For years, we made a strange compromise with software.
We bought products that were 70% right. Then we changed our processes, hired people to operate the tool, built spreadsheets around it, and called the remaining 30% “configuration.”
That was acceptable when building software was expensive.
It is becoming much less acceptable now.
My current belief is simple: the era of mass SaaS is ending.
Not all software. Not every software company. Marketplaces, networks, systems of record, regulated infrastructure, products with proprietary data, and products that coordinate many parties will remain very hard to replace.
But generic workflow software, the kind sold to thousands of businesses with essentially the same pitch, is entering a brutal repricing.
The reason is not that AI will make software disappear. The reason is that AI is making custom software cheap enough to be the default answer again.
The old bargain is breaking
The traditional SaaS model worked because the alternative was painful.
If you wanted a custom internal tool, you needed a product manager, designers, engineers, a budget, and months of work. Even then, the tool would be frozen at the moment you shipped it. Every new exception became a ticket. Every workflow change became a project.
So companies accepted the mass product. They adapted their business to the tool because the tool was cheaper than building something that fit.
AI changes that equation.
A small team can now understand a workflow, build an interface, connect systems, create automations, and iterate with the people actually doing the work at a speed that would have sounded ridiculous two years ago.
When a company can have software that speaks its language, follows its process, understands its customers, and changes next week when the process changes, “best practice” software starts to look like a constraint.
Users do not really want more features. They want the software to fit their business, instead of adjusting their business to fit the software.
That is a much bigger shift than a new UI or a better chatbot.
The pricing power test
We recently went through the familiar exercise of reviewing a major SaaS contract.
The initial annual price was $80,000.
We asked for $20,000.
They offered $30,000. We said no.
They offered $20,000. We said no again.
They eventually came back with a number below $10,000.
By then, we had already moved away.
This is not a story about one vendor negotiating badly. It is a signal.
When a company believes it can replace a generic workflow with a combination of its own systems and AI, the vendor no longer has the same leverage. The switching cost falls. The “standard” feature set matters less. And the price that used to look normal suddenly looks absurd.
This will happen across categories.
Not overnight. Enterprises still have contracts, compliance requirements, messy data, legacy systems, and very rational fear of breaking critical processes. But those are implementation problems, not proof that the old model is safe.
The product is no longer about writing code
There is another uncomfortable implication for product companies.
When everyone can create software, shipping software is no longer the scarce skill. Not one that Silicon Valley has any edge over.
The scarce skill is product discovery: finding the real problem, understanding how work actually gets done, deciding what should change, and staying close enough to the customer to keep making the system useful.
This is why I am skeptical that standalone agents will capture most of the value on their own.
An agent is not a business transformation. A model is not a workflow. A demo is not adoption.
The hard part is getting inside a company: understanding the incentives, the data, the handoffs, the exceptions, the politics, and the real definition of a good outcome. Then building and operating the system until it actually produces that outcome.
That requires people who can do the work with the customer, not just sell a licence and send documentation.
Call them forward-deployed engineers. Call them AI operators. The title matters less than the operating model.
The winning teams will have to be deeply embedded in the business. They will discover the problem, build the custom system, connect it to reality, and continuously improve it. In many cases, they will be paid from a labour or transformation budget because that is where the economic value is created.
This looks more like services. But it is not traditional services.
Traditional services scale by adding people and billing hours. The AI-native version should scale by delivering outcomes with very small teams, while reusing the infrastructure, patterns, and learning gained from each deployment.
That is where I think a lot of the money will move: away from generic SaaS subscriptions, open-source projects without a clear business model, and horizontal tools that can be copied quickly, and toward AI transformations that change how a business operates.
What still has a moat?
I do not think every company should build everything itself. That would be another expensive mistake.
Some software will become more valuable in this world, not less:
Marketplaces and networks, where the value comes from participants and liquidity.
Systems of record and regulated infrastructure, where trust, reliability, compliance, and historical data matter.
Products with unique proprietary data or a hard-to-recreate distribution advantage.
Software that coordinates many independent parties, rather than merely digitising one company’s internal workflow.
These are real moats. A prettier interface is not.
For everyone else, the question gets harder: why should this be a product sold to the whole market, rather than a capability built specifically for one business?
The next generation of great products
I do not think this is the end of product building.
It might well be the golden age of products.
I think it is the end of pretending that a generic product is always the highest form of it.
The next generation of great products may be multiple products, built continuously, close to the people doing the work. They may begin as a custom system inside one company and become reusable only after the underlying problem is deeply understood.
That is less romantic than the old SaaS story: build once, sell forever, add seats, watch the chart go up.
It is also more honest.
AI gives every company the ability to build more of its own software. The companies that win will not be the ones with the most impressive agent demo. They will be the ones that understand a business well enough to make AI useful inside it.
Mass SaaS is not going to vanish tomorrow.
But the default assumption is changing: software should adapt to the business, not the other way around.
this felt like reading Seth Godin just a bit longer.