A large share of US healthcare spending goes to administration rather than care. Research published in Health Affairs puts it at roughly a quarter to a third of the total. Prior authorization alone takes about 12–13 hours a week of physician and staff time, according to the American Medical Association's survey.
Most of that work is reading documents that already sit in your systems, checking them against rules someone already wrote down, and assembling a packet for a person to send. A language model can help with that. The problem is that the obvious way to use one sends the chart to a public model API, and most compliance leads stop the project right there. They are right to.
This note covers the narrower version: an administrative agent that runs inside the environment where protected health information already lives.
The scope is administrative work
Four kinds of work fit:
- Intake. Reading referral and registration documents, checking fields against the eligibility response, and listing what is missing so a person can chase it.
- Billing support. Checking a claim against the payer rules your team already maintains before it goes out, and drafting the appeal packet for a denial so a biller can review and send it.
- Prior-authorization packets. Pulling the documents a payer asks for from records your staff already open for this job, putting them in the payer's format, and flagging gaps. A person submits. The agent does not judge medical necessity.
- Internal policy lookup. Answering staff questions from your own policies and payer manuals, with the source quoted.
Where the data goes
The model and the agent run inside the environment you already treat as in scope for PHI. That can be your own servers or a hosting arrangement your counsel has already reviewed. At inference, no prompt, document, or output goes to a public model API.
Be honest about what each workflow touches. An eligibility check needs demographics and coverage data. A prior-authorization packet needs clinical documentation, because that is what the payer asks for. So the packet workflow sits fully inside the boundary, and the agent gets the same access as the staff role that builds packets today, or less. It never gets broader access to the EHR than that role has.
Every read and every draft goes to an audit log, so a compliance lead can answer "what did the agent see, and for which patient" without asking an engineer.
What we do not certify
We do not issue a HIPAA certificate, and we make no compliance claim about the system. Your counsel decides what is in scope and whether the setup meets your obligations.
A Business Associate Agreement is a contract between a covered entity and whoever creates, receives, or handles PHI on its behalf. If our people or a hosting provider will touch PHI during the work, your counsel decides whether a BAA is needed and what it covers. It is a contract you sign with a party. It is not a property of the software.
Counsel will usually look past HIPAA too: state privacy laws such as California's CMIA, the X12 standards for claims and eligibility transactions, and CMS rules if Medicare or Medicaid billing is involved. Start that review at the same time as the technical scoping, not after it.
Start with a baseline on one workflow
Pick one workflow, one service line, and if possible one payer. Prior-authorization packets for a single specialty are a common choice. Then measure it as it runs today:
- packets or claims per week
- staff minutes per item, including rework
- how often an item comes back incomplete or denied, and why
- days from request to decision
Next, agree a failure budget with the people who own the work. Write down what the agent may get wrong, how often, and what stops it. A packet sent to a person with a gap flagged is fine. A packet that invents a missing document is not, and one such error stops the workflow until someone fixes the cause.
The staff who build packets today are the reviewers. They also supply the evaluation set: past packets with the right answer known, including the awkward ones. The agent runs against that set before it touches live work, and after every change.
Two things slow these projects more than the model does. One is data quality: duplicate records and free-text notes with codes buried inside them. The agent should flag those, not guess. The other is copying the manual process step for step. Some steps exist only because a person had to do them by hand, and those are worth dropping before you build.
What you keep
A running system inside your boundary, a data-flow map, an access model, and an audit log. You still own the policy and the sign-off. The work runs on a fixed milestone, a senior practitioner leads it, and it ends when the milestone is met.
If charts cannot go to a public model API and you want an agent on administrative work, read Sovereign AI for the HIPAA boundary, or book a scoping call. It takes thirty minutes to find out whether the environment you already have is enough. If the blocker turns out to be a clinical decision, we will say so.
Sources: Health Affairs, "The Costs of Health Care Administration in the United States and Canada" (2020). American Medical Association, Prior Authorization Physician Survey (2024).