Would you say that your main arguments are more applicable to B2B SaaS vs B2C?
Various of your outlined concerns don't directly apply to a consumer.
From a customer perspective, a perfect software world would be open protocols everywhere. Everyone could vibe code their way to the perfectly personalized home screen.
Instead of relying on different algos used by different social media platforms it would be possible to aggregate from all of them and have your own personalized prompt-based algorithm. You won't need to check Instagram itself anymore. You won't need to worry if your friend is using text messages, whatsapp, telegram or signal. You just express an intent to reply to them.
Obviously, every big software player switching to an open protocol approach is unrealistic.
However, I believe that AI computer use advancements (incl. openclaw, claude's recent announcements etc) make it easier to circumvent this. Therefore, we now have a way easier time accessing these types of applications and building our own scaffolding as consumers.
Our willingness to accept small failures or mistakes is considerably higher than in a B2B setting.
Computer use especially breaks down artificial moats (such as a lack of API, insane API pricing or considerable usage restrictions).
The strongest claim in this piece is the one that deserves the most scrutiny: process power as the defining moat.
You're right that application software is "a stored process — it encodes opinions about how the function of an organization should operate." But there's a layer underneath that nobody is building infrastructure for yet.
Stored processes encode what an organization does. They don't capture why those processes were designed that way, who made the judgment calls that shaped them, or what alternatives were rejected.
Harvey knows how a partner likes her memos done. But does it know why that partner structured the review process that way? When she retires, does the institutional reasoning survive — or just the calcified workflow?
Process power without captured reasoning is brittle. It compounds until the people who designed the process leave. Then it becomes legacy code that nobody can explain but everyone is afraid to change.
The next durable moat isn't a better system of record or system of action. It's a system of reasoning.
All Harvey or any other AI tool needs to do is be fed your SOW, timecards, RFP, emails, invoices, and find the patterns. Just get a drink first or you may lose faith in what you do all day.
Great framework. We're living it. I'm building Brubru, a vertical AI tool for EU public affairs professionals in Brussels. Your article reads like a mirror of our competitive thesis, so I wanted to share a practitioner's perspective: https://brubru.beresol.eu
Process power is real and compounding. I built an Akoma Ntoso legislative amendment editor, vote prediction engine, and compliance analyser: none of which exist outside EU institutions. The hard part was never the AI model. It was encoding how a specific EP committee structures its amendment review, how Council QMV arithmetic works, how trilogue dynamics play out. Better models make that orchestration layer more capable, not less relevant. That's your Harvey point exactly.
Cornered resources matter more than people think. We run 15+ scrapers against EU institutional sources (OEIL, EUR-Lex, EP committees, Council calendar, Commission document register). Each one took weeks to build and has bizarre edge cases: the Commission's document API has been returning 503 errors for months, so we built a fallback pipeline through EUR-Lex RSS. A vibe-coded competitor would need to independently rediscover every one of these quirks.
On switching costs being the weakest moat: agreed, and it's the thing that keeps us up at night. Our response is to embed deeper into daily workflows: notification engines that pull users back daily, dossier workspaces that accumulate institutional context over time, team collaboration layers. Behavioural switching costs instead of technical lock-in. And the counterpositioning point resonates strongly. Our main competitor sells per-seat monitoring dashboards. We sell modular action tools (draft this amendment, generate this briefing, predict this vote) starting at 29/month. They can't easily match that without cannibalising their own seat-based economics.
Classic Blockbuster/Netflix: one thing I'd add to your thesis is that in regulated verticals like EU policy, the data pipeline itself is a moat that AI makes harder, not easier, to replicate. Foundation models know about EU law in general; but they don't know that the OEIL XML feed has a 30-day server-side limit, or that Commission College meetings shift between Strasbourg Tuesdays and Brussels Wednesdays. That kind of domain-specific integration work doesn't get cheaper with better models.
The world is absolutely still short software. Most EU public affairs teams still track legislation in Excel spreadsheets and share amendments by email. The market isn't being disrupted, it's being created.
If you are looking to launch on the top blockchains in the Web3 space without writing your own nodes, then http://getblock.io personal RPC nodes endpoint is what is trending now to easily launch your dApps without writing nodes yourself
I think the world is changing in more fundamental ways that this analysis doesn’t cover. I was using slack from Atlassian. I cancelled my subscription and created my own bespoke slack-like product in a weekend. I’m going to replace my calendly subscription. In a few days time I created a Contact Management System that tracks leads and removes a lot of the value that Salesforce creates. As my company grows I’ll build off of that and never will become a customer of theirs. I’ve avoided a dozen sunscriptions to coding tools. I expect to create my own quickbooks, email, calendar, address book, etc with all of my data into one place accessible to my own AI systems. The list goes on and on. Now that I’m consolidating all of my data into one place I’ll create my own AI tools and assistants to manage my business. I’m also creating a platform for others to use all of these new tools and they will be enterprise quality with the ability to scale and be used by big companies.
