The most important part of the VA watchdog report is not that clinicians are using AI.
Everyone in healthcare should assume that is already happening.
The real problem is that generative AI tools are moving into clinical work faster than governance programs can classify, monitor, and control them. That is the lesson from the VA Office of Inspector General’s review of generative AI chat tools used by Veterans Health Administration staff.
According to the watchdog, VA clinicians and staff were using general-purpose tools such as VA GPT and Microsoft 365 Copilot Chat for work that may involve patient information, clinical documentation, and patient care support. The concern was not simply that these tools existed. The concern was that the agency had not treated those tools with the same level of risk oversight it applied to more obviously clinical AI systems.
That distinction matters.
Healthcare organizations are used to thinking about AI risk in terms of formal products: an ambient scribe, a diagnostic aid, a revenue cycle tool, a population health model, or a clinical decision-support system. Those tools usually pass through procurement, security review, legal review, vendor diligence, and some form of implementation planning.
But general-purpose AI chat tools are different. They spread sideways.
They appear inside the productivity stack. They show up in Microsoft 365, collaboration tools, browsers, note-taking workflows, internal knowledge bases, and employee-facing platforms. They are not always introduced as “clinical AI.” They are framed as writing assistants, summarization tools, research helpers, drafting tools, or productivity accelerators.
Then clinicians start using them.
That is where the risk begins.
The “It’s Just a Search Engine” Defense Is Not Good Enough
One of the most revealing points from the VA review is that leaders reportedly minimized the risk of general-purpose AI chat tools by comparing them to search engines and emphasizing user-level responsibility.
That is a common but flawed way to think about AI in healthcare.
A search engine retrieves information. A generative AI system creates output. It can summarize, infer, reformat, prioritize, draft, and present information in a confident clinical tone. A clinician may still be responsible for the final decision, but the system can shape what the clinician sees, how the clinician understands it, and what ultimately gets copied into a patient record.
That is not the same risk profile as a search bar.
When an AI tool helps draft a clinical note, summarize a patient history, produce discharge instructions, translate medical information, suggest follow-up language, or organize diagnostic considerations, it is no longer merely a convenience tool. It has entered the clinical workflow.
Once AI-generated language can influence documentation or care decisions, governance cannot stop at a warning label that says “use your judgment.”
Healthcare already knows this from HIPAA, security, privacy, and patient safety. Policies that rely entirely on individual discretion are not controls. They are risk transfers.
A compliant organization needs system-level safeguards, not just user-level reminders.
AI Governance Must Follow the Use Case, Not the Product Label
The VA report highlights a critical governance failure that applies far beyond the federal government: organizations often classify AI tools based on what the vendor says the tool is, rather than how the workforce actually uses it.
That is backward.
If an AI product is marketed as a general-purpose assistant but clinicians use it to summarize patient care, draft medical record content, or support clinical reasoning, the risk classification should follow the use case. The name of the tool matters less than the workflow it enters.
This is the same mistake organizations make with privacy and tracking technology. A pixel may be sold as analytics. A chat widget may be sold as customer support. A session replay tool may be sold as website optimization. But if it collects health data, financial data, children’s data, or sensitive consumer behavior, its legal and compliance risk changes.
AI is no different.
Healthcare organizations need to ask a simple question: What is the tool actually doing inside our environment?
Not what does the vendor call it. Not what did procurement approve it for. Not what did the internal launch email say. What are employees putting into it? What outputs are they relying on? Is protected health information being used? Is the output entering the medical record? Could the output influence diagnosis, treatment, triage, patient communication, billing, or care management?
If the answer is yes, the tool belongs in a governed AI risk program.
Prompt Governance Is Now a Patient Safety Issue
The VA watchdog identified clinical prompts shared across the organization and noted that prompt techniques can affect the accuracy of AI outputs in medical use cases. That point should get the attention of every hospital, health system, clinic, and digital health company.
Prompts are not harmless.
A prompt can tell the model what information to prioritize, what assumptions to make, what format to use, and what clinical tone to adopt. A poorly designed prompt can omit context, overstate certainty, compress nuance, or create a false sense of completeness. In healthcare, those errors can matter.
This means prompt libraries, prompt sharing, and prompt reuse need governance.
If clinicians are informally sharing prompts through Teams, Slack, email, intranet pages, or department documents, that is no longer just employee collaboration. It is the creation of an unofficial clinical workflow layer. The organization should know which prompts are being used, who approved them, whether they have been tested, whether they involve PHI, and whether the output is allowed to enter documentation.
The same is true for AI-generated templates.
A hospital would not allow an unreviewed clinical form, consent template, medication instruction sheet, or discharge workflow to spread across departments without oversight. AI prompts and outputs should not be exempt simply because they feel lightweight.
The Missing Control: Traceability
The most dangerous AI failure may not be a dramatic hallucination. It may be an untraceable one.
If AI-generated text enters a clinical note without any marker, audit trail, label, or reporting mechanism, the organization loses the ability to investigate patterns. It cannot easily determine whether an error came from a clinician’s independent judgment, a copied AI output, a bad prompt, an outdated model response, or a workflow design issue.
