What I Learned About AI in Healthcare at Harvard — And Why I Came Back to the Data

Between December 2025 and March 2026 I went through two Harvard courses on agentic AI in healthcare, back to back. I walked in expecting to build AI systems. I walked out building a data transparency platform. This post is the honest version of why.

Course 1 — Agentic AI Intensive, December 2025

The first course was a 2.5-week Harvard Data Science Review intensive on agentic AI in clinical settings. The project I took through it was a multi-agent clinical decision support system for addiction medicine consults in the emergency department — essentially an attempt to model the way our consult service actually reasons through a patient who presents with intoxication, withdrawal, co-occurring psychiatric illness, and an urgent disposition problem.

The course introduced the AGENT framework for thinking about AI governance in healthcare: where humans need to remain in the loop, where AI can contribute, and critically, where AI should not be making autonomous decisions. The framework takes safety seriously because the consequences of being wrong in clinical settings are not like the consequences of being wrong in consumer software. You cannot quietly patch a mistake in someone's chart.

The biggest takeaway from that course had nothing to do with model architecture. It was about oversight. It was about the discipline of mapping, for every step of a workflow, who is responsible and what failure looks like. That discipline matters. It is easier to write than to practice.

The interesting question about AI in healthcare is not “what can it do?” It is “at which step are we willing to let it decide alone, and how will we know when that step fails?”

My addiction medicine prototype was cited by Prof. Xiao-Li Meng, Editor-in-Chief of the Harvard Data Science Review, in Issue 8.1 (Winter 2026), in an article titled “Navigating the AI Safari.” That meant something to me, because the citation was not about the model. It was about taking governance seriously in a high-stakes clinical domain.

Course 2 — Agentic AI and AI Evaluation in Healthcare, February 2026

The second course ran 2.5 weeks starting February 24, 2026. That happens to be the same day OversightReports.com went live. I was running both at once, which was stupid in the traditional sense and useful in the specific sense that it forced the comparison I am about to make.

I walked into the second course carrying something concrete: 1.9 million rows of CMS nursing home data. Inspection citations. Complaint records. Penalty histories. Payroll-based staffing logs. Ownership disclosures. HCRIS cost report data. Eighteen different federal source categories — and almost no usable way for a family, a discharge planner, or a plaintiffs' attorney to look at a single facility and see the full picture.

In the first few days of the course, the nursing home transparency concept crystallized. I was sitting in a lecture about evaluation frameworks for agentic AI systems in healthcare. I was listening carefully. And I realized something uncomfortable.

The realization

The real bottleneck in nursing home safety is not that we lack an algorithm. It is that the data we already have — data the federal government is already collecting with public money — is not accessible in any form that lets a non-specialist act on it.

When I look at OIG reports documenting antipsychotic prescribing gaming. When I look at the payroll-based staffing data showing thousands of facilities with zero-bedside-RN days in violation of 42 CFR 483.35(b). When I look at the $11.26 billion in related-party transactions across 13,324 facilities in HCRIS Worksheet A-8. When I look at the 599 chain-ownership groups I mapped for OversightReports. None of that required an LLM to surface. It required someone to join the tables, read the regulations, and render the results in a human-readable format.

You do not need agentic AI to do that. You need a data engineer who is also a clinician, and you need the willingness to read the CMS documentation end to end.

That is not a rejection of AI in healthcare. I think clinical AI is a serious discipline and I want to work on it for the rest of my career. What I am rejecting is the order of operations. Transparency first. Interface first. AI second. If the raw data is still trapped in regulator-formatted spreadsheets, no amount of model sophistication on top of it will make it accessible. You will just get a very fluent model describing data nobody else can verify.

Where I came down

By the end of the second course I had made a specific commitment: OversightReports.com is a transparency project first. There will be analytics. There will probably, eventually, be some well-scoped AI features — the kind of features that survive the AGENT framework, where humans remain in the loop on every consequential decision. But I am not leading with AI. I am leading with the data, the citations, the methodology, and the interface.

The Harvard Data Science Initiative featured my testimonial in advertising for the program. I am grateful and I stand behind what I said about the course. What I would add, and what I'm saying here for the first time, is that the most important thing I learned at Harvard about AI in healthcare is when not to use it.

The short version

I did not come back from Harvard disillusioned about AI. I came back convinced that in nursing home oversight specifically, we are one step earlier than the AI question. We are still at the data transparency step. Someone has to do that step, do it carefully, and do it with the same discipline that the AGENT framework applies to clinical AI. That is what OversightReports.com is.

— Robert Benard, NP
March 7, 2026