CMS Needed Five Pages to Say Who Owns Its AI

Your AI policy says what is permitted. It does not say who answers.

On June 11 the Centers for Medicare and Medicaid Services published five pages in the Federal Register establishing a new Office of Health Technology and Products. The Secretary had approved it two days earlier. Most coverage read it as a reshuffle. It is.

The useful part is how carefully it gives things away.

The summary paragraph hands the new office enterprise leadership for technology modernization, then immediately subordinates it to the Chief Information Officer for governance, cybersecurity, architecture and capital planning. The artificial intelligence function reads the same way. The office leads enterprise AI strategy and advises the Administrator on AI risks and governance, "in coordination with the CIO as required under statutory responsibilities." Purchase approval stays under FITARA. The Centers stay the business owners.

I counted that qualifier seven times in the list of the office's own functions, and once more in the summary paragraph.

Read quickly, that is bureaucratic throat-clearing. Read as a document, it is an accountability map. Strategy here. Security there. Architecture there. Spend approval somewhere else again. Nobody who reads those five pages thinks one office owns it all.

Now find your own version of that document.

Most health systems have an AI policy. They often look broadly alike: acceptable use, a review committee, a nod to bias. They describe permission. What I have not seen is a document naming who answers, by name and title, when a specific model produces a specific bad output on a Tuesday afternoon.

That gap stays theoretical until the model is one you bought.

The AI found in health systems did not come from a data science team. It arrived inside products. Triage and acuity logic in the EHR module. The coding assistant in the revenue cycle contract. Sepsis alerting bundled into a monitoring agreement. None of it was classified as AI at purchase, because the purchase order said module, or upgrade, or annual release. An organization can hold a completely defensible position on every model it built and nothing at all on the ones it bought.

Many people have experienced the exact shape of that gap. In the health system where I ran business intelligence and data analytics, I could trace the lineage of any metric on any executive dashboard we published back to the source system. I could not have told you, without going and asking each vendor, which of those source systems had a model running inside it.

The force underneath this is organizational rather than technical. Accountability that has never been written down defaults, at the moment of failure, to whoever happens to be in the room. Written and split beats unwritten and assumed. CMS just demonstrated the first in public.

Three moves, in the order I would actually do them.

Start where you have full control, your own warehouse. When a purchased system writes a risk score or a predicted length of stay into your environment, it arrives as a column, and nothing in the schema tells an analyst it was predicted rather than measured. It gets picked up, built on, and eventually informs a decision, carrying no record of where it came from and no sense of how often it is wrong. Add a provenance attribute marking any field a model produced. This is not free. It touches ingestion, the modeling layer and the catalog. It is still the cheapest of the three and the only one needing nobody's permission but yours.

Then write the map for the twenty or thirty vendors that matter. Not every agreement in the building. Take your largest clinical and revenue cycle systems by spend, and for each one answer two questions in writing: does this product make or rank a decision, and who here answers for it. You will need an executive sponsor to get the contract files opened, and my own research is direct about what happens without one. A short list with a named sponsor beats a complete inventory nobody funded.

Then move the terms at renewal, where you can. Ask what the model does, what it was trained on, what change obligates the vendor to notify you, and what you can see in the logs. Mid-tier vendors and net-new purchases are the winnable ones. No single customer moves standard terms with the dominant EHR vendor, and that is exactly where the largest embedded exposure sits, so that conversation belongs in a user group or a customer coalition rather than in your own redlines.

CMS needed five pages and eight hedges to write down who owns what, and the result is not elegant. It exists, which is the point. Your version does not have to be good. It has to name people.

If you found this briefing valuable, share it with a colleague in data and analytics.