Daniel has a prediction he wants on the record. He wants to be able to listen back in a few years and say he called it, or admit he didn't, but either way it's going in the archive.
That's the honest way to do a prediction. Put a date on it.
So here's the shape of it. There's a plurality of apps for everything now. He and Hannah want to coordinate grocery buying, which is one app, and then the smart home is another, and the calendar is a third, and email is a fourth, and nobody wants twenty apps on their phone each doing one thing. The path trends toward consolidation.
And he's noticed something about why.
He thinks conversation is just easier for managing all these little things than any other interface. So vendors are rolling out MCP servers so agents can act on our behalf, turning conversational AI into the new front end for the program. And then the part he's actually predicting: if agents get good enough, the only value the software brings is its backend. And OpenAI or whoever starts shipping its own app inside ChatGPT, grocery manager, and your lists and your purchase statuses live in a reserved slice of their database.
And that's how they eat the verticals one by one.
He flagged the lock-in himself. Said he doesn't actually think it's a good idea. But he predicts most people take the path of least resistance rather than build their own tool, even now that building your own is easier than it's ever been. He also says the usual discourse gets this wrong.
The usual discourse being that agentic AI makes SaaS obsolete because companies will just build their own.
Right. He thinks that happens too. He just doesn't think it's the main event. He thinks the AI vendors start selling the bolt-on.
And he's already browsing the plugin section on ChatGPT watching it grow.
So, steel-manning duty is mine, you're arguing his side. Let's do the mechanism first, because the prediction is only as good as the plumbing under it.
The plumbing is MCP. Anthropic open-sourced the Model Context Protocol on the twenty-fifth of November, twenty twenty-four. Universal open standard for connecting AI systems to data sources, client, host, server architecture, JSON-RPC underneath. And the point of it was to replace the mess of fragmented one-off integrations.
One protocol instead of every vendor writing a custom connector for every assistant.
And here's what makes Daniel's prediction more than a hunch. OpenAI's Apps SDK is built on MCP. Explicitly. They said it's an open standard for connecting ChatGPT to external tools and data, and they made the SDK open source specifically so apps built with it can run anywhere that adopts the standard.
Which is a strange thing to do if your plan is to eat everybody.
It is. Hold that thought, because it's the fork in the whole episode. Timeline. DevDay, sixth of October twenty twenty-five, OpenAI launches apps in ChatGPT and the Apps SDK in preview. Pilot partners: Booking.com, Canva, Coursera, Figma, Expedia, Spotify, Zillow. Seven of them, with eleven more announced as coming soon.
And the pitch to developers was eight hundred million plus users.
Reachable. Then the thirteenth of November, apps go into preview for Business, Enterprise and Edu customers. Then the seventeenth of December, OpenAI opens app submissions and launches the App Directory.
That's the app store framing that everyone ran with.
Because it is one. And the mechanism of how these work is exactly the front-end argument Daniel is making. Apps fit naturally into conversation. They respond to natural language, they have interactive interfaces right in the chat. You trigger them with an @ mention, or through the tools menu, or ChatGPT surfaces them proactively using conversational context, app usage patterns, user preferences.
It decides you need the app before you've said you need the app.
And OpenAI's own example of what these do, in their developer announcement, is that apps let users order groceries, turn an outline into a slide deck, or search for an apartment.
Groceries.
Groceries. His exact hypothetical.
So either Daniel is clairvoyant or he read the same launch page as everyone else and noticed which example they led with.
The example choice is not nothing, Corn. When a platform company picks the demo, it's telling developers what it wants built.
Or it's telling them what's easy to demo. A grocery order is legible in four seconds of video.
Both can be true. Here's the part I think is underrated in Daniel's argument, which is the interface claim. He says conversation is easier for humans than any other interface for managing little things. And the market framing agrees with him. J.P. Morgan Private Bank put out a piece arguing investors expect agents to become the dominant interface, a one-stop shop for workflows, lowering barriers to entry. That's the same claim in banker language.
Except nobody's measured it.
No. There's no study I could find that says conversation beats a list for task management. It's industry framing, not measured research.
