000 if we match you with your wife, and if we don't, don't give us anything.

So instead they monetize premium features like hiding your age, super likes, Hinge's digital roses.

Willingness to pay is decreased by the ambiguity of whether this is going to work, so historically these businesses haven't been great.

The key takeaway, and how this relates to AI, is that with dating apps the attempt to monetize comes before they're proving any value, because after you get someone's number and hop off the platform, Hinge can't come back and ask for another $500.

So the point of monetization is different for every business.

For a typical consumer product company, it's when they buy the product, so the product and advertising convey the value ahead of time, sometimes with money-back guarantees.

Services are a little different, sometimes based on success, like lawyers taking a percentage of a winning case, but there's the same issue of not knowing the quality ahead.

How AI Is Shifting the Point of Monetization.

The point of monetization is shifting, and this is already happening.

Traditionally, if you looked up investing questions on Investopedia, they'd put banner ads on their website and collect a couple cents or fractions of a cent, perfectly aligned with your usage of the web and their revenue model.

Instead, AI is now taking the information from the webpages and circumventing the advertising revenue they used to get, so you're just not visiting the actual websites anymore.

The value used to come at the time you visited the site, and now it comes first instead, so no one clicks into the site, and their revenues have been dropping a lot.

Take Airbnb for example.

For alternative accommodation you could do Airbnb or Booking, both have millions of listings, Airbnb has a little more unique supply.

So you go to Muse and say, find me an Airbnb in Lake Tahoe.

Muse knows you don't actually care if it's Airbnb, Booking, or VRBO, you just want a home, so it searches those sites and cross-checks whether there's a direct listing from the property manager at a lower price.

Airbnb is still providing value, but now Muse is giving it to you, and when you're actually monetizing, the transaction portion gets de-paired, now Airbnb provides the information and you go book directly or through VRBO instead.

Marketplaces, Sponsored Ads, and the Consumer Hierarchy of Preferences.

There's other issues with sponsored ads, one of the biggest headwinds for marketplace businesses.

A typical marketplace takes a commission on the product sold, but most incremental profits come from ad revenue, pushing a sponsored listing up in the results.

Say you type Clorox wipes into Amazon's search bar, and Clorox buys a sponsored ad to be top of the results, because if they don't, a cheaper knockoff buys that slot and wins the sale.

What Amazon is monetizing is effectively your impatience, and the same exists on Google Search, where FootLocker has to pay for a sponsored ad so you're not accidentally clicking on someone else's store.

This ties back to the consumer hierarchy of preferences.

If you recall,, the consumer hierarchy of preferences is any time you buy something you have preferences, and once you fulfill enough of them you get a sale.

Want lunch that's cheap, close, and relatively healthy, meeting those 3 criteria gets a sale, and a free drink and chips on top builds consumer surplus, things that could have been charged for but weren't.

In the marketplace model it's flipped, consumers don't have a preference for sponsored ads, at best they're indifferent, at worst annoyed scrolling past them.

Yelp is the worst offender, you scroll down and about halfway through realize every listing was a sponsored ad, very annoying, a company taking away consumer surplus.

Tying this to agentic AI, marketplaces like Amazon, apartments,com, and Booking won't have the ability to show sponsored ads anymore, because they won't have a captive user, it's Muse, and Muse doesn't get tired or care how much it has to scroll past sponsored ads, it's going to just skip that part.

Something that doesn't provide value is the most at risk under this idea.

Why Airbnb Still Has an Edge.

Going back to Airbnb, they provide a lot of value with unique listings, and more important, a trust layer.

When you're booking a home with a stranger, you want to know that if something goes wrong, someone has your back, and that it's vetted, and that's what Airbnb provides.

There's still risk, because that home might also be listed on VRBO or booking(.)com, and Muse might be able to price check the two.

But since Airbnb has so many unique listings, they have much more vetted supply, a way of providing value, whereas an AI agent search doesn't tell you the quality or how it compares across platforms.

That's not to say it's not disruptive, it is, but the point is it's still providing value, so they don't have to be as worried about getting disintermediated, but if you're doing something other people cannot do easily, you still have a good shot at charging for it.

This ties into a similar example of the debate between software companies vs AI.

AI Software Debate.

One of the tricky things about the debate over whether AI is going to replace software is there's a lot of software out there that does functions it seems like AI would eventually be able to do.

The tricky part in the moats of that software has to do with a few things:

1. Habits and Switching Costs: People and businesses are very unwilling to switch existing providers because of habit, they're used to a certain UI, and don't want to force employees to retrain on a new one.

