Healthcare AI: Regulated, Siloed Data Meets Search

TL;DR: The state of AI in healthcare shows that regulated, siloed data is a leading proving ground for specialized AI search, not a barrier to it. Healthcare organizations now adopt AI 2.2x faster than the broader economy, and 22% deployed domain-specific tools in 2025, because dense unstructured records reward systems that retrieve, attribute, and reason over institutional knowledge. The lesson for any regulated industry: connect the silos and put search over them, instead of buying one more disconnected app.

For years, healthcare was the example people reached for to describe a slow-moving industry. That story no longer holds. The $4.9 trillion sector now deploys AI at more than twice the rate of the broader U.S. economy (Menlo Ventures, 2025). What changed is not the regulation or the data sprawl. What changed is that AI search finally got good enough to work on dense, messy, regulated knowledge, and healthcare had the most of it to gain.

This post looks at why the state of AI in healthcare makes regulated, siloed data a proving ground for specialized search, what the adoption numbers actually say, and what other regulated industries can copy.

Why does regulated, siloed data favor specialized AI search?

A data silo is information trapped in one tool, team, or system that others cannot easily find or reuse. Healthcare runs on silos by design: electronic health records, imaging systems, claims platforms, lab results, prior-authorization queues, and clinical notes rarely share context. Most of that content is unstructured institutional knowledge — free-text notes, scanned forms, PDFs, and dictated summaries that a keyword search handles poorly.

General chatbots struggle here. They do not know which record is authoritative, who is allowed to see it, or where an answer came from. Specialized AI search is retrieval built for a domain: it indexes the institution’s own sources, returns answers with citations back to the original document, and respects access controls. In a regulated setting, that traceability is the point. An answer a clinician or auditor cannot trace back to a source is not usable.

That is why dense, regulated data turns out to be an advantage rather than a blocker. The harder the retrieval problem, the more a purpose-built system beats a generic one. Menlo Ventures found that fewer than one in ten companies across the broader economy (9%) has implemented AI, and most rely on general tools like enterprise ChatGPT instead of domain-specific solutions (Menlo Ventures, 2025). Healthcare went the other way and bought for the problem.

What the adoption numbers actually say

The shift happened fast. Healthcare went from 3% AI adoption to a leadership position in roughly two years (Menlo Ventures, 2025). The current picture:

• 22% of healthcare organizations deployed domain-specific AI tools in 2025, a 7x increase over 2024 and 10x over 2023 (eMarketer, 2025).
• Adoption splits by segment: health systems lead at 27%, ahead of outpatient providers (18%) and payers (14%) (HIT Consultant, 2025).
• Healthcare AI spending reached $1.4 billion in 2025, nearly tripling the prior year, with providers accounting for about 75% of the total (eMarketer, 2025).

The methodology matters for trusting those figures. Menlo Ventures partnered with Morning Consult to survey more than 700 executives involved in AI decisions across provider, payer, and pharma organizations in August and September 2025 (Menlo Ventures, 2025). This is a sized survey of buyers, not a vendor estimate.

One more detail separates healthcare from the average AI story: most of this money is in production, not pilots (HIT Consultant, 2025). The pilot phase is largely over for the organizations that committed.

Why the biggest use cases are retrieval problems

Look at where the budget went and a pattern appears. The two largest categories are ambient clinical documentation and coding and billing automation (Menlo Ventures, 2025). Both are, at their core, problems of reading unstructured text and pulling the right facts out of it: turn a messy visit into a structured note, or read a chart and assign the correct billing codes.

These categories won first because the return is obvious and the work is high-volume. They also share a dependency. Accuracy depends on retrieving the right context — the patient’s history, the relevant policy, last year’s exception — from systems that do not normally talk to each other. Get the retrieval wrong and the output is confidently wrong, which in a clinical or financial setting is worse than no output.

The opportunity ahead is larger than what has been captured. Total U.S. healthcare administrative spending runs around $740 billion a year, yet IT represents under 10% of it (Digital Health Wire, 2025). Menlo Ventures estimates roughly 80% of that market is still untapped (Digital Health Wire, 2025). The next wave is people-intensive workflows — prior authorization, patient engagement, front-office revenue cycle — that only become automatable once a system can search across the silos that feed them.

A concrete example: Vantage Health

Picture a regional health system, Vantage Health, with about a dozen clinics and a 400-bed hospital. A revenue-cycle analyst gets a denied claim for a complex procedure. To appeal it, she needs the original authorization, the clinical note that justified the procedure, the payer’s policy version in effect that month, and any similar appeal the team won last year. Today those four things live in four systems, and finding them takes most of an afternoon of asking colleagues.

This is the retrieval problem in miniature. The knowledge exists; it is just scattered and unstructured. A connective layer that indexes those sources and links the entities — patient, claim, policy, prior decision — turns the afternoon into a single query with citations to each source document.

That connective layer is the role SemanticOS is built for: a knowledge-graph and AI-search “operational brain” that ties fragmented tools together so both people and AI agents can find and reason over institutional knowledge, with attribution and permissions carried through. In a regulated setting, the citations and access controls are not extras. They are what make the answer safe to act on. The healthcare adoption curve is early evidence that this approach works where the data is hardest.

What other regulated industries should take from this

Healthcare is the loud example, but the mechanism is general. Any industry with dense, regulated, siloed records — financial services, insurance, legal, pharma manufacturing, government — has the same shape of problem and the same opening. The pattern that is working in healthcare is portable:

• Buy for the specific retrieval problem, not a generic assistant.
• Insist on source attribution and access controls from day one.
• Connect the silos first; search and agents are only as good as what they can reach.
• Start where the work is high-volume and the return is measurable.

The state of AI in healthcare is not a story about one vertical getting lucky. It is a demonstration that regulated, siloed data is exactly where specialized AI search earns its keep.

Key takeaways

• Regulated, siloed data is a proving ground for specialized AI search, not an obstacle; healthcare adopts AI 2.2x faster than the broader economy (Menlo Ventures, 2025).
• 22% of healthcare organizations deployed domain-specific AI in 2025, a 7x jump over 2024, with spending near $1.4 billion (eMarketer, 2025).
• The leading use cases — ambient documentation and coding automation — are retrieval problems over unstructured institutional knowledge.
• Roughly 80% of the administrative market is still untapped, and reaching it requires search across siloed systems (Digital Health Wire, 2025).
• The winning pattern is portable to any regulated industry: connect the silos, then put attributed, permissioned AI search over them.