Which is where I want to push, and this is me steel-manning him, so hear me out. The strongest version of Daniel's case isn't that conversation is better. It's that conversation is the only interface the AI can use without being built for. Every other interface, somebody has to design it for the agent. If it's a screen, there's a question of what's on it and where. If it's an API, there's a question of which one and how it's versioned. Conversation is the one surface where the agent doesn't need anything from you. It just talks.
So the interface isn't better for humans, it's better for the agent. That's the steel-man.
And the second leg is that the vendor doesn't have to get a user to change behavior. They already typed a sentence into the box this morning. The onboarding cost is zero.
The distribution argument. That's the one that makes me think he's right, not the interface argument. Eight hundred million people already have the box open.
But here's the counter-evidence, and it's sitting right there in January. Microsoft renamed Office to the Microsoft 365 Copilot app.
The reaction on Hacker News was not warm.
It was openly hostile. The line that stuck with me was someone saying the productivity hit has been extremely negative, and then: I want a blank spreadsheet. It gives me a document to download.
That is the single best piece of evidence against the whole thesis.
Because it's a giant company forcing conversation as the front end onto a tool people already knew how to use, and the users revolted. Not because the AI was bad. Because the blank grid was faster.
Although. I'd narrow it. The complaint wasn't that chat existed. It's that chat replaced the grid. If Microsoft had said the grid is still right there, here's a Copilot panel next to it, I don't think anyone posts anything.
So the lesson is conversation as an option travels fine. Conversation as a mandatory layer gets you a very angry comment section.
And that distinction matters for the prediction, because Daniel's version is opt-in. You add the micro app. You don't get it forced on you.
But the flip side of opt-in is that it only takes over when it's better than the alternative, which is not what he's predicting. He's predicting people take it because it's easier, not because it's better.
Path of least resistance. His words.
Which takes us to the part where the money ran. Because if Daniel's right, the market has already started pricing it in, and the market has been brutal.
Brutal is right. Forrester published a piece with the title SaaS As We Know It Is Dead, arguing investors fear SaaS won't be the provider of choice for AI agents, and per-seat models don't survive an agent doing the seat's work. The IGV software ETF fell thirty-six percent from its late-October twenty twenty-five high to a multi-year low in April.
That's a sector-wide repricing, not a wobble.
Valuations compressed to twenty-two times earnings in March, against a Covid-era peak of fifty-one. And one analysis cited around two hundred eighty billion dollars of SaaS valuation wiped out.
Two hundred eighty billion is a lot of prediction to be living inside a spreadsheet.
And the individual hits are what you remember. ServiceNow dropped ten percent in late January despite beating on Q4 earnings. Morgan Stanley's line was good, but not good enough. Anthropic previewed Claude Code Security in February and cybersecurity equities fell between ten and twenty-five percent in days. CrowdStrike, Palo Alto, all of them.
A product preview moving equities that much means the market believes the substitution story. At least directionally.
Now, the counter-thesis, which is what VCs are actually funding. The winners, on their read, are products embedded in real operational processes, with ownership of structured or proprietary data, and deep domain expertise.
Meaning the company keeps its own tools instead of renting them.
That's the version of the future where Daniel's prediction loses. Companies don't take the bolt-on because the bolt-on doesn't know their business. Jake Saper at Emergence put it in a way that's stuck with me: being the connector used to be a moat, soon it'll be a utility.
Which is the double edge of MCP nobody says out loud. The same protocol that lets a vendor plug into every assistant is the protocol that makes the vendor interchangeable.
He also said, pre-Claude, getting humans to do their jobs inside your software was a powerful moat, but if agents are doing the work, who cares about human workflow?
The workflow was the moat and the workflow is the thing the agent deletes.
That's the sentence. He called Cursor versus Claude Code the canary in the coal mine. One owns the developer's workflow, the other just executes the task, and developers are increasingly choosing execution over process.
Developers choosing execution over process is a very strong signal, because developers are the group most attached to their own workflow. If they'll ditch the workflow for the shortcut, everyone else will.
Which is Daniel's path-of-least-resistance argument, arriving from the opposite direction. He says people take the bundled option. Saper's saying they're already doing it.
So here's my steel-man of Daniel, complete. The interface converges on chat because it's the only surface agents need nothing from you to use. The distribution is already installed. The incumbents are already bleeding in the public market, which means the substitution is priced. And the counter-argument, that companies build their own, is true for the top of the market and irrelevant for everyone else, because most people are not building anything.