That creates patient safety exposure. It also creates compliance exposure.
Healthcare organizations need traceability around AI-assisted documentation. That does not mean every AI interaction needs to be treated as a crisis. It does mean that organizations need a way to understand when AI is being used in regulated workflows, what data is being entered, what output is being produced, and where that output is going.
Without traceability, the organization cannot answer basic questions after an incident:
Was AI used?
Was PHI entered into the tool?
Was the tool approved for that purpose?
Was the prompt reviewed?
Was the output validated by a clinician?
Was the output copied into the medical record?
Have similar errors occurred before?
Were staff trained to report the issue?
Those questions are not theoretical. They are the questions legal, compliance, privacy, security, risk management, and patient safety teams will be forced to answer after something goes wrong.
HIPAA Programs Need to Expand Their AI Lens
AI governance in healthcare is not only a patient safety issue. It is also a privacy and security issue.
If clinicians or staff enter patient information into an AI tool, the organization needs to understand whether the tool is approved to receive PHI, whether the vendor is acting as a business associate, whether a business associate agreement is required, whether data is retained, whether data is used for training, whether access is logged, and whether minimum necessary principles are being respected.
A healthcare organization cannot answer those questions with a one-page AI policy.
It needs an inventory.
It needs data-flow mapping.
It needs vendor review.
It needs workforce training.
It needs monitoring.
It needs incident reporting.
It needs rules for permitted and prohibited use.
Most importantly, it needs to connect AI governance with the existing privacy, security, and compliance program. AI should not sit in a separate innovation silo run only by IT or digital transformation. In healthcare, AI governance needs cross-functional ownership from privacy, security, compliance, legal, clinical leadership, procurement, quality, and patient safety.
That is the only way to manage the full risk surface.
The Board-Level Lesson: Shadow AI Is Already Here
The VA report should make healthcare executives uncomfortable because it shows how quickly AI adoption can outpace oversight even inside a large, sophisticated government health system.
Private healthcare organizations should not assume they are different.
If a hospital has enabled Microsoft Copilot, Google Gemini, ChatGPT Enterprise, internal GPT tools, ambient scribe pilots, AI call center tools, AI coding tools, AI search assistants, or AI documentation support, then the organization already has an AI governance problem to solve.
The question is not whether employees are using AI.
The question is whether leadership knows how they are using it.
That requires more than an annual policy update. It requires continuous governance. Healthcare organizations need to know which AI tools are active, which departments are using them, which workflows involve PHI, which uses are clinical, which are administrative, and which have never been approved.
The highest-risk category may be the tools that were not purchased as clinical systems but became clinical systems through everyday use.
What Healthcare Organizations Should Do Now
Every healthcare organization should treat this report as a practical checklist.
First, create an AI inventory that includes both formal AI systems and general-purpose AI tools embedded in productivity software. Do not limit the inventory to tools branded as clinical AI.
Second, classify AI use cases by risk. A chatbot used to rewrite an internal newsletter is not the same as a chatbot used to summarize patient history or draft medical record language.
Third, define permitted and prohibited uses. Staff should know whether PHI can be entered, whether outputs can be copied into medical records, whether AI can be used for patient communication, and when human review is mandatory.
Fourth, review vendor and data protection terms. Healthcare organizations need to know whether AI vendors can access PHI, whether data is retained, whether it is used to train models, where it is stored, and what contractual protections apply.
Fifth, govern prompts and templates. Any prompt or workflow intended for clinical use should be reviewed, tested, versioned, and monitored.
Sixth, create AI-specific reporting channels. Staff should know how to report hallucinations, unsafe outputs, privacy concerns, documentation errors, or inappropriate use.
Seventh, connect AI monitoring to patient safety, HIPAA, and compliance programs. AI risk should not live in a policy folder. It should be operationalized through training, audits, incident response, vendor management, and governance meetings.
AI Risk Is Becoming Expensive Compliance Risks
Healthcare organizations are under pressure to adopt AI because the productivity upside is real. Clinicians are burned out. Documentation burden is high. Administrative workflows are inefficient. Patients expect faster service. Leadership wants automation.
AI can help.
But healthcare cannot repeat the same pattern it followed with pixels, trackers, chat tools, call recording, and marketing automation: deploy first, govern later, then discover the compliance exposure after regulators, plaintiffs, or the press point it out.
The VA report is a reminder that AI governance must begin at the workflow level. The risk is not just the model. The risk is the combination of sensitive data, clinical context, user behavior, vendor terms, documentation practices, and weak oversight.
For healthcare organizations, the standard should be clear: if AI touches patient information, clinical documentation, patient communication, or care delivery, it needs governance before it becomes invisible infrastructure.
The organizations that get this right will not be the ones that ban AI. They will be the ones that make AI usable, auditable, accountable, and safe.
That is the future of healthcare compliance.
And it is arriving faster than most compliance programs are built to handle.
