SaaS gave birth to AI

Artificial intelligence is often pitted against the subscription model, as if the former were disrupting the latter. Publishers who spent a decade migrating their products to SaaS now hear that the real disruption lies elsewhere, and that they may have chased the wrong train.

It is almost the opposite. Without twenty years of SaaS and cloud, there would be no industrial AI integrated into software as we are beginning to see it. There would be powerful models, impressive demonstrations, and capabilities accessible via API. But far less artificial intelligence embedded in line-of-business products at scale, fuelled by real-world enterprise context and continuously improved. SaaS is not the victim of the current wave: it helped build part of its foundations.

This claim deserves to be demonstrated rather than asserted. It changes what a publisher must do this year.

Sixty years of waiting, and two locks

Artificial intelligence did not begin in 2022. The Turing test dates back to 1950, and the first conversational agent to the 1960s. At the time, some already predicted that human-level intelligence would be replicated within a few decades.

It was not, and the reasons are well documented. Early approaches were largely symbolic — rules were written by hand, one by one, in the hope that they would eventually cover enough situations to mimic intelligence. They ran into the complexity of the real world. Two essential resources were also missing: computing power and large-scale exploitable data. The subsequent periods of disillusionment remain etched in the field’s history as the AI winters.

Then a computer beat the world chess champion, statistical methods and machine learning advanced, and AI worked — but often in specific cases, on closed problems, and with carefully curated datasets.

The first lock gradually fell with the explosion of computing power, notably thanks to graphics processors: thousands of operations executed in parallel, calculations that were once out of reach now industrialised.

The second lock is more interesting for publishers: some of the data and infrastructure that today’s line-of-business AI needs were organised while we pursued a very different goal.

The second lock was US.

Large generative models learn from vast corpora of text, code, images, and other content from multiple sources. But to turn a general-purpose model into a line-of-business product, something else is required: giving it access to context. To customers, transactions, documents, histories, workflows, permissions, past decisions, and, progressively, the results of its own actions.

The question then becomes: why is an increasing share of this data now available in a form that is accessible, structured, and exploitable by software?

Because for twenty years, the software industry has progressively moved enterprises towards a model where applications, data, and usage are increasingly centralised, connected, and continuously accessible.

On-premise software also produced data. But it was scattered across thousands of installations, sometimes in different versions, behind different systems and firewalls. Aggregating it, observing it, and exploiting it at scale was infinitely harder.

That gap was not bridged for artificial intelligence: it was bridged to reduce costs, simplify operations, improve availability, and accelerate software delivery.

The shift to SaaS and cloud changed this without it being the objective. It delivered six things:

• Availability

  Data is there, continuously — not in a quarterly backup.

• Aggregation

  A thousand customers on one platform make a corpus. On a thousand installations, a thousand silos.

• Learning

  A model needs volume and diversity: exactly what a shared platform provides.

• Cost control

  Cloud elasticity made previously lab-only computation affordable.

• Scalability

  Delivering a capability to an entire user base at once, rather than installation by installation.

• Industrialisation

  Continuous delivery, supervision, measurement: deploying an imperfect capability and correcting it quickly.

None of these six conditions were built for artificial intelligence. All become critical when transforming AI into a product.

Thirty years of industrialisation, and a shift in purpose

There is a simple way to summarise the movement.

For thirty years, we have industrialised software: moving it from craftsmanship to reproducibility, scalability, observability, and large-scale exploitability. This took time.

1. 1970-1985 — Centralisation
2. Computing resides at the centre. The user sits at the end of a terminal with very little local power.
3. 1985-1995 — Client-server
4. Power moves to the desktop. Software is installed and must be deployed, maintained, and updated installation by installation.
5. 1995-2005 — Web
6. The browser gradually becomes the interface. The server regains a central role and software begins to be delivered remotely.
7. 2005-2020 — SaaS and cloud
8. A single platform can serve thousands of customers. Updates become continuous. Data and usage become observable at a scale that did not previously exist.

Each of these four steps was taken for reasons largely unrelated to artificial intelligence: reducing costs, simplifying operations, accelerating updates, improving availability. The result, however, was the same — bringing software, data, users, and systems closer together.

Today begins the industrialisation of intelligence: making reproducible and scalable a capability that no longer merely executes written rules, but interprets a situation, proposes a decision, and increasingly executes an action.

The first project was a prerequisite for the second. It is hard to industrialise a probabilistic system if you cannot already deploy continuously, measure its behaviour, observe its errors, and correct them quickly.

