Rey Columna
All guides

What Google's Read-Only Rule Gets Right About AI in the Chart

Google launched Gemini for Healthcare on Nov 10, 2026. It reads Epic and Oracle Health but doesn't write anything back to the chart. I'm Rey Columna, CTO at Welby Health, and I think that limit is the smartest part of the launch.

By Rey Columna, CTOPublished 3 min read

On this page

What did Google launch?

A set of AI agents for hospitals and health plans, built on Gemini Enterprise. Google announced it on November 10, 2026.

It can pull a patient's records into one summary, prep a clinician before an appointment, review coverage and policy rules, and work out what a payer will cover. It plugs into Epic and Oracle Health, plus data like the NPI registry, ICD-10 codes and CMS coverage data.

It's a preview. It isn't generally available yet.

Why does read-only matter?

The line I keep coming back to is in Google's press release. It connects to Epic and Oracle Health "without writing back to the EHR."

So the agent can read the chart, but it can't change it. A person still makes the decision and updates the record.

I think that's the right call. If an agent reads something wrong, a person sees a bad summary and usually catches it. Writing is different. A wrong medication, a wrong allergy, a note filed on the wrong patient. Once it's in the chart, other people and other systems trust it. It gets copied into referrals and claims, and it can follow a patient for years.

A bad summary costs someone a few minutes, while a bad write can cost a lot more and is much harder to find.

How should any AI agent handle a chart?

The same way. Google's rule is a good default for anyone building agents that touch patient data.

  • Start read-only. Let the agent read, summarize and draft, and let people make the change.
  • Show where every answer came from. If the data is missing, the agent should say so instead of guessing. That's how I built Markus, our clinical AI teammate for nurses. In clinical work, a confident wrong answer is worse than no answer.
  • Give the agent the person's permissions and nothing more. Google says its agents inherit the user's role permissions, which is the right model.
  • If you do let an agent write, start with one narrow task and have a person approve each change. Widen it only after it has a clean track record.

Agents will write to the chart eventually. I'd want a long record of clean reads before I let one.

What should you check before using it?

A few things before anyone gets excited.

  • It's a preview, and Google hasn't given a date for general availability.
  • Google says it "supports HIPAA compliance." That isn't the same as your BAA covering it. Confirm that before patient data goes anywhere near it.
  • It isn't only Gemini. It supports models from Google, Anthropic and OpenAI, and sends high-volume admin work to cheaper models. Most of what Google built here is the connections to EHRs and payer data, not the model.
  • NextGen support is still in development. If your clinics run on another EHR, check before you plan around it.

Google also says customer data stays in the organization's own Google Cloud environment and isn't used to train models. Get that in writing for your own setup.

Where will it help first?

Look at who signed on: BCBS of Massachusetts, Hartford HealthCare, HCA Healthcare, Highmark Health, Northwell Health, Optum and Stanford School of Medicine. That list mixes hospital systems and health plans.

Then look at the work Google lists. Coverage and policy review, benefit checks, eligibility, records prep. That's back-office work, where a lot of staff hours go, and a read-only agent can help with it today without touching the chart.

That's where I'd start too. Find the paperwork that eats your team's week, let the agent read and draft it, and keep people on the decisions.

If you're working through the same questions, email me at hello@reyco.me.

Sources