More than 15,000 Department of Veterans Affairs staff have been using general-purpose AI chatbots to draft clinical notes and summarize patient care — and according to the VA Office of Inspector General, nobody is centrally checking what those bots are actually saying. That’s not a hypothetical risk. That’s a review window between October 2025 and January 2026, with clinical prompts already running in production against tools that were never designed for medicine.
The VA OIG report, released this month, is a case study in how fast shadow AI adoption outruns governance — and it should be required reading for anyone shipping healthcare software, building compliance tooling, or selling AI into regulated industries.
What the VA OIG Actually Found Inside the VHA
The OIG’s review of GenAI usage across the Veterans Health Administration identified more than 15,000 staff members using two authorized chat tools: VA GPT and Microsoft 365 Copilot Chat. Inside an internal prompt-sharing application, reviewers found 135 prompts in circulation — 79 of them clinical. Drafting clinical notes and summarizing patient care were among the most common applications, though neither tool was built for clinical work.
Why it matters: this is the gap between “authorized” and “governed.” The VA allowed staff to use these tools, gave them general training, and stopped there. According to the OIG, the agency does not centrally curate or evaluate the prompts being used, nor does it evaluate the generative output that may feed into clinical decisions. The OIG also cited research showing that prompt techniques themselves can drive output errors with downstream consequences for diagnosis and care management.
Practical example: if you’re a hospital CIO who has “approved” a general-purpose copilot for staff productivity, you may have already replicated the VA’s situation without realizing it. The moment a nurse pastes a patient summary into a prompt to “clean it up,” you’ve crossed from productivity tool into clinical workflow — and you probably don’t have the healthcare-specific guardrails to catch it.
The take: “authorized” is the new “shadow IT.” Approval without monitoring is just plausible deniability with a paper trail.
Why the High-Impact Classification Question Is the Whole Ballgame
The Office of Management and Budget’s 2025 memorandum, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust, requires federal agencies to identify high-impact AI use cases and apply safeguards to manage their risk. The VA did not classify VA GPT or Copilot Chat as high-impact, which meant the required risk management actions — pre-deployment testing, human oversight protocols, ongoing monitoring — did not apply.
Meanwhile, the VHA did classify Ambient AI Scribe as high-impact. That tool listens to clinical visits and drafts medical record notes. The OIG points out the uncomfortable parallel: staff were using VA GPT and Copilot Chat for documentation tasks with functionality similar to Ambient AI Scribe, but only the purpose-built tool got the safety regime.
Why it matters: classification drives obligation. If you can keep a tool labeled “general productivity,” you avoid pre-deployment testing, mandatory oversight, and integration with patient safety programs. That creates a perverse incentive across every regulated industry — keep your AI deployments vague, and you keep the compliance bill cheap.
Practical example: if your team is shipping an AI agent for internal knowledge or support into a healthcare customer, expect their procurement team to start asking exactly which OMB or HHS risk tier applies. The vendor that can answer that question with documentation will close the deal. The one that says “it’s just a chatbot” will lose it.
The take: the regulatory definition of “high-impact” will broaden across 2026, and any general-purpose chatbot touching clinical data will be reclassified upward — voluntarily or otherwise.
The Three Recommendations and the April 2027 Clock
The VA OIG made three recommendations to the VHA: evaluate VA GPT and Copilot Chat as high-impact tools, implement the required safeguards, and integrate monitoring of AI-related risks into existing patient safety programs. The VHA concurred in principle with the first and concurred with the other two. The action plan targets April 2027 for completion.
Why it matters: that’s a 16-month remediation window for a known patient safety risk involving 15,000+ users. The healthcare sector outside the VA has even less structured oversight. Earlier this year, Health-ISAC issued its white paper Policies and Safeguards for a Safe Use of AI, and the Health Sector Coordinating Council published the HSCC Health Industry AI Cyber Governance Framework Implementation Guide. Both documents exist precisely because private-sector hospitals are running into the same shadow-adoption pattern, often without an inspector general to flag it.
Practical example: imagine you’re a regional health system rolling out Microsoft 365 Copilot Chat for administrative staff. The VA’s experience tells you that within a quarter, clinicians will start using it for note drafting whether you sanctioned that or not. Your governance framework needs to assume clinical use from day one, not patch it in after an OIG-style audit.
The take: organizations that wait for their own internal version of this report will spend three times what they would have on proactive governance. The decision between custom and off-the-shelf AI is no longer purely about features — it’s about whether you can prove what the model did, to whom, and when.
What This Means for Healthcare Software Vendors
For anyone building healthcare AI, the VA OIG report is a buyer’s checklist in disguise. Expect RFPs to start asking for: prompt-level audit logs, output evaluation pipelines, evidence of pre-deployment testing against clinical scenarios, and integration hooks into existing patient safety reporting systems. Generic productivity AI will not pass these procurement filters in 2027.
Why it matters: the differentiator is no longer what the model can do — it’s whether you can prove what it did. A vendor that can demonstrate prompt curation, output monitoring, and high-impact-tier compliance documentation will outsell a more capable but unaccountable competitor — healthcare first, then financial services, then anywhere identity verification touches a regulated outcome.
The take: 2027 will be the year “AI governance tooling” becomes a budget line item separate from “AI tooling.” Vendors who don’t ship governance features will be repackaged as features inside vendors who do.
FAQ
Q: What is a high-impact AI classification under the OMB 2025 memorandum? A: The OMB’s Accelerating Federal Use of AI through Innovation, Governance, and Public Trust memorandum requires federal agencies to identify AI use cases that carry meaningful risk and apply safeguards to manage them. High-impact classification triggers obligations such as pre-deployment testing and human oversight before the tool is used in production.
Q: Why did the VA OIG flag VA GPT and Copilot Chat specifically? A: Because more than 15,000 VA staff were using them, including for clinical documentation tasks, even though the tools were not developed for clinical use and the VA did not centrally evaluate the prompts or outputs being applied to patient care decisions.
Q: What should healthcare organizations do now? A: Review which general-purpose AI tools are authorized in their environment, audit how staff are actually using them, and consult published frameworks such as the Health-ISAC white paper Policies and Safeguards for a Safe Use of AI and the HSCC Health Industry AI Cyber Governance Framework Implementation Guide to build oversight before regulators or auditors require it.
Key Takeaways
- Treat any AI tool with clinical-adjacent capabilities as high-impact by default; reclassifying upward is cheaper than an OIG-style remediation cycle.
- Prompt-level monitoring and output evaluation will move from nice-to-have to procurement requirement across regulated industries through 2026 and 2027.
- General-purpose copilots will increasingly fail healthcare RFPs unless paired with documented governance tooling.
- The VA’s 16-month remediation timeline is a signal: private-sector organizations should expect their own governance buildouts to take a year or more, meaning organizations that haven’t started are already behind.
- In healthcare AI, provable accountability now outweighs raw model performance — audit logs, evaluation pipelines, and patient safety integration.