Software Is Eating Labor: Why the Next Wave of Companies Will Be 10x Larger_

From "Eating the World" to "Eating Labor"

In August 2011, Marc Andreessen published his now-iconic essay in the Wall Street Journal: "Why Software Is Eating the World." The thesis was straightforward. Every industry — from retail to transportation to agriculture — would eventually be run by software. Borders would fall to Amazon. Blockbuster would fall to Netflix. The taxi industry would fall to Uber. Software would replace analog processes with digital ones, and the companies that understood this earliest would win.

Fifteen years later, in their Big Ideas 2026 collection, Andreessen Horowitz (a16z) has updated the thesis. The new formulation is more precise and far more consequential: software is eating labor.

The distinction matters enormously. "Eating the world" described digitization — moving processes from paper to screens, from phone calls to APIs, from spreadsheets to dashboards. The output was still produced by humans. The software just made humans faster.

"Eating labor" describes something qualitatively different. AI systems are now capable of producing outputs — written analysis, code, design mockups, legal research, customer service interactions, financial models — that previously required human labor. The software isn't making a human faster. It's replacing the human's output entirely.

This is not automation in the traditional sense. Robotic process automation (RPA) could handle rule-based, repetitive tasks. What large language models and agentic AI introduce is the ability to handle judgment-based, variable tasks — the kind of work that has historically been the exclusive domain of college-educated knowledge workers.

The economic implications are staggering. Global spending on enterprise software is roughly $300 billion annually. Global spending on the labor that enterprise software supports — knowledge workers in finance, legal, marketing, operations, engineering, customer success — exceeds $3 trillion. When software can address the labor budget and not just the tooling budget, the ceiling for category-defining companies grows by an order of magnitude.

Three Phases of AI Evolution

To understand where we are in this shift, it helps to map the three distinct phases of AI capability that have unfolded — and are still unfolding — since the release of GPT-3 in 2020.

Phase 1: Information Retrieval (2020-2024)

The first wave of commercially viable AI was fundamentally about information retrieval. ChatGPT, Perplexity, Copilot in its early form — these systems could find, synthesize, and present information faster than a human searching through documents, codebases, or knowledge bases.

This was valuable. It saved time. But it didn't replace work. A lawyer still needed to read the AI's research, verify it, and write the brief. A developer still needed to review the AI's code suggestions, test them, and integrate them. The human remained the bottleneck and the decision-maker.

In this phase, the business model was additive. Companies charged for AI as a feature on top of existing workflows — a copilot, an assistant, a search upgrade. The addressable market was bounded by the existing software market, plus a modest premium for AI capability.

Phase 2: Reasoning (2025)

The release of reasoning-capable models — OpenAI's o1 and o3, Anthropic's Claude with extended thinking, DeepSeek R1 — marked a genuine capability discontinuity. These models don't just retrieve and recombine information. They can hold multi-step chains of logic, evaluate trade-offs, identify edge cases, and produce outputs that reflect genuine analytical reasoning.

This changed the nature of what AI could replace. A reasoning model can draft a complete legal brief, not just find relevant precedents. It can architect a software system, not just suggest code snippets. It can produce a financial model with assumptions, sensitivity analysis, and narrative commentary.

The practical impact in 2025 was that AI began handling complete tasks, not just subtasks. The human shifted from "doing the work with AI assistance" to "reviewing the AI's work and making final decisions." For many categories of knowledge work, the labor content of a task dropped by 50-80%.

Phase 3: Multiplayer/Agentic Mode (2026)

The phase we are entering now — and the one that a16z's "software eating labor" thesis is built on — is the emergence of agentic AI systems. These are not single-model, single-prompt interactions. They are orchestrated systems where multiple AI agents collaborate, delegate subtasks, use tools, access external systems, and pursue multi-step goals with minimal human intervention.

In this phase, the unit of AI work shifts from "answer a question" or "complete a task" to "achieve an objective." An agentic system tasked with "prepare the Q1 board deck" doesn't just draft slides. It queries the data warehouse, generates financial summaries, pulls competitive intelligence, drafts narrative sections, creates visualizations, cross-references with the previous quarter's deck for consistency, and presents a complete deliverable for human review.

This is what "software eating labor" looks like in practice. The system doesn't assist a human performing labor. It performs the labor, with a human serving as a reviewer and decision-maker rather than a producer.

The Theory of Well: Where Value Accrues

One of the most compelling frameworks in a16z's 2026 analysis is what they call the Theory of Well. The argument draws an analogy to the oil industry: in any resource-extraction economy, the entity that controls the chokepoint — the well — captures a disproportionate share of the economics. Refineries, pipelines, and gas stations all participate in the value chain, but the well owner has structural leverage over all of them.