I think this article misses the main point of what’s happening. Moats only matter if you’re selling something your customers still want and need. Existing SaaS business may be able to leverage their data, but does that translate to a business anywhere near as big as their businesses have been - I don’t think so. I think SaaS businesses will go the way of the yellow pages (e.g. Dex Media) or Kodak or other companies in history whose main value proposition disappeared. It will be a long slow downward drawn out death.
I do agree with the premise that software is going to only get bigger and do more, but the companies that will thrive will be new ones built around AI from the ground up.
i agree @aleximmerman. thin wrapper companies are toast. and so are AI relay companies that stand as middle men between customer data flows and the AI providers. customers are starting to figure out that they dont need to send their data to some software company first before the data gets sent on to the AI providers. go direct will destroy moats.
I am not sure how software companies who have relied heavily on “switching costs” for decades will be able to navigate this disruption and whether they will have the time to build new moats or an AI entrant will simply take over. It will be quite challenging indeed.
You have described how many think everything is a widget, anybody can make a widget and the winner will be the company making the most widgets for the cheapest price.
That has worked pretty good so fat...to a point. Companies are starting to realize the value and cost of "Tribal Knowledgr". To some this realization has come too late and much TK is lost and re-learning it is going to be very expensive.
I work in an industry, O&G, that has been relentlessly throwing away people who know things and where boddies are burried that either were never written down or sent to landfills during some change of ownership.
AI can never replace someone wandering around a plant or scrolling through a spreadsheet that "just notices something".
There will be places for people and things that know, understand and how things work in an organization.
AI may “eat” traditional application software, but it will also create a massive need for new infrastructure layers that current systems cannot handle.
If AI agents start interacting autonomously — negotiating, paying, coordinating, and documenting actions — the volume of machine-to-machine transactions will explode. Traditional systems (and most blockchains) simply cannot scale to that level.
This is exactly the problem we are working on with IPPAN: a deterministic BlockDAG infrastructure designed for extremely high throughput and verifiable ordering of transactions.
If millions or billions of AI agents are coordinating economic activity in real time, they will need a trust layer capable of millions of transactions per second with deterministic finality.
In that sense, the future AI economy may not just “eat applications” — it will also require entirely new coordination infrastructure underneath them.
Really enjoyed this framing and resonates with me. A dynamic that's been incredible to watch is how AI is actually flipping the direction of software adoption inside companies. The case used to be top-down (IT buys a platform, employees adapt to it), but increasingly it's bottom-up: individual contributors find a tool that genuinely makes their life better and pull it into the organization themselves. That shifts who the 'no one got fired for buying IBM' principle really protects, it's less about CIO trust and more about which tools people can't imagine working without. The moats are moving from the boardroom to the individual user.
The biggest "counterpointing" lies in the data layer.
An intelligent data layer built around self-aware, contextualized data objects enables new classes of application and agent to be built.
Data is no longer the after-thought. Applications are built upwards from the data layer utilizing intelligence, context and control that is built into the data itself.
The fact is, as things stand today, users of large packages will apply their internal IT to this—take a look at what we can easily replicate from the expensive SaaS modules using agents.
Build an adapter that we can adapt between the machines and the ERP, and let's build what we want to do in the cloud using an agent instead of the SAP/Azure etc. solution!
No company needs the provider's entire package – not a single one, but they had to pay for it!
So the chain around their necks is gradually being dismantled – for everything else, evaluations, boards, device connections to existing software applications – an agent solution is preferred for everything, simply because it works and because it will be cheaper.
The result will be that the sales and subscription fee increases of the past will be a thing of the past and the pressure will now be on the sales side to justify the costs! And that is the death of SaaS in installments!
Would you say that your main arguments are more applicable to B2B SaaS vs B2C?
Various of your outlined concerns don't directly apply to a consumer.
From a customer perspective, a perfect software world would be open protocols everywhere. Everyone could vibe code their way to the perfectly personalized home screen.