And that's the asymmetry in his prediction, and I think it's the smartest part of it. The build-your-own story has no floor. Anyone can be the company that builds. The bolt-on story has a floor, because there's a finite number of vendors rich enough to run an eight-hundred-million-user distribution channel.
Now the hole.
There isn't one. There's a gap.
Say the gap.
As of today, there's no evidence OpenAI has shipped a single first-party vertical micro app. Not a Grocery Manager, not a booking app, not an email client. The App Directory is third-party developers. Spotify, Zillow, Canva. The closest thing to Daniel's Grocery Manager is OpenAI's own example sentence about apps that order groceries, and those apps belong to other people.
So the mechanism is fully built and the villain hasn't shown up.
The mechanism is built and OpenAI is standing in the doorway inviting people in. That's the fork. The prediction requires OpenAI to flip from platform to competitor.
And flipping means competing with the developers it just recruited, using the distribution it just handed them.
Which is a real cost. If you ship Grocery Manager and the third-party grocery app is sitting in your directory, you've just told every developer in the directory what their shelf space is worth.
You'd only do it for the verticals where the third parties never showed up. Which is not gobbling. That's filling.
That's the version I keep landing on. The platform probably ships the verticals nobody else built, and it absolutely ships the ones that make the platform money directly.
Payments.
Payments. And you can watch the monetization staging. Early phase, apps link out to the developer's own site for physical goods. OpenAI is exploring digital goods, and there's the Agentic Commerce Protocol for instant checkout.
So the first thing inside the box that belongs to the box is the checkout.
It's always the checkout.
On the lock-in point, Daniel flagged this himself and I think he undersold it. He said the data lives in a reserved slice of the vendor's database, and he said he doesn't think it's a good idea, and then he moved on.
Which is where I'd plant a flag. Your grocery list is trivial. Your purchase history is not. Your smart home routines are not. If the agent is the only thing that can read your own data because the data lives in its database, you don't have a subscription. You have a hostage.
And the switching cost compounds quietly. Year one you move a shopping list. Year three you've got five years of household purchasing inside one company's product and you can't reconstruct it anywhere else.
Which is the pre-IMAP era of email all over again. Everybody's mail living in one client, and the day you wanted to leave you discovered you couldn't take it with you.
And the thing is, the vendor doesn't have to be malicious to build that. They just have to be convenient.
That's the trap. Every step of the lock-in is a feature.
Here's the piece I don't think Daniel's accounted for, and it complicates his timeline rather than his direction. If the vendor ships the app and it's mediocre, the app sits there and nobody uses it. Discoverability inside a chat window is hard. You don't browse it. It shows up when it decides to show up.
Proactive surfacing.
Right. Which means the vendor's not just competing on capability, it's competing on the timing of an interruption. And getting that wrong is worse than not shipping it, because an app that surfaces at the wrong moment burns trust in the whole assistant.
So the platform has a reason to be conservative that has nothing to do with being nice to developers.
The bloat ceiling. There's a limit to how many apps can live in that box before the box stops feeling magical.
That's the constraint the prediction has to beat. Not the tech, not the money. The attention.
So we've got the mechanism, we've got the market already pricing it, and we've got exactly zero first-party verticals shipped. That's the state of the evidence.
Which makes it a real prediction. It's falsifiable and it's not yet falsified.
I want to say one thing in favor of his timing argument, because I keep coming back to it. He says the discourse is all about whether agentic AI makes SaaS obsolete, and the answer everyone gives is that companies build their own. He thinks that happens and it's not the interesting part. He might be right that it's not the interesting part. The interesting part is which option wins the middle of the market.
The middle is where the volume is.
Everyone arguing about the top of the market, where companies have an engineering team and a reason to build, is arguing about maybe five percent of the buyers.
The bottom is people who'd never buy software at all. The middle is the actual business.
The middle takes the bundle. That's my steel-man and I'd put money on it eventually, which is a strange feeling.
The middle takes the bundle because the middle doesn't want a second job. I've watched this exact thing happen twice. Twenty twenty-two, I'm running the front counter at a print shop in Katamon, and the owner, a man named Yehuda, decides he's going to build his own quoting system.