SaaS is the body of software; artificial intelligence can now become its brain. A brain without a body can reason. It cannot act.

Software ceases to be a tool

There is a second disruption, and it is not about the origin of the wave but about the nature of what is delivered.

For thirty years, most line-of-business software has been deterministic: rules were written, and the software executed them. Its quality was largely measured by predictability — the same input had to produce the same output, and unexpected behaviour was generally a defect to reproduce, identify, and correct.

What is happening now is different in nature, and the difference lies in the form of the output.

This difference underpins much of the rest — and it runs deeper than it appears. A wrong answer is no longer necessarily a reproducible bug. It may stem from the model, the context provided, the data retrieved, poor orchestration, or a probabilistic decision that will differ at runtime. Correction is no longer just about delivering a patch: it also involves what the model sees, the tools it accesses, the rules governing its actions, and the situations in which a human must take over.

The consequences are visible role by role.

Software is no longer just a tool that is used. It becomes an actor that executes. And an actor is not sold, tested, contracted, or billed like a tool.

What this means for a publisher this year

From this genealogy comes a direct consequence, and it is uncomfortable.

A publisher still predominantly on-premise can, of course, integrate artificial intelligence. But it starts with a handicap: it is harder to aggregate context, observe usage, measure outcomes, rapidly improve the system, and deliver improvements simultaneously to its entire user base. It can buy a model; making it an industrial learning loop and a durable competitive advantage will be far harder.

Three questions allow an honest assessment, and they can be answered in half a day:

• What share of your user base is truly mutualised?

  Not “available in SaaS”: truly migrated. It is this figure, not the existence of an offer, that measures your ability to observe and industrialise usage.

• Are your data aggregable?

  Legally, technically, and semantically. Data that exists but cannot be linked or used contractually remains a silo.

• How long does it take to deliver to your entire user base?

  If the answer is measured in quarters, the learning loop that drives the value of an AI-augmented product will turn too slowly.

Those who answer these questions well will accelerate. Others risk falling behind — not because they lack access to artificial intelligence (the same models will be accessible to almost all), but because they lack the foundations to turn it into a product and a competitive advantage.

The next cycle is already underway

This genealogy is not a retrospective. It serves to situate what is coming, and what is already here for some.

Three tiers stand out, and they do not require the same maturity:

• Copilots and automation

  The system advises, drafts, summarises, or executes certain steps. The human retains decision-making and largely remains in the loop.

• Agents

  A task can be executed end-to-end. The human sets the objective and supervises the outcome rather than each step. Software truly begins to become an actor.

• Autonomous and adaptive systems

  The system observes, decides, and executes within a defined framework; the human intervenes mainly on rules and exceptions. This tier demands all the foundations — data, integrations, observability, trust — and will be the hardest to catch up.

An American investment house summed up the economic shift in a sentence we have not improved upon:

This captures the shift in foundation: we are progressively moving from billing solely for a user to billing for usage, a workflow, a transaction, or an outcome. Seat → Usage → Workflow → Transaction → Outcome. And an outcome is measurable — therefore contestable, therefore provable.

This creates a paradox for the SaaS publisher. If its product tomorrow enables fifty people to do the work of a hundred, it has created enormous value for its customer. But if it continues to bill solely by number of users, it may simultaneously shrink its own revenue base.

Publishers venturing into this space discover that they have changed business before they have changed pricing: the challenge is no longer just creating value with AI, but knowing how to capture some of it.

And then

This genealogy explains where the wave comes from. It does not yet say everything it changes about what you sell.

Because if software ceases to be a tool that is used and becomes an actor that executes, then the commercial promise changes in nature: you are no longer just renting a tool to do the work, you are beginning to sell a system that performs part of it. This is the shift from Software as a Service to Service as a Software — and it moves pricing, value measurement, and the customer relationship.

It also shifts the competitive advantage. If all publishers gradually gain access to the same large models, the moat does not necessarily reside in the model itself. It moves to what the publisher owns around it: distribution, line-of-business data, workflows, integrations, and trust. A general-purpose model knows a line of business. A vertical publisher knows the files, transactions, histories, rules, exceptions, and permitted actions of thousands of customers. The model may become a commodity; the line-of-business context, however, can become a strategic asset.

This is the essence of Service as a Software: what AI truly changes about the promise of software.

One final remark, to end where we began. Artificial intelligence is often presented as a threat to those who work. The reality is more precise: AI does not decide what happens to the time it frees up. It makes that time available. A company can convert that time into growth, service quality, additional capacity, margin — or headcount reduction. What it does with it next is a question of management, not technology.