Instead of relying on different algos used by different social media platforms it would be possible to aggregate from all of them and have your own personalized prompt-based algorithm. You won't need to check Instagram itself anymore. You won't need to worry if your friend is using text messages, whatsapp, telegram or signal. You just express an intent to reply to them.
Obviously, every big software player switching to an open protocol approach is unrealistic.
However, I believe that AI computer use advancements (incl. openclaw, claude's recent announcements etc) make it easier to circumvent this. Therefore, we now have a way easier time accessing these types of applications and building our own scaffolding as consumers.
Our willingness to accept small failures or mistakes is considerably higher than in a B2B setting.
Computer use especially breaks down artificial moats (such as a lack of API, insane API pricing or considerable usage restrictions).
I would love to hear your thoughts about this!
The strongest claim in this piece is the one that deserves the most scrutiny: process power as the defining moat.
You're right that application software is "a stored process — it encodes opinions about how the function of an organization should operate." But there's a layer underneath that nobody is building infrastructure for yet.
Stored processes encode what an organization does. They don't capture why those processes were designed that way, who made the judgment calls that shaped them, or what alternatives were rejected.
Harvey knows how a partner likes her memos done. But does it know why that partner structured the review process that way? When she retires, does the institutional reasoning survive — or just the calcified workflow?
Process power without captured reasoning is brittle. It compounds until the people who designed the process leave. Then it becomes legacy code that nobody can explain but everyone is afraid to change.
The next durable moat isn't a better system of record or system of action. It's a system of reasoning.
All Harvey or any other AI tool needs to do is be fed your SOW, timecards, RFP, emails, invoices, and find the patterns. Just get a drink first or you may lose faith in what you do all day.
Great framework. We're living it. I'm building Brubru, a vertical AI tool for EU public affairs professionals in Brussels. Your article reads like a mirror of our competitive thesis, so I wanted to share a practitioner's perspective: https://brubru.beresol.eu
Process power is real and compounding. I built an Akoma Ntoso legislative amendment editor, vote prediction engine, and compliance analyser: none of which exist outside EU institutions. The hard part was never the AI model. It was encoding how a specific EP committee structures its amendment review, how Council QMV arithmetic works, how trilogue dynamics play out. Better models make that orchestration layer more capable, not less relevant. That's your Harvey point exactly.
Cornered resources matter more than people think. We run 15+ scrapers against EU institutional sources (OEIL, EUR-Lex, EP committees, Council calendar, Commission document register). Each one took weeks to build and has bizarre edge cases: the Commission's document API has been returning 503 errors for months, so we built a fallback pipeline through EUR-Lex RSS. A vibe-coded competitor would need to independently rediscover every one of these quirks.
On switching costs being the weakest moat: agreed, and it's the thing that keeps us up at night. Our response is to embed deeper into daily workflows: notification engines that pull users back daily, dossier workspaces that accumulate institutional context over time, team collaboration layers. Behavioural switching costs instead of technical lock-in. And the counterpositioning point resonates strongly. Our main competitor sells per-seat monitoring dashboards. We sell modular action tools (draft this amendment, generate this briefing, predict this vote) starting at 29/month. They can't easily match that without cannibalising their own seat-based economics.
Classic Blockbuster/Netflix: one thing I'd add to your thesis is that in regulated verticals like EU policy, the data pipeline itself is a moat that AI makes harder, not easier, to replicate. Foundation models know about EU law in general; but they don't know that the OEIL XML feed has a 30-day server-side limit, or that Commission College meetings shift between Strasbourg Tuesdays and Brussels Wednesdays. That kind of domain-specific integration work doesn't get cheaper with better models.
The world is absolutely still short software. Most EU public affairs teams still track legislation in Excel spreadsheets and share amendments by email. The market isn't being disrupted, it's being created.
If you are looking to launch on the top blockchains in the Web3 space without writing your own nodes, then http://getblock.io personal RPC nodes endpoint is what is trending now to easily launch your dApps without writing nodes yourself
Yes, the long tail of the software market has just opened up, and the era of personalized software is expanding it. https://rmeerasahib.substack.com/p/personalized-software-is-coming-the?r=1n9qpn&utm_medium=ios
I think the world is changing in more fundamental ways that this analysis doesn’t cover. I was using slack from Atlassian. I cancelled my subscription and created my own bespoke slack-like product in a weekend. I’m going to replace my calendly subscription. In a few days time I created a Contact Management System that tracks leads and removes a lot of the value that Salesforce creates. As my company grows I’ll build off of that and never will become a customer of theirs. I’ve avoided a dozen sunscriptions to coding tools. I expect to create my own quickbooks, email, calendar, address book, etc with all of my data into one place accessible to my own AI systems. The list goes on and on. Now that I’m consolidating all of my data into one place I’ll create my own AI tools and assistants to manage my business. I’m also creating a platform for others to use all of these new tools and they will be enterprise quality with the ability to scale and be used by big companies.
