Shadow IT with AI in health: 7 in 10 organizations lose control of patient data

Seven out of ten health organizations have already detected tools generated by AI processing data from patients out of control of the IT department. The data, released by Help Net Security, describes a governance problem that begins before any policy: no one can regulate what does not yet know exists.

Shadow IT with AI in health: what the survey found

The proportion of 70% refers to health sector organizations that at some point identified AI tools dealing with patient information without going through IT. It's worth reading the number for what it is: a floor. These are confirmed detections, not estimates of total use. That is, the actual volume of tools running in the margin tends to be greater than that pointed out by research — each tool discovered suggests others that remain invisible.

The phenomenon is a variation of the old shadow IT, now powered by generative models that anyone can trigger without installing anything, without calling IT and without contract with supplier.

Code Assistants: the most difficult vector to close in Shadow IT with AI

Code review, the main quality control of software development, covers only what is formally sent to the repository. She does not see the path taken before that — and it is precisely in this passage that the risk focuses.

A developer who needs to debug a query usually pastes data samples for the wizard to understand the format. The code that reaches the pull request looks clean; the content shared with the external tool does not. As the record of what has been glued rarely exists, the subsequent audit runs out of material to investigate.

Why Inventory Governance Fails in Shadow IT with AI in Health

The industry’s classic controls — catalogue of homologated applications, supplier evaluation, domain blocking, data classification — all start from a silent premise: the organization knows what’s running. Shadow IT reverses that logic. The most critical item is precisely what does not appear on the list, and detection, when it happens, is usually reactive — after an incident, an audit or an internal complaint.

The result is an uncomfortable cycle for the health sector: the discovery of a tool does not generate control over the others, only about that one.

The impact of Shadow IT with AI on health in Brazil

LGPD classifies health data as sensitive and requires specific legal basis and restricted treatment, which elevates the rule of any tool that touches this type of information. Brazilian hospitals, operators and laboratories usually operate with lean IT teams and increasing pressure for productivity — a combination that favors informal adoption of assistants.

The news does not bring a Brazilian cut, so there is no way to say that the proportion of 70% is replicated here. What can be said is that the governance model based on formal inventory has the same blind spot in any country. Those who already work with the discipline of agent engineering, as in the AI-DLC method taken to AWS Marketplace by SoftwareOne, part of an advantage: the record of what the AI access becomes part of the process, not a byproduct of it.

What Shadow IT research with AI in health does not answer

The survey does not detail which tools were found, which suppliers, how many patient records were processed or whether there was a confirmed leak. Nor does it explain how organizations did the detection — whether by network monitoring, internal auditing or accidental discovery. Without these data, it is not possible to measure gravity case by case, only the size of the phenomenon.

The central message, however, does not require details: while AI governance in health depends on knowing what has already been found, it will continue to chase.

Frequently asked questions about Shadow IT with AI in health

What is shadow IT in health?

It is the use of technology tools and services — including generative AI assistants — by employees out of the IT department’s knowledge and control. In the health sector, the risk is higher because these tools often process patient data.

The code review doesn't solve the problem?

It only partially resolves. The review covers what is formally sent to the repository, but not what was shared with AI assistants during development, before uploading.

How to reduce this risk in practice?

The path pointed out by the logic of the problem itself is to make visible what is invisible today: to record what the AI tools access, rather than rely only on formal inventories that only see what has already been approved.