In the AI economy, the "wells" are infrastructure chokepoints. These include:

• Foundation model providers (OpenAI, Anthropic, Google DeepMind) who control the core reasoning capability
• Compute infrastructure (NVIDIA, cloud hyperscalers) who control the hardware layer
• Data infrastructure (Snowflake, Databricks, MongoDB) who control the structured access to enterprise data
• Orchestration layers (LangChain, CrewAI, emerging agentic frameworks) who control how agents are composed and deployed
• Governance and observability layers (emerging category) who control compliance, auditing, and cost management for agentic systems

The Theory of Well argues that applications built on top of these infrastructure layers are structurally disadvantaged in the same way that a gas station is structurally disadvantaged relative to an oil well. Applications can be replicated. The AI wrapper problem — where a thin application layer over an API gets displaced by the API provider adding the same feature — is a manifestation of this dynamic.

The durable value capture happens at the infrastructure layer. Not because infrastructure is inherently more valuable, but because infrastructure creates switching costs, network effects, and data moats that applications cannot easily replicate.

Agent-Speed Architecture: The Infrastructure Problem No One Is Talking About

There is a critical technical problem lurking beneath the agentic AI thesis that most market analysis ignores: existing enterprise backends were not built for agent-speed interaction patterns.

Today's enterprise software architectures are designed around a 1:1 human-to-system interaction model. A human clicks a button, the system processes a request, the system returns a response. The entire stack — APIs, databases, rate limiters, authentication systems, billing models — is calibrated for human-speed interactions. A busy human might generate 50-100 API calls per hour.

An agentic system pursuing a single goal can generate 5,000 recursive sub-tasks in milliseconds. Each sub-task may involve multiple API calls, database queries, and inter-system communications. A single agentic workflow for "reconcile this month's accounts payable" might generate more API calls in 30 seconds than a human accountant generates in a month.

This is not a theoretical problem. Companies deploying agentic AI in production are already hitting rate limits, triggering fraud detection systems, overwhelming database connection pools, and generating cloud computing bills that exceed the labor cost they were trying to replace.

The architecture problem has several dimensions:

Concurrency. Enterprise databases are designed for hundreds of concurrent human users, not thousands of concurrent agent sub-tasks. Connection pooling, locking strategies, and query optimization all assume human-speed interaction patterns.

Cost. LLM inference costs are declining rapidly, but agentic systems compound inference costs through recursive sub-task generation. A single agentic workflow might invoke the underlying LLM 200+ times. At current pricing, this means that an agentic system replacing a $75/hour analyst might cost $30-50 per workflow execution — viable, but only with careful architecture.

Observability. When a human uses software, the audit trail is straightforward: the human took an action, the system recorded it. When an agentic system takes 500 actions in pursuit of a single goal, the audit trail becomes a debugging problem. Which sub-task failed? Which decision point led to an incorrect output? Enterprise governance requires answers to these questions.

Authentication and authorization. Existing enterprise identity and access management (IAM) systems were designed for human users with roles and permissions. Agentic systems require a fundamentally different authorization model: what can this agent do, on whose behalf, with what constraints, and with what escalation paths when it encounters an action outside its authorized scope?

The companies that solve these infrastructure problems — agent-speed databases, agentic orchestration platforms, AI-native observability tools, agent-aware IAM systems — will capture a disproportionate share of the value created by the agentic AI wave. This is the Theory of Well in action.

Why Category Winners Will Be 10x Larger

The claim that AI-native category winners will be 10x larger than their SaaS predecessors is not hyperbole. It's arithmetic.

Consider the market for customer support software. The global market for customer support tools — help desks, ticketing systems, chatbots, knowledge bases — is roughly $15 billion annually. The global market for customer support labor — the salaries, benefits, training, and management overhead of customer support teams — exceeds $150 billion annually.

A SaaS company selling customer support software can address the $15 billion market. An AI-native company that replaces the customer support labor itself can address the $150 billion market. The category is 10x larger because the addressable market has expanded from "what companies spend on tools" to "what companies spend on the function."

This pattern repeats across every knowledge work category:

Category	Tool Market	Labor Market	Multiplier
Customer Support	$15B	$150B	10x
Legal Services	$10B	$350B	35x
Financial Analysis	$20B	$200B	10x
Software Engineering	$25B	$500B	20x
Marketing & Content	$15B	$300B	20x

Y Combinator's thesis on vertical AI agents maps directly to this logic. When YC president Garry Tan says vertical AI agents could be "10x bigger than SaaS," he is pointing at the same arithmetic. A vertical AI agent for legal work doesn't sell a $50/month tool to lawyers. It replaces $200/hour of paralegal and associate labor. The revenue per customer is higher, the addressable market is larger, and the value proposition is more defensible because the customer is buying an outcome, not a tool.

