Software Is No Longer a Product

For a while, artificial intelligence appeared to pose an existential problem for SaaS.

The logic was persuasive. Why keep paying a monthly subscription for a relatively simple piece of software when an AI agent can write something similar for you? A small CRM, an internal dashboard, a scheduling system, a reporting application or a specialized administrative tool no longer necessarily requires a software company behind it. Increasingly, it can be assembled from prompts, APIs and generated code.

The uncomfortable question for the SaaS industry was obvious: if software becomes cheap to create, what exactly are we subscribing to?

That question has not gone away. In fact, the pressure on conventional SaaS is becoming more visible. Software built around a narrow collection of features, a proprietary interface and a per-user subscription suddenly looks much less defensible when an agent can reproduce a significant portion of its functionality.

But something interesting happened at the same time.

AI created another generation of SaaS.

ChatGPT is SaaS. Claude is SaaS. Suno is SaaS. Cursor operates substantially through a subscription service. Perplexity does too. Runway provides sophisticated generative video capabilities through the same basic relationship: create an account, choose a plan, use a capability maintained somewhere else.

They look different from Salesforce, Dropbox or the SaaS products of the previous decade, but economically the family resemblance is unmistakable.

And their success suggests that SaaS itself was never the problem.

What changed was what the service provides.

From software to capability

Traditional SaaS generally gave the customer access to an application.

The provider developed the software, hosted it, updated it, secured it and made it available through a browser. The customer avoided installing and maintaining a local system and paid continuously for that convenience.

AI-native SaaS goes considerably further.

When someone subscribes to Suno, they are not really paying for a music application. The visible application is almost incidental. They are buying access to the capability to generate music.

The same distinction applies to ChatGPT. The value is not primarily the chat window. Anyone could build a chat interface. The value is everything operating behind it: models, inference infrastructure, tools, memory, multimodal processing, orchestration, continuous model improvements and the enormous operational machinery required to make those capabilities available within seconds.

Cursor is not valuable because someone invented another text editor. Its attraction comes from bringing continuously evolving models into the software-development workflow.

This is a fundamentally different proposition.

The old SaaS model could often be summarized as software you do not have to run yourself.

The emerging model is closer to capability you cannot reasonably maintain yourself.

Yes, you could build it

This is where the apparent contradiction disappears.

AI dramatically lowers the cost of writing software. It does not necessarily lower the cost of providing a sophisticated AI service.

A competent developer can create an interface connected to a language model remarkably quickly. A company can build its own internal AI application. An individual can automate workflows that would once have justified purchasing another subscription.

That is real, and it will eliminate some SaaS products.

But reproducing Suno is not the same as creating a web page that sends a prompt to a music model. Reproducing ChatGPT is not the same as connecting an application to an LLM. Operating a production service requires models, compute, storage, inference optimization, evaluation, security, identity, billing, failure handling, scaling and constant adaptation as the underlying technology changes.

The code around the service may actually become less important.

The service itself becomes more important.

That distinction creates a useful dividing line between the SaaS businesses that AI threatens and those it strengthens.

If the principal value of a product is its interface and a collection of relatively reproducible business rules, AI is dangerous.

If its principal value is providing continuous access to a difficult capability, AI can make SaaS extraordinarily attractive.

The disappearing seat

There is another important change.

For twenty years, one of the most successful concepts in SaaS was the seat.

Ten employees use the software, so the company buys ten licenses. Hire another employee and revenue expands almost automatically.

Agents disrupt that relationship because the entity performing the work may no longer correspond to a human sitting in front of a screen. Analysts are consequently questioning the durability of seat-based pricing as software moves from assisting workflows toward executing them.

AI-native services are already experimenting with a different vocabulary: tokens, generations, credits, compute, tasks and usage limits.

Suno sells the ability to produce a certain quantity of creative work. AI platforms meter computational consumption. Other services increasingly charge according to executions or outcomes.

The unit of software economics is gradually moving away from the user.

It is moving toward work performed.

That may eventually prove to be a much larger market.

The interface becomes secondary

There is an additional irony here.

For traditional SaaS, the application interface was a significant part of the product. Companies invested enormous amounts in dashboards, navigation systems, forms and workflow design.

With AI, some of that value migrates underneath the interface.

A user may simply describe what needs to happen.

The system determines how.

That makes the visual application thinner while making the infrastructure behind it substantially more sophisticated. SaaS does not disappear; it moves down the stack.

This also explains why established systems of record are proving harder for agents to eliminate than some early predictions suggested. Agents still require reliable places to retrieve information, execute transactions, enforce permissions and preserve an audit trail. In many enterprise environments, AI is being layered over those systems rather than replacing them outright.

The valuable SaaS platform of the next decade may therefore be something humans barely interact with directly.

Its customers could increasingly be other software and agents.

A different moat

Cheap software creation also changes what constitutes defensibility.

Code alone is becoming a weaker moat.

A company cannot assume that having spent three years building an application gives it three years of protection when competitors can now reproduce significant portions of software in weeks or days.

The durable assets move elsewhere: proprietary data, distribution, workflow integration, model orchestration, reliability, specialized knowledge, infrastructure, trust and the accumulated understanding of what users are actually trying to accomplish.

This is already visible in the emerging divide between generic AI applications and deeply embedded vertical systems. Ease of building increases competition; it does not eliminate the advantages of products that own valuable data or become tightly integrated into customer workflows.

The strange consequence is that AI simultaneously makes software easier to replace and services harder to reproduce.

Software after software

The first generation of SaaS transformed ownership into access.

The emerging generation is performing another conversion.

Access is becoming capability.

Instead of buying software that allows us to perform a task, we increasingly subscribe to a system capable of performing some portion of the task itself.

That is why the predicted death of SaaS now looks premature. There will certainly be casualties. Thin applications whose value consisted largely of packaging relatively simple operations behind a convenient interface have a serious problem.

But Suno, ChatGPT, Claude and the growing collection of AI-native services reveal another trajectory.

We may build more of our own software than ever before.

And simultaneously subscribe to more services than ever before.

The distinction is that we will not necessarily be paying them for their software.

We will be paying them for what their machines can do.