2. Fear of Something Going Wrong: Whenever you're switching out your systems, there's risk that something goes wrong, and that's not comforting to take on.

3. Mission Critical: A lot of times these systems are mission critical and very embedded with everything else going on in the business, so switching could be a big headache

That's why I think a lot of software companies are going to be fine, staying in the back end as infrastructure, because they'll just have the AI use the software.

If it's a mission critical piece, I'm thinking of a ServiceNow or a Salesforce, they'll be fine because the AI is still going to use the software plumbing.

Over time, amongst new entrants, that could change, because if you don't already have an embedded platform, you may just trust the AI to build it out for you.

The point is these companies are still providing value, so they're not leaving, and whether it's subscription or usage based doesn't matter, as long as it's getting used, they continue to profit.

So the big takeaway is whether a service is still providing value, and whether it's provided in a way that lets them monetize at the point they're providing it, because like Investopedia, if you provide value up front and no one visits your site, you never get the chance to monetize it.

This shifts any time the customer journey or interface changes, so we have to think about it example by example, because it could be different for 2 seemingly similar businesses.

Apartments(.)com.

If you're familiar with Costar's apartments(.)com, it's one of the most popular ways to rent an apartment in the US, free on the user side.

You search, get listings, and they monetize by charging bigger property managers to buy gold and platinum advertising packages, which push the listing to the top.

However, you never rent an apartment online, you always visit in person, so what if an AI agent scrapes all of that data and says, “Here are the 5 listings you should want.”

Apartments(.)com doesn't give you the lead information unless you log in.

The AI agent could find the same listings on other websites, cross-checked with MLS and Zillow, got the phone numbers and emails, and set up the visit times, which circumvents the need to ever visit the site.

So that point of monetization is no longer relevant, you can't monetize it the same way with usage.

That's overstating the case, though, there's content on apartments.com that's harder to pull out, like floor plan tours and lease applications you can get pre-approved for right there.

Something like Muse web-scraping is very compute intensive, and it's not going to do it right now.

I asked Muse, can you find me hotel rooms in Florence, and it pulled up 5 listings, all from booking,com.

Then I said, price check these, and it went direct to each hotel's website, taking 14 minutes for just 5 hotels, so you can imagine scraping the entire web would take even longer.

Things like Booking and apartments,com exist because they already aggregated supply, and the path of least resistance is simply using them.

It's tricky, because what if AI uses apartments,com as a search engine to find the listing, then price checks quicker across other platforms, it's still providing value, just leaking out where it is.

Booking is in a better position because the transaction still happens on the platform, and a lot of times they're equivalently priced or cheaper.

Minimum Viable Products vs. Maximum Possible Products.

Now we're going to talk about the maximum possible product versus the minimum viable product.

An MVP is basically the worst product you could bring to market while still proving its purpose, users give you feedback and you iterate and make it better.

AI is an MVP product, especially when it initially rolled out, the best worst product it could make while people would still use it.

If you remember ChatGPT-3, it was fun and kind of helpful, but made all sorts of hallucinations, then got better over time.

It's a unique thing that a company as big as Anthropic or OpenAI can put out MVPs, because becoming big usually brings bureaucracy that prevents minimum viable products, a weird advantage for them, people just expect things don't always work and they'll get better.

Google had this issue, they'd researched AI for decades and bought DeepMind 15 years ago, but sat on it for fear of chatbots lying to people, so they never launched a product that quickly, and ChatGPT took their lunch, a new channel for search that's never going to go away.

Apple is in a similar position, and its painful to watch because they keep announcing AI features they can't deliver on, since Apple has a reputation of being super polished.

If we compare that to Meta’s Muse, it isn’t super polished. It breaks all the time, generates the same image over and over.

But many consumers see AI as this technology that is truly incredible and free, but come to expect it doesn't work a lot of the time, and that's the reputation it's giving.

That was fine for Meta, but it's not what people expect out of Apple, a tough position, because in contrast to an MVP is a maximum possible product, and Apple basically only makes MPPs.

Any time they create a product like the iPod or iPhone, they create the best possible product at the time, pushing limits farther than people thought possible, like Steve Jobs pushing Corning to create a glass they said was impossible, because their products had to be perfect.

Dyson is another MPP business, making products so good they convince people to want something they didn't know they wanted, a fancy vacuum cleaner, unnecessary until you see it work.

That's the world of MPPs, while AI so far has been in the world of MVPs, creating issues for incumbents like Google and Apple.

The Shifting Positions in the AI Race.

Google is wavering in the middle, but Apple hasn't made their AI move because they feel it has to be perfect.

When they do, it could be a killer to agentic AI apps like Muse, which kind of works and kind of lies to you sometimes.