Sequoia's Arc product-market fit (PMF) framework provides a useful lens for evaluating ventures in this landscape. The framework identifies three PMF archetypes:

1. Hair on Fire — The customer has an urgent, well-understood problem and is actively seeking a solution. In the AI labor context, this maps to categories where labor costs are already a C-suite concern: customer support, data entry, basic legal review, routine financial reporting.
2. Hard Fact — The customer may not be actively seeking a solution, but an undeniable market shift makes the solution inevitable. The "hard fact" in 2026 is that AI can now perform knowledge work at 10-20% of the cost of human labor, with comparable or superior quality for routine tasks. Companies that ignore this fact will be structurally disadvantaged against competitors that embrace it.
3. Future Vision — The customer doesn't know they need the solution yet because the solution creates a category that didn't previously exist. Agentic AI systems that can autonomously manage entire business functions — not just assist with tasks — represent a Future Vision opportunity. The companies building toward this vision today will define the categories of tomorrow.

The Gartner Reality Check

Not all of this will proceed smoothly. Gartner's forecasts provide a necessary counterweight to the optimism.

By the end of 2026, Gartner projects that 40% of enterprise applications will integrate AI agents in some capacity. This is a remarkable rate of adoption for a technology category that barely existed 18 months ago.

But Gartner also projects that more than 40% of agentic AI projects initiated before 2028 will be scaled back or cancelled due to three primary factors:

Cost overruns. Agentic AI systems are compute-intensive. The recursive sub-task generation described above means that inference costs compound in ways that are difficult to predict during proof-of-concept phases. Many enterprises are discovering that their agentic AI pilots work beautifully in controlled environments but become prohibitively expensive at production scale.

Governance failures. Regulated industries — financial services, healthcare, legal, government — require auditability, explainability, and compliance documentation for consequential decisions. Current agentic AI systems provide limited transparency into their multi-step reasoning processes. When an agent makes 500 micro-decisions to produce an output, explaining why it reached a particular conclusion is a non-trivial technical challenge.

Integration complexity. Agentic AI systems need to interact with existing enterprise software — CRMs, ERPs, HRISs, financial systems. These integrations are brittle, poorly documented, and designed for human-speed interaction patterns. Building reliable, production-grade integrations between agentic systems and legacy enterprise software is significantly harder than most vendors acknowledge.

These failure modes are real, and they will create significant market disruption. But they also represent opportunities. The companies that solve cost optimization for agentic workloads, governance for multi-agent systems, and reliable enterprise integration patterns will be building the infrastructure layer — the "wells" — of the agentic economy.

What This Means for Venture Studios

The implications for how ventures should be built — and where venture studios should focus — are clear.

The application layer in AI is structurally vulnerable. AI wrappers get displaced. Features get absorbed by platform providers. Thin application layers over foundation model APIs face the same margin compression that mobile apps faced when Apple and Google continuously expanded the capabilities of their operating systems.

The infrastructure layer is where durable value accrues. Studios that build companies at the infrastructure level — the chokepoints where switching costs are high, data moats compound, and network effects create defensibility — are positioned for the 10x wave.

At AI Gens, our portfolio reflects this thesis:

Moonxi builds AI delivery infrastructure — the orchestration and operational layer that enables companies to deploy, monitor, and govern AI-powered delivery systems at scale. This is infrastructure, not application. It sits at a chokepoint between AI capabilities and real-world delivery operations, accumulating operational data that creates a compounding moat.

HOST360 builds hospitality operating infrastructure — the backend systems that enable hospitality businesses to manage operations, compliance, and guest experience at scale. In a world where agentic AI will manage guest interactions, dynamic pricing, and operational logistics, HOST360 provides the infrastructure layer those agents need to operate.

Both companies address the infrastructure layer. Both create switching costs through data accumulation and operational integration. Both are positioned to capture value as the application layer above them becomes increasingly AI-driven.

The Category Reimagination Opportunity

The most important takeaway from the "software eating labor" thesis is not about adding AI to existing categories. It is about reimagining categories entirely around AI-native economics.

Adding an AI chatbot to a traditional help desk is not "software eating labor." It's adding a feature. Building an AI-native customer success platform that autonomously handles 90% of customer interactions, escalates the remaining 10% to humans with full context, and continuously improves from every interaction — that is software eating labor. And the company that builds it is not competing in the $15 billion help desk market. It is competing in the $150 billion customer support labor market.

The same logic applies across every vertical. The opportunity is not to build a better tool for accountants. It is to build a system that performs accounting. Not a better tool for legal researchers. A system that performs legal research. Not a better tool for software engineers. A system that engineers software.

These are not incremental improvements. They are category redefinitions. And the companies that define these new categories — built on infrastructure that can handle agent-speed workloads, with governance that satisfies enterprise requirements, and with data moats that compound over time — will be the 10x companies that this thesis predicts.

The race is not to add AI to the old world. It is to build the new one.