The vast "blue ocean market for agents": software programming accounts for half, while fields like healthcare, finance, and law are "few and far between."

```

A recent study on the actual application of AI agents reveals an extremely imbalanced market landscape: software engineering dominates nearly half the market, while over a dozen verticals such as healthcare, law, finance together account for the other half, with each individual field representing less than 5%. This reality points the way for entrepreneurs—the real opportunity lies not in the fields that have been developed, but in those untouched blue ocean markets.

A comprehensive study released by Anthropic shows that software engineering accounts for as much as 49.7% in AI agent tool calls through its API. In contrast, healthcare takes only 1%, law 0.9%, education 1.8%. These are not saturated markets, but rather markets that barely exist.

The study also sheds light on a key finding: the actual capabilities of AI models far exceed the level of trust users have in them. METR’s capability assessments show that Claude can solve tasks requiring nearly five hours for a human, but in actual use, 99.9th percentile session duration is only about 42 minutes. This huge gap between capability and deployment is exactly the product opportunity entrepreneurs can seize.

Y Combinator President Garry Tan and Box CEO Aaron Levie both believe this landscape indicates that 300 vertical AI unicorns will emerge in the future, compared to the 170+ unicorns from the previous SaaS era. The scale of the AI version may grow 10-fold, as they not only replace software, but also operators.

Software engineering dominates, vertical fields nearly blank

Anthropic's data shows that software engineering occupies half of all AI agent activity, and the other half is scattered across 16 vertical fields, none exceeding 9%. Healthcare, law, education, customer service, logistics and other fields each hold single-digit market share.

This distribution is not 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 naturally being early adopters of AI tools and lower technical barriers.

In contrast, verticals like healthcare and law involve proprietary data, regulatory constraints, and complex organizational processes. These seemingly barriers in fact constitute defensible competitive moats. Anyone can build a generic wrapper, but few can deeply understand workflows like medical billing, legal discovery, or construction permits.

Capabilities ahead of trust: The deployment gap

The “lagging deployment” phenomenon revealed in the study is worth entrepreneurs’ attention. The capacity models already have far surpasses what users are willing to let them use.

From October 2025 to January 2026, the 99.9th percentile session duration nearly doubled, rising from less than 25 minutes to over 45 minutes. This increase remains steady across multiple model versions. This is not only a sign of model capability improvement but also of accumulated user trust—users learn to collaborate with agents through repeated sessions.

Anthropic researchers such as Miles McCain point out that from August to December, the success rate of Claude Code on the most challenging internal user tasks doubled, while average human intervention per session dropped from 5.4 to 3.3 times. This suggests that as users better understand agent capabilities, they are willing to grant more autonomy.

The capabilities are already present but deployment hasn’t caught up. This is not a problem, but a product opportunity.

The paradoxical pattern of trust evolution

The study found a phenomenon in the evolution of user trust: experienced users both automatically approve more sessions and intervene more as well.

New users automatically approve about 20% of Claude Code sessions. After 750 sessions, this rate rises to over 40%. However, while new users intervene in only 5% of rounds, experienced users’ intervention rate reaches 9%.

This is not contradictory. The research team explains it as a shift in supervision strategies. Novices approve before every step, while experienced users delegate and then intervene afterwards, moving from pre-approval to proactive monitoring.

The study also found a key safety feature: for complex tasks, Claude Code proactively requests clarification more than twice as often as human intervention. Agents pause to confirm when uncertain, instead of pushing ahead blindly. Researchers believe, "The autonomy exercised by agents in practice is jointly built by the model, user, and product. Claude limits its independence by pausing to ask when uncertain."

73% of tool calls involve human participation, and only 0.8% of operations are irreversible. The highest-risk deployment scenarios, such as API key extraction or autonomous cryptocurrency trading, are mainly for security evaluations rather than real production environments.

Defensible strategies for vertical AI

Aaron Levie's vertical AI strategy reveals the path to building defensible enterprises: create agent software able to access proprietary data; ensure the software truly solves real problems; fully utilize context to maximize intelligent output; and the key step most founders overlook—driving change management for the customer.

This last point is precisely why vertical AI is defensible. In verticals, mastering traditional workflows, regulatory constraints, and organizational friction is the key to differentiating defensible companies from generic wrappers.

The SaaS industry grew ten-fold every decade over the past several decades. Over the past 20 years, more than 40% of venture capital went to SaaS companies, producing over 170 unicorns. Vertical AI is similar: each SaaS unicorn has a corresponding vertical AI version waiting, and the AI version could be 10 times larger, as it replaces not just software but operators as well.

Researchers note that policies requiring “approval of every action” will kill productivity gains without increasing safety. A better goal is to ensure humans can monitor and intervene, not to mandate specific approval workflows.

The hiding places of 300 unicorns

The market map is already clear. Software engineering is spoken for, while verticals like healthcare, law, finance, education, customer service, logistics, etc. each hold single-digit market share, waiting for someone to embed domain expertise into agents.

Models are already capable of working for five hours, but users only let them work for 42 minutes. This gap shows the market is still in its early stages, with plenty left to do, and many fields have not even seen a minute of intelligent application.

Previously 300 SaaS unicorns emerged; next will be 300 vertical AI unicorns. Founders who choose a vertical, embed domain expertise into agents, and solve the change management challenge will lead the next decade of enterprise software.

Risk Warning and DisclaimerThe market involves risks; please invest cautiously. This article does not constitute personal investment advice, nor does it take into account the unique investment goals, financial statuses, or needs of individual users. Users should decide whether opinions, views, or conclusions in this article fit their specific circumstances. If investing based on this, the responsibility is theirs alone. ```