A recent study on the practical application of AI agents reveals a highly imbalanced market structure: software engineering alone accounts for nearly half of the activity, while over a dozen vertical sectors, including healthcare, law, and finance, combined make up the other half, with no single sector exceeding 5% share. This situation points entrepreneurs towards the real opportunity—not in already cultivated fields, but in the virtually untouched blue ocean markets.

Comprehensive research from Anthropic indicates that software engineering constitutes a dominant 49.7% of AI agent tool calls via its API. In contrast, healthcare accounts for only 1%, legal 0.9%, and education 1.8%. These are not saturated markets; they are markets that barely exist yet.

The study also uncovered a key finding: the actual capabilities of AI models far exceed the level of trust users currently place in them. Evaluations by METR show that Claude can solve tasks that would take a human nearly five hours to complete. However, in practical use, the 99.9th percentile session length is only about 42 minutes. This significant gap between capability and deployment represents a tangible product opportunity for entrepreneurs.

Both Y Combinator President Garry Tan and Box CEO Aaron Levie believe this landscape signals the future emergence of 300 vertical AI unicorn companies, compared to the 170+ unicorns created during the SaaS era. The scale of the AI versions could be ten times larger because they replace not just software but also operational personnel.

Software engineering dominates, while vertical sectors are nearly blank. Data from Anthropic shows software engineering accounts for half of all AI agent activity, with the other half distributed across 16 vertical sectors, none exceeding 9%. Market shares in sectors like healthcare, law, education, customer service, and logistics are all in the single digits.

This distribution is not because these verticals don't need AI agents, but because relevant applications haven't been truly developed yet. Software engineering's dominance stems from developers being natural early adopters of AI tools and the relatively lower technical barriers.

In contrast, verticals like healthcare and law involve proprietary data, regulatory constraints, and complex organizational processes. These factors, which appear as obstacles, actually form defensible competitive moats. Anyone can build a general-purpose wrapper, but few can deeply understand the specific workflows of medical billing, legal discovery, or building permits.

The deployment gap: Capability outpaces trust. The "deployment lag" phenomenon revealed by the study warrants attention from entrepreneurs. The capabilities models already possess far surpass the level to which users are currently willing to let them operate.

From October 2025 to January 2026, the 99.9th percentile session length nearly doubled, growing from under 25 minutes to over 45 minutes. This growth remained steady across multiple model versions. This indicates not just improved model capability but also accumulating user trust—users learn to collaborate with agents over successive sessions.

Researchers from Anthropic, including Miles McCain, noted that from August to December, Claude Code's success rate on internal users' most challenging tasks doubled, while the average number of human interventions per session dropped from 5.4 to 3.3. This suggests that as users gain understanding of the agent's abilities, they grant it more autonomy.

The capability is already there; deployment hasn't caught up. This isn't a problem but a product opportunity.

A paradoxical pattern in the evolution of trust. The study observed a phenomenon in user trust evolution: experienced users both auto-approve more sessions and also intervene more frequently.

New users auto-approved roughly 20% of Claude Code sessions. After 750 sessions, this rate increased to over 40%. Simultaneously, new users intervened in only about 5% of turns, whereas veteran users had an intervention rate of 9%.

This is not a contradiction. The research team explains it as a shift in supervision strategy. Novices tend to approve before each step, while experienced users adopt a delegate-and-intervene-later model, moving from pre-approval to proactive monitoring.

The study also identified an important safety feature: in complex tasks, Claude Code proactively requested clarification more than twice as often as human interventions occurred. The agent pauses to confirm when uncertain, rather than blindly pushing forward. The researchers believe "the autonomy exercised by agents in practice is co-constructed by the model, the user, and the product. Claude constrains its own independence by pausing to ask questions when uncertain."

73% of tool calls involved human participation, and only 0.8% of operations were irreversible. The highest-risk deployment scenarios, like API key extraction or autonomous cryptocurrency trading, were primarily for security assessment, not actual production use.

Defensible strategies for vertical AI. The vertical AI strategy proposed by Aaron Levie outlines a path to building defensible businesses: build agent software that can integrate proprietary data; ensure the software genuinely solves real problems; leverage context fully to maximize output intelligence; and a key step most founders overlook—driving change management for the customer.

This last point is precisely what makes vertical AI defensible. In vertical sectors, navigating legacy workflows, regulatory constraints, and organizational friction is key to differentiating a defensible company from a generic wrapper.

The SaaS industry grew 10x every decade over recent decades. Over the past 20 years, over 40% of venture capital flowed to SaaS companies, creating more than 170 unicorns. The logic for vertical AI is similar: for every SaaS unicorn, a corresponding vertical AI version is waiting, and the AI version could be 10x larger because it replaces both software and operators.

The researchers note that policies requiring "approval for every action" would kill productivity gains without enhancing safety. A better goal is ensuring humans can monitor and intervene, rather than mandating specific approval workflows.

Where the 300 unicorns are hiding. The market map is clear. Software engineering is already claimed. Sixteen vertical sectors—including healthcare, law, finance, education, customer service, and logistics—each hold single-digit market shares, waiting for someone to embed domain expertise into agents.

Models are capable of working for five hours, but users only let them work for 42 minutes. This gap indicates the market is still in its very early stages, with immense work remaining, and many sectors haven't even seen a minute of agent application.

Just as 300 SaaS unicorns were born previously, the next wave will see 300 vertical AI unicorns. Founders who choose a vertical sector, build domain expertise into their agents, and solve the change management problem will dominate the next decade of enterprise software.