The Care Stack

Fix the boring data before you buy the AI

Victor Wilson · · 4 min read

Every health system I talk to has an AI list right now. Some of it is in pilots. Some of it is in vendor demos that look very good. The pressure to move is real, and boards are asking about it.

The adoption is real too. An AJMC study found that 62.6% of US hospitals on Epic had adopted ambient AI documentation as of June 2025. The next wave is AI agents, software that can take a series of steps on its own, like filling in a form or sorting a work queue.

I’m not against any of it. But I’ve noticed a pattern. The AI projects that struggle usually aren’t struggling because of the AI. They struggle because of the data underneath, and that data was a problem long before anyone said the word “model.”

What I mean by boring data

By boring data, I mean the reference data that everything else depends on. Nobody puts it in a board deck, and it rarely has a clear owner.

A few examples:

  • Your provider list. Is each physician one record, with the right specialty, the right locations, and the right group? Or do they show up three different ways across your EHR, credentialing, and payroll?
  • Your locations and departments. Does a clinic have one name and one code everywhere, or does finance call it one thing and operations another?
  • Your cost centers. Do they map cleanly to the units that actually deliver care?
  • Your supply item master. Are items described consistently, with the right units, so that two records for the same product don’t look like two different products?
  • Your basic definitions. What counts as an encounter? A bed day? An active patient? If two reports disagree, is it because the data is wrong or because they define the measure differently?

None of this is exciting. All of it is what an AI tool reads when it tries to help you.

Why AI makes this worse

A dashboard built on messy data is wrong in a visible way. People see a strange number, question it, and someone goes to fix it. The mess slows you down, but it gets noticed.

AI is different. It takes messy inputs and produces confident, well-written outputs. A model that summarizes spend by department will happily summarize the wrong departments if your mapping is off. A forecasting tool will forecast from duplicate records. The answer will look polished, and fewer people will think to question it.

So AI doesn’t fix bad data. It hides it behind a better user interface. That’s a real risk in a setting where people make staffing and purchasing decisions from what they see.

Agents raise the stakes again. A wrong answer on a screen is a problem someone might catch. An agent that routes work or updates a record acts on the wrong answer. If your department mapping is off, the agent sends work to the wrong place, every time, and quickly.

Fix only what the AI needs

Fixing this data doesn’t require a multi-year program. You don’t have to clean everything. You have to clean what your first AI use cases depend on.

That’s the trick. Pick the AI project you most want to do. Then trace backward: what data does it need, where does that data come from, and which pieces of it would you trust today? You will usually find a short list of fixes. Maybe it’s the provider list and the department mapping. Maybe it’s the item master for a few categories.

Fix those, and give each one an owner who keeps it clean after the project ends. That last part matters most. Data that gets cleaned once and then left alone drifts back to where it was within a year.

There’s a second benefit. The same fixes help everything else you do. Clean provider data makes your productivity reports right. A shared definition of an encounter ends a lot of arguments in leadership meetings. Even if the AI pilot never ships, this work pays for itself.

The hard part is often internal. Asking for money to clean reference data is a tough sell next to an AI demo. It sounds like delay.

I’d frame it differently. The data work is part of the AI project, and should be in the same budget and plan. If a vendor says their tool will work on your data as is, ask them to prove it on your data, not on a demo set. Most will find the same gaps you would.

It also helps to be honest with your board. “We’re doing AI carefully, starting with the data it depends on” is a stronger position than “we bought a tool.” Boards understand that foundations come first. They usually just need someone to say it plainly.

There’s now an outside yardstick as well. In June, the Joint Commission launched a voluntary AI certification for health systems. It looks at how an organization governs and monitors AI and trains its people, not at individual tools. Knowing what data each tool reads, and who owns that data, fits naturally into that work.

Take the AI project at the top of your list. Before signing anything, ask your data team to write one page: the data this tool will read, where it comes from, who owns it, and how confident they are in it, on a simple scale from low to high.

If most of that page says low, you have your first project, and it isn’t the AI.

Discussion

    Join the discussion