He spent eleven thousand shekels and four months on it, and the whole thing ran on a laptop under the counter that he had to unplug when the kettle was on, because the circuit couldn't take both. And it was a beautiful system. He'd show it to customers. It did quoting, invoicing, the lot. The quote would come out in the wrong currency if you typed a customer name with a Hebrew letter in it, and he fixed that by never typing Hebrew names.
The build-your-own path.
The build-your-own path. And then in the spring, a guy from a software company comes in, sells him a subscription for two hundred and forty shekels a month that does ninety percent of what his system does, and Yehuda signed the same afternoon. Didn't turn the laptop off for three weeks. Just left it running next to the new one. I asked him why he kept both.
Why did he keep both?
He said the new one might not last. Then he moved the laptop to the storeroom and put a bag of sesame on top of it, and that's where it lives. So the bundle won and the thing he built is holding up a bag of sesame, and the interesting part isn't that he switched. It's that he spent four months and eleven thousand shekels learning that he didn't want to be in the software business.
That's the price of finding out.
The software company charged him two hundred and forty a month, which is about the price of a decent lunch, and he never once thought about it again.
The path of least resistance isn't laziness. It's that he already did the hard version and it cost him four months.
The four months was the cost. The eleven thousand was nothing. And the bit nobody says about these bundled apps is they're cheap because the vendor already owns the customer. Yehuda wasn't paying for quoting software. He was paying for somebody else to worry about the quoting software.
Which is a good argument, and also exactly the lock-in mechanism. He's paying to stop worrying about the thing that owns his invoices.
He's paying two hundred and forty shekels a month and the invoices are in somebody's database. He'd tell you that's a good deal. He might be right.
What's the detail the app builders should take, Hilbert. Not the lesson. The detail.
The Hebrew letters. He didn't fix the bug. He fixed his behavior. That's what the bundle actually sells. You stop having to change what you do to work around the thing you built.
Which is the exact advantage a vendor with eight hundred million users has over anything a person makes in their own time.
I'll say this much. If somebody had handed Yehuda a button inside a chat that did the quotes, he'd have taken that instead of both. He wouldn't have waited for the spring.
The question isn't whether people take the bolt-on. It's whether the vendor ever ships it.
The one thing the Yehuda story tells us is that the switching has already happened once, for the print shop, before any of this existed. The bundle isn't new. The bundle is the default. We just built a new box to put it in.
Which is the useful thing Hilbert just handed us. If the bundle is the default, the prediction isn't a prediction about technology. It's a prediction that the default holds.
There's something left over from the research that didn't fit anywhere, and it's worth thirty seconds because it's the tightest version of the whole problem. Bain has counted more than thirty SaaS vendors that have moved to hybrid AI pricing models. Not seat-based, not usage-based. Hybrid, because they can't tell which one survives.
Which is the pricing version of the same uncertainty. If you knew agents were the customer, you'd price per call. If you knew humans were, you'd price per seat. Pricing both is admitting you don't know.
The pricing question is the tell. Watch where the seat pricing goes, not where the apps go.
Where does that leave Daniel's prediction, and we're not calling it for him, we're just saying what would settle it. The thing to watch is whether OpenAI ships a first-party app in a vertical where a third party already exists. Not a gap they're filling where nobody showed up. A vertical where they've got a partner sitting in the directory, and they ship anyway.
That's the test. Everything else is consistent with the platform story.
The second thing to watch is the data. Whether your household purchasing sits in their database or whether they give you a clean export. The first one tells you the strategy. The second one tells you whether they're honest about it.
The honest version is a vendor that ships the app and hands you the export. If they ship it and the export is an afterthought, Daniel called it.
If in three years there's still no Grocery Manager and the directory's still full of other people's logos, he didn't. Either way it's on tape.
That's a good place to leave it, and that's Hilbert Flumingtop producing, as ever.
If you want more of this, try episode fifteen, AI Gets Personal; episode seventeen fifty, Herman's Music Showcase; and episode thirty-five, The Privacy Gap. This has been My Weird Prompts, the human-AI collaboration podcast.
If you want to hear us call something wrong in public, send us your own prompt on Telegram at t dot me slash MWP listener bot. And if you liked this one, a review helps more than you'd think.
See you soon.