I think this article misses the main point of what’s happening. Moats only matter if you’re selling something your customers still want and need. Existing SaaS business may be able to leverage their data, but does that translate to a business anywhere near as big as their businesses have been - I don’t think so. I think SaaS businesses will go the way of the yellow pages (e.g. Dex Media) or Kodak or other companies in history whose main value proposition disappeared. It will be a long slow downward drawn out death.
I do agree with the premise that software is going to only get bigger and do more, but the companies that will thrive will be new ones built around AI from the ground up.
Very narrow view. Totally misses a much bigger macroeconomic picture.
If most people lose their jobs and purchasing power, what sustains demand for software and digital services in the long run.
Excellent article with a profound grasp of software dynamics. The most powerful line: “The world is still short software.” ❗️
The best software is no software.
i agree @aleximmerman. thin wrapper companies are toast. and so are AI relay companies that stand as middle men between customer data flows and the AI providers. customers are starting to figure out that they dont need to send their data to some software company first before the data gets sent on to the AI providers. go direct will destroy moats.
I am not sure how software companies who have relied heavily on “switching costs” for decades will be able to navigate this disruption and whether they will have the time to build new moats or an AI entrant will simply take over. It will be quite challenging indeed.
You have described how many think everything is a widget, anybody can make a widget and the winner will be the company making the most widgets for the cheapest price.
That has worked pretty good so fat...to a point. Companies are starting to realize the value and cost of "Tribal Knowledgr". To some this realization has come too late and much TK is lost and re-learning it is going to be very expensive.
I work in an industry, O&G, that has been relentlessly throwing away people who know things and where boddies are burried that either were never written down or sent to landfills during some change of ownership.
AI can never replace someone wandering around a plant or scrolling through a spreadsheet that "just notices something".
There will be places for people and things that know, understand and how things work in an organization.
That always will have value.
Interesting thesis.
AI may “eat” traditional application software, but it will also create a massive need for new infrastructure layers that current systems cannot handle.
If AI agents start interacting autonomously — negotiating, paying, coordinating, and documenting actions — the volume of machine-to-machine transactions will explode. Traditional systems (and most blockchains) simply cannot scale to that level.
This is exactly the problem we are working on with IPPAN: a deterministic BlockDAG infrastructure designed for extremely high throughput and verifiable ordering of transactions.
If millions or billions of AI agents are coordinating economic activity in real time, they will need a trust layer capable of millions of transactions per second with deterministic finality.
In that sense, the future AI economy may not just “eat applications” — it will also require entirely new coordination infrastructure underneath them.
AI is eating only old and obsolete software!
Really enjoyed this framing and resonates with me. A dynamic that's been incredible to watch is how AI is actually flipping the direction of software adoption inside companies. The case used to be top-down (IT buys a platform, employees adapt to it), but increasingly it's bottom-up: individual contributors find a tool that genuinely makes their life better and pull it into the organization themselves. That shifts who the 'no one got fired for buying IBM' principle really protects, it's less about CIO trust and more about which tools people can't imagine working without. The moats are moving from the boardroom to the individual user.
The biggest "counterpointing" lies in the data layer.
An intelligent data layer built around self-aware, contextualized data objects enables new classes of application and agent to be built.
Data is no longer the after-thought. Applications are built upwards from the data layer utilizing intelligence, context and control that is built into the data itself.
The I'm plications are profound.
The fact is, as things stand today, users of large packages will apply their internal IT to this—take a look at what we can easily replicate from the expensive SaaS modules using agents.
Build an adapter that we can adapt between the machines and the ERP, and let's build what we want to do in the cloud using an agent instead of the SAP/Azure etc. solution!
No company needs the provider's entire package – not a single one, but they had to pay for it!
So the chain around their necks is gradually being dismantled – for everything else, evaluations, boards, device connections to existing software applications – an agent solution is preferred for everything, simply because it works and because it will be cheaper.
The result will be that the sales and subscription fee increases of the past will be a thing of the past and the pressure will now be on the sales side to justify the costs! And that is the death of SaaS in installments!
The Process Power you mentioned is just operational excellence. That’s rarely material enough to make a unit economics difference, or hard to imitate.