Apple could release an AI product that has a consumer experience where you never had it break on you, connected with all your apps because of the OS layer, so it could work better than anything Meta, OpenAI, or Anthropic has.

Something I've observed is how quickly AI company positions have changed.

First it was ChatGPT, and people said Google was dead, then Google was suddenly number one, then Anthropic, and now Meta's Muse is having its time in the sun, and SpaceX is out there with Grok.

Apple hasn't made their move yet, and may not need to, because what matters is the product, not the model.

Products Over Models.

It doesn't mean that much how a model tests on benchmarks, what matters is the value it provides, and the only people who really care about models are research scientists.

If Muse's model is worse than Anthropic's, it doesn't really matter, Muse does things Claude can't do, so it's more useful to me.

ChatGPT was great because it was very user-friendly, reminiscent of early Google's simple white page with a search bar, and everyone copied that clean format.

The question is whether Apple comes in later, since OpenAI, Anthropic, or SpaceX would all be happy to partner under the right terms.

There are enough options these AI models could get commoditized already to an extent, the capabilities we had 6 or 9 months ago are already open source for free.

Frontier models are maybe half a year ahead, and their moat of staying on the leading edge is going to get harder to do.

I think they'll realize the rate of improvement is becoming smaller, and what it means for the user experience too.

OpenAI and Anthropic will realize they have to productize AI better and move down into verticals, the scariest day for software companies, because model intelligence alone is not a moat, and there'll be enough good enough options that people swap out models, Apple included, powering Siri with Gemini or OpenAI or Anthropic.

As long as Apple controls the consumer through the device, not the model, OpenAI and Anthropic have to productize better too.

Claude and OpenAI already did this once with coding tools, Claude Code and Codex, productizing their models and competing against existing competitors.

I think they'll do this in many other verticals, because if you don't control the consumer endpoint, going through something like Cursor, your loyalty is to Cursor, not the model underneath.

OpenAI cut access to their model for Cursor users, and Cursor just switched models, exemplifying who's more powerful in the value chain, the end user cared about Cursor, not Codex.

It doesn't look like we're heading to a world where one model is so superior to all the others.

As long as that's the case, there will be more push to go deeper into the vertical.

I could see ChatGPT getting more into small business work, marketing, accounting, finance.

The more the model layer gets commoditized, the more they'll press on the distribution they already have, for better lock-in with a stronger product.

If not, they're running a rat race trying to have a better model than someone else, very hard to sustain.

Watch the Consumer Surplus.

The last thing I want to touch on is the consumer surplus.

Some AI companies are better positioned to offer free usage without taxing it with ads, because Meta's bet is to subsidize Muse, and ChatGPT or Claude wouldn't be able to do that, they don't have an ads business pumping out tens of billions a quarter in revenue each quarter.

If Anthropic or OpenAI rolled out a similar product, they'd have to pay for it, meaning Meta is better positioned for the masses, and Google is similar, just slower and more reluctant.

ChatGPT originally was free, then they started adding ads and they did that because they need that for revenue.

Meta doesn't, they'll just keep subsidizing.

Anthropic is already struggling with what to give away versus monetize, since they have a leading edge model, Fable, and only gave people access for a couple weeks on trial, then a weird credit arrangement, not a great consumer experience.

All these AI companies are giving you a product, in many cases for free, and then taking some of it back, because the financials of the businesses aren't quite balancing properly.

Other businesses, like Meta, don't have that issue, they're better positioned to keep giving consumers something free for a long period of time.

Apple is also in a similar position and we'll see whether they come out of the gate with a subscription or make it free for AI users.

Summary.

The important question for investors is not simply which companies will be disrupted by AI, but whether they can continue to provide and monetize valuable services as the customer interface changes.

Companies that primarily aggregate information, sell access to a captive audience, or depend on advertising may be particularly vulnerable if AI agents become the new interface between consumers and the web.

By contrast, businesses with unique supply, trusted relationships, mission-critical infrastructure, or control over the customer relationship may remain well positioned.

As AI models become increasingly commoditized, the durable advantages may sit less in the model itself and more in the product, distribution, and vertical application built around it.

When evaluating an AI-exposed business, I think the most useful questions are simple: What value does the company provide, can an AI agent replicate it, and does the company still control the point at which that value is monetized?

For More on AI Agent Risks, check out the video below.

YouTube FAQ

Nothing in this newsletter is investment advice nor should be construed as such. Contributors to the newsletter may own securities discussed. Furthermore, accounts contributors advise on may also have positions in companies discussed. Please see our full disclaimers here.