"SaaS Is Dead" Is Half Right

In late 2024, Satya Nadella suggested on a podcast that traditional business applications were “just CRUD databases with business logic” and that AI agents would dissolve them into something new. The clip went everywhere; investors debated it, product managers forwarded it, and a thousand social posts declared the death of SaaS.

Most of them were wrong. A few were right. And the difference matters a great deal if you build, buy, implement, or invest in enterprise software.

The uncomfortable truth is this. SaaS, as a category, is not dying. But SaaS as we have known it for twenty years, per-seat licensing, forms and menus, humans driving workflows through software, is being unbundled.

And it has happened before.

Software has an unbundling cycle. It runs roughly every fifteen to twenty years, and it looks the same each time.

In the 1990s, large enterprises ran monolithic ERPs. One vendor, one suite, one throat to choke. The critique at the time was that these systems were too big, too rigid, too expensive, and too slow to change. They were.

The first wave of SaaS unbundled the monolith. Salesforce did not try to replace SAP. It took one slice, CRM, put it in a browser, and charged per seat. Workday did the same for HR. NetSuite for finance. ServiceNow for IT operations. The monolith came apart into dozens of specialised clouds, and the buyer’s problem flipped from “customise the beast” to “integrate the best-of-breed.”

Then came the rebundle. Suites returned, not as monoliths but as platforms. Salesforce became a platform. Microsoft rebuilt around Dynamics and later the Power Platform. App stores and marketplaces quietly reassembled the stack. By the late 2010s, the enterprise no longer bought thirty apps. It bought three platforms and two hundred apps running on them.

That is the pattern. Unbundling is what the industry talks about. Rebundling is where the money is made.

Call it the unbundling cycle. It is the single most important lens for understanding what AI is actually doing to enterprise software, and it is almost entirely absent from the current debate.

If you put the hot takes down and look at what AI changes under the hood, three things seems to be moving.

1-The interface layer. Menus, forms, and dropdowns exist because scarce human attention needs scannable structure. They are an artefact of constrained compute. Once software can be asked in natural language and act, the menu becomes optional, then burdensome, then absurd. The UI of SaaS will erodes first.

2-The workflow layer. The same reason people still run half of finance out of Excel is the reason workflows sprawl across apps. No one designs them, they accrete. An agent that can read across those apps and stitch workflows together needs an API (read: MCP). This is where per-seat pricing begins to buckle.

3-The integration layer. Fifteen years of SaaS produced an industrial ecology of iPaaS vendors, connectors, and middleware that exists only because apps could not talk to each other. Agents with tool use collapse much of that plumbing. They will compress, but not eliminate it.

Now notice what has not moved.

The system of record still holds the canonical data. Compliance and audit still demand trails. Enterprise buyers still want SLAs, contracts, and someone to sue. The expensive, hard parts of enterprise software are not the parts AI makes trivial.

This is why “SaaS is dead” is the wrong framing. The parts of SaaS that are dying are the parts that were always going to die. The parts that persist are the parts that were always the product.

Every unbundling reorganises around a new primitive. In the ERP-to-SaaS cycle, it was the module. The CRM. The HR cloud. The ITSM suite. Each one was a unit of software that a buyer could evaluate, procure, and adopt independently.

The primitive this cycle is the agent.

Not “an AI feature inside my existing app.” An agent. A persistent, goal-directed actor that can read, reason, call tools, and take consequential action. It has a name. It has permissions. Eventually it has a budget and a manager.

If you take this seriously, the implications are awkward for most current product roadmaps. The app is no longer the destination. It is a capability the agent calls. The UI matters less. The API matters more. The per-seat licence makes no sense to something that does not have a seat.

It is a frame shift, not another feature.

And it is where most “AI-powered” SaaS pitches in 2025 and 2026 quietly break down. Bolting a chat panel onto the side of an app treats the agent as a feature. The agent is not a feature. It is the centre of the new architecture, and the app is the feature.

The interesting question is not what unbundles. It is what rebundles, and around what.

Three candidates are already clear.

The horizontal assistant platform. Microsoft is the obvious contender, with Copilot threaded across the productivity surface. The bet is that the agent lives inside the workspace, with access to mail, documents, calendars, and chat, and everything else becomes a plugin. It is a platform play against vertical apps. It works, or it does not, depending on whether the assistant can credibly reach into systems of record without becoming a lowest-common-denominator interface.

The vertical workflow operator. For industries with real domain depth, mining safety, clinical coding, superannuation administration, construction claims, the winners will be agents trained on the vocabulary, data, and constraints of the domain. These are actual operators, not mere chatbots. They compete not with horizontal assistants but with consulting hours and offshore back offices. Australia, with its concentrated industry structures, is unusually well-positioned for this.

The agent-native back end. The least visible but most interesting category. New companies are being built whose product is not a UI at all. It is a set of tools, policies, and evaluations that other agents consume. Think of it as the Azure / AWS / GCP layer of the agent economy. It is early, it is fragmented, and it is where a disproportionate share of the enterprise value in this cycle will accrue.

None of these replaces SaaS. All of them absorb its workflows.

A few concrete implications follow, and they are worth sitting with.

If you build software. If the product assumes a human is the user, it is being built for the last cycle. Design for an agent as the primary caller and a human as the supervisor. This changes everything about API shape, rate limits, audit logs, and error handling. The “user” in the analytics is increasingly not a user.

If you buy software. The next two years of procurement will be bruising. Every incumbent will add an AI SKU at a twenty to forty per cent uplift. Most of that uplift will not be justified by measurable outcomes. The question to ask any vendor showing a Copilot-style add-on is simple. Does this remove work, or does it just add a new tab? If they cannot answer in hours saved per workflow per month, walk.

If you invest. The SaaS comps in current use are about to mislead. A forty per cent gross margin AI-application company is not a broken SaaS business. It is a new category with different multiples. The sooner the market prices this honestly, the sooner capital flows to the right places. That is the subject of the next piece.

If you operate inside an incumbent. The worst strategy is to treat this as a feature release cycle. The second worst is to treat it as a complete rebuild. The right strategy, almost always, is to pick one workflow, instrument it honestly, and let the agent architecture grow from there. Patience is a competitive advantage in a market full of theatre.

Here is the part that does not make it into conference keynotes.

Unbundling cycles are brutal on incumbents but rarely terminal. SAP survived the SaaS wave. Oracle survived. IBM, despite everything, survived. They shrank in relative importance. They lost category leadership. They reorganised. They did not disappear. Enterprise software moats are built from things AI does not change: (a) Distribution relationships, (b) Procurement inertia, and (c) Compliance baggage.

So the honest conclusion is that SaaS, as a business model, will persist for longer than the hot-take crowd believes, and change more than the incumbent crowd believes. Both can be true. Both usually are.

What dies is a specific version of SaaS. Menu-driven, per-seat, human-piloted software sold to a named buyer for a named workflow.

What lives is software as enterprise substrate. APIs, data, policies, and audit trails, consumed increasingly by agents on behalf of humans.

The unbundling cycle is back. It will take longer than the bulls think and move faster than the bears believe. By the time there is a consensus name for what rebundles, the next cycle will already be starting.

In the next piece in this series I want to take on the thing the “SaaS is dead” debate keeps avoiding. What happens to eighty per cent gross margins when inference is not free.

This is the first article in a series on how enterprise SaaS is being reshaped by AI.