News Analysis: Yes, Software is Still Eating the World

$1 Trillion wipeout in a week

Enterprise software stocks had been drifting for months as investors quietly questioned whether the traditional SaaS model was sustainable. Then the drift turned into a snap.

It took nearly 2 years to erase $5 trillion of market value during the dot-com crash, with Nasdaq falling almost 80% from peak to trough. In early February 2026, it took roughly a week of trading to wipe out nearly $1 trillion of value from enterprise software and related stocks.

Markets rarely reprice an entire category that quickly unless something structural has changed. Sparked by a wave of highly autonomous agent releases, investors began recalculating the economics of the traditional SaaS model itself.

Seat-based model was always linear

For most of the SaaS era, enterprise software scaled on a simple premise: add features, add users, add revenue. Subscription pricing made growth predictable and compounding-like, which justified premium multiples.

Yet the mechanism carried a constraint that was easy to ignore in a rising market. Seats scale with labor, and labor is the cost base.

Despite years of investment in automation, most systems did not eliminate work so much as reorganize it. Processes became digital, but someone still had to interpret information, make decisions, and execute each step. As customers grew, they bought more licenses and hired more operators to run them. Revenue increased, but labor rose alongside it, keeping output tied to human effort.

What looked like leverage was linear economics.

The seat model wasn’t flawed; it was extraordinarily effective for its era. Predictable recurring revenue, massive cloud distribution, and network effects drove explosive growth and justified 10 to 20x revenue multiples. But in a world where AI can decouple output from labor, that same linearity now looks like a ceiling.

As investors recognized that growth remained coupled to labor, the valuation math became harder to defend.

Breaking the constraint

If growth is constrained by labor, the only durable way to expand capacity is to reduce how much human execution each workflow requires. That is the constraint AI begins to break.

Earlier platform shifts improved delivery but left execution intact. Cloud simplified deployment and adoption, APIs eased integration, and mobile expanded access, yet humans still performed the work. The boundary between what software could do and what required human execution remained fixed.

AI moves that boundary.

Modern models can interpret messy inputs, reason through ambiguity, plan actions, and operate directly across systems within defined policies. In routine workflows, they are reliable enough to act on their own.

Software that assists makes humans faster. Software that executes changes the role of humans, moving them from operators to governors of the system.

As execution shifts to machines, output scales with compute rather than labor. Because compute costs fall while labor costs do not, margins expand.

Assistance is not autonomy

That distinction explains why many “AI-powered” tools still feel incremental. Copilots draft, summarize, and recommend, yet humans approve every action, so work moves faster while the operating model stays the same.

Autonomy changes the structure.

In an autonomous system, software monitors conditions, adjusts plans, triggers actions, and resolves most routine decisions without waiting for human input, with humans remaining explicitly in the loop to set policy, supervise outcomes, and intervene where judgment matters.

Instead of manually triaging hundreds of routine operational decisions each day, including production adjustments, supplier exceptions, quality alerts, and compliance checks, the system resolves the vast majority automatically and escalates only the few that truly require judgment. Execution scales with machines, while accountability stays with humans.

Intelligence without context can’t act

Reliable autonomy depends on more than intelligence. Enterprise decisions rarely live inside a single application; they depend on context scattered across operational systems, financial systems, documents, and policies. Without understanding how those pieces connect, even the best model can suggest but cannot safely act.

Traditional software required deterministic inputs. Data had to be clean, structured, and standardized. When reality was messy, humans translated it first, normalizing names, categorizing requests, and filling out fields. Software could only process what fit its schema.

AI-native systems work differently. They can ingest messy reality directly - unstructured emails, inconsistent data, ambiguous requests - and still produce deterministic, reliable outputs.

To execute safely at scale, these systems need a unified view of the business itself: how assets, constraints, and processes relate. A coordinating layer is therefore emerging above existing applications that builds this shared understanding - essentially a business graph that maps how everything in the enterprise connects - and orchestrates actions across the stack.

Over time it begins to function less like another tool and more like an operating system for the enterprise, directing work continuously rather than waiting for humans to stitch systems together. Because every action generates feedback and every outcome refines policy, the system improves with continued use. Context compounds, autonomy expands, and the gap between organizations that build this capability and those that do not widens.

Trust is the gate

Autonomy only works if it is trustworthy.

When a human executes a task, accountability is implicit. When software executes thousands of decisions a day, accountability has to be designed. Humans must define the policies. Humans must authorize what the system is allowed to do. Every action must be logged, explainable, and traceable back to the data that informed it. High-risk actions require explicit human approval. And the system itself must be continuously tested so errors are caught early rather than compounded at scale.

These safeguards are not features; they are foundational. They allow organizations to delegate routine execution while retaining control over risk and responsibility.

As execution shifts into software, humans move up the stack into governance, setting policy, supervising outcomes, and remaining accountable, while software handles the mechanics. Seats per employee matter less than throughput per employee.

Where value compounds

As models improve and costs decline, model capability becomes abundant and general reasoning commoditizes.

But context does not.

Durable advantage accrues to whoever grounds intelligence in deep, enterprise-specific understanding and gives it the authority to act within policy. Models may be interchangeable, but accumulated operational knowledge compounds with every decision.

The enduring value lies not in the smartest brain, but in the system that understands the business well enough to run it.

The next operating system

The capital leaving enterprise software is not disappearing; it is migrating. It is shifting away from tools that digitize human effort and toward operating systems that execute routine work autonomously while keeping humans firmly in control.

Seats were the unit of pricing. Work becomes the unit of value.
 The companies that win will not merely digitize operations; they will run them.

Software is still eating the world - no longer one seat at a time.