Start by reading a year of your team's email, read-only, to see where the time goes. Then automate the work that repeats, one job at a time, with a person approving anything that leaves the building.
8 questions owners asked us, from 3 calls. Updated 2026-10-23.
The load depends mostly on how the agents are built. Ask whoever builds them to size it. Our preference is to avoid running two copies of the same data, and to connect through a scoped API key, which limits what each tool can reach.
Start with a master list of every AI program your team uses, and a short usage policy that names the approved tools and who signs off on each new one. Set up a team account for AI tools. Then a read-only review of a year of email shows where the time goes and what to automate first.
Keep health information in its own database, and control access at the API level, the point where each tool asks for data. Limit access to the few people who need it, and have a technically skilled person in the loop for anything that touches it. Every outside service that handles the data should sign a business associate agreement (a contract that binds it to health privacy rules). The law on AI and health data is still unclear, so keep a record of where each piece of data went.
The bigger need is a clear record more than layers of locks: you should be able to show an auditor exactly what happened to the data and where it went. Give each agent as little information as its task needs, and keep its access scoped to that task.
Usually the tool meets a case it has not seen before, or a field the software handles in an unusual way. To find out, run the same test case many times, often a hundred runs, and see whether the result changes. Each difference points to something to fix, either in the tool or in the software, and a person reviews the exceptions it flags.
For now, new files often have to be added by hand before a tool can use them. Check this before you rely on a shared library. Some tools cap how much you can upload at once. Start with smaller batches. A person reviews what the tool produces before anything is published.
Simple, repeated questions can be answered from a knowledge base the team builds over time, using the questions patients actually ask. Complex medical questions take longer, and expect that build to take a year or more. Keep a person on any call that goes beyond what the knowledge base covers.
Checking that coverage is active is different from getting the benefit details, such as whether the deductible is met and what the patient will pay. An agent can sign in to each insurer's portal and read those fields, but every portal is set up a little differently, so each one is built and tested on its own. A person checks the result before anything is used for a claim.
Today. Each fax is opened by hand, checked against existing patient records, and the patient and order details are typed in.
What changes. Once the audit shows how these faxes arrive and repeat, an agent could read each one, pull out the patient and order details, and draft the record. A person would check each draft before it is saved, and any fax the agent does not understand would go to a person.
Today. A staff member signs in to each insurer's portal with their own login, copies the member details, and notes the benefits by hand.
What changes. Once the audit confirms how often this repeats, an agent could sign in to each portal, read the benefit details, and file them in the patient record. A person would check the notes before the patient is contacted.
Today. Staff call each patient to book a time once the order and benefits are ready. Patients often ask about their benefits, the test and the process during that call.
What changes. Once the audit confirms how often this repeats, an agent could draft the booking messages and offer open times, with a person approving each one before it is sent. Any call that raises a medical question would stay with a person.
Today. Staff open each newly scheduled patient, check whether the intake forms are signed and whether a copay was paid, then change the status by hand so the next step can run.
What changes. Once the audit confirms it, an agent could check the forms and payment records and move each patient to the next status once both are complete. A person would review any patient the agent marks incomplete before anyone contacts them.
Today. Staff pick the billing code and the diagnosis on each record, add the insurer, and submit the claim. Some fields fill in automatically and others have to be keyed by hand.
What changes. Once the audit confirms it, an agent could fill in the codes and fields from the chart and prepare the claim. A person would review every claim before it is submitted.
Everything that matters in a business goes through an inbox at some point. Read a year of it, read-only, and you can see where each person's week goes. Then automate the work that repeats, one job at a time.
