The huge "Intelligent Agent Blue Ocean Market": software programming accounts for half, while healthcare, finance, legal, and other sectors are "scarce"

A recent study on the practical applications of AI agents reveals a highly imbalanced market landscape: software engineering dominates nearly half of the market, while medical, legal, financial, and over a dozen other vertical sectors combined account for only the other half, with each individual sector holding less than 5%. This situation points entrepreneurs toward a clear direction—true opportunities lie not in already saturated fields, but in the nearly untouched blue ocean markets.

A comprehensive study by Anthropic shows that 49.7% of AI agent tool calls on their API are in software engineering. In contrast, healthcare accounts for only 1%, legal 0.9%, education 1.8%. These are not saturated markets but areas where applications are still virtually nonexistent.

The research also uncovers a key insight: the actual capabilities of AI models far surpass user trust in their performance. METR’s capability assessments show that Claude can solve tasks that would take humans nearly five hours, yet in real-world use, the 99.9th percentile session length is only about 42 minutes. This huge gap between capability and deployment represents a product opportunity for entrepreneurs.

Garry Tan, President of Y Combinator, and Aaron Levie, CEO of Box, both believe this pattern indicates that 300 vertical AI unicorns will emerge in the future—compared to over 170 unicorns created during the SaaS era. The scale of AI versions could expand tenfold because they not only replace software but also substitute for operational personnel.

Software Engineering Dominates, Vertical Fields Are Nearly Empty

Anthropic’s data shows that software engineering accounts for half of all AI agent activity, with the other half spread across 16 vertical sectors, none exceeding 9%. Markets like healthcare, legal, education, customer service, and logistics each hold single-digit market shares.

This distribution isn’t because these verticals don’t need AI agents, but because relevant applications have not yet been truly developed. Software engineering’s dominance is due to developers being natural early adopters of AI tools and the relatively low technical barriers.

In contrast, vertical fields like healthcare and legal involve proprietary data, regulatory constraints, and complex organizational workflows. These seemingly obstacles actually form defensible competitive barriers. Anyone can build a general-purpose wrapper, but few can deeply understand the specific workflows of medical billing, legal discovery, or building permits.

Capabilities Outpace Trust: The Deployment Gap

The phenomenon of “deployment lag” revealed by the study warrants attention from entrepreneurs. The models already possess capabilities far beyond what users are willing to let them perform.

From October 2025 to January 2026, the 99.9th percentile session length nearly doubled, from less than 25 minutes to over 45 minutes. This growth remained steady across multiple model versions. It reflects not only improvements in model ability but also the accumulation of user trust—users are learning to collaborate with agents through successive sessions.

Anthropic researchers Miles McCain and others note that from August to December, Claude Code’s success rate on the most challenging internal tasks doubled, while the average number of human interventions per session dropped from 5.4 to 3.3. This indicates that as users better understand the model’s capabilities, they are willing to grant it more autonomy.

Capabilities are in place; deployment has yet to catch up. This is not a problem but a product opportunity.

The Paradox of Trust Evolution

The study uncovers a paradox in how user trust evolves: experienced users tend to approve more sessions automatically while also intervening more.

New users automatically approve about 20% of Claude Code sessions. After 750 sessions, this rises to over 40%. But at the same time, new users intervene in only 5% of turns, while seasoned users intervene in 9%.

This is not contradictory. The research team explains it as a shift in supervision strategy. Beginners approve each step before execution, whereas experienced users delegate more and intervene afterward, moving from pre-approval to active monitoring.

The study also highlights an important safety feature: in complex tasks, Claude Code requests clarifications more than twice as often as it receives human interventions. The agent pauses to confirm when uncertain rather than blindly proceeding. Researchers believe that “the autonomy exercised by the model in practice is jointly constructed by the model, the user, and the product. Claude limits its independence by pausing to ask questions when unsure.”

73% of tool calls involve human participation, and only 0.8% are irreversible. High-risk deployment scenarios, such as API key extraction or autonomous cryptocurrency trading, are mainly for security assessments rather than actual production use.

Defensible Strategies for Vertical AI

Aaron Levie’s vertical AI strategy reveals pathways to building defensible companies: develop agent software capable of accessing proprietary data; ensure the software truly solves real problems; fully leverage context to maximize intelligent output; and a key step most founders overlook—drive change management for clients.

This last point is precisely why vertical AI can be defensible. In vertical sectors, mastering traditional workflows, regulatory constraints, and organizational friction is crucial to differentiating a defensible company from a general wrapper.

The SaaS industry has grown tenfold every decade over the past few decades. Over 40% of venture capital in the past 20 years has flowed into SaaS companies, producing more than 170 unicorns. The logic of vertical AI is similar: each SaaS unicorn has a corresponding vertical AI version waiting to be built, and the AI version could be ten times larger because it not only replaces software but also substitutes operational personnel.

Researchers warn that policies requiring “approval for every operation” will stifle productivity gains without enhancing safety. A better approach is to ensure humans can monitor and intervene, rather than enforce strict approval workflows.

The Hidden Potential of 300 Unicorns

The market landscape is now clear. Software engineering has its domain, and the 16 vertical sectors—medical, legal, financial, education, customer service, logistics—each hold single-digit market shares, waiting for domain expertise to be integrated into intelligent agents.

Models can already work for five hours, but users only allow about 42 minutes of actual work. This gap indicates the market is still in its very early stages, with much work to be done and many fields yet to see even a minute of intelligent application.

Over the past two decades, 300 SaaS unicorns have been born; the next step is the emergence of 300 vertical AI unicorns. Founders who choose a vertical, embed domain expertise into their agents, and solve change management challenges will lead the next decade of enterprise software.

Risk Disclaimer and Legal Notice

Market risks are present; invest cautiously. This article does not constitute personal investment advice and does not consider individual users’ specific investment goals, financial situations, or needs. Users should evaluate whether any opinions, viewpoints, or conclusions herein are suitable for their particular circumstances. Investment is at your own risk.