01The problem
The easiest way to build AI for clinics is to build eight products: a phone bot, a fax reader, a scribe, a coder, a billing assistant, a portal, a chatbot, an authorization tool. Each is good at its job. None of them know what the others did, so the clinic staff carry information between them by hand.
A practice we visited had four vendors and a shared spreadsheet. The scribe heard the physician order a lab; the front desk learned about it when the patient called to ask where to go; billing learned about it when the claim was denied for a missing diagnosis code.
02What we saw
Every hand-off was a place where context was lost. The reminder text did not know the visit had been rescheduled. The authorization request did not know the note already documented six weeks of physical therapy. The copilot answering a question about a patient had never seen the fax that arrived that morning.
The errors were not dramatic. They were small, constant and expensive: a repeated question, a re-typed value, a denied claim for information that existed all along.
03How Spriggan handles it
We made a different bet. Every agent in Spriggan reads and writes the same chart, the same schedule and the same rules. When the scribe hears the physician say recheck the A1c in three months, the order exists for the receptionist who books the lab, for the billing agent who codes the visit, and for the copilot that answers, months later, whether the patient is due.
Ask the copilot a question and the answer comes with its sources: a Quest result from March, a care plan from January, a medication list from August. Ask it to act and the action is drafted from the same sources, then waits for a person’s OK.
04What AI made possible
Shared context is not a new idea; every EHR has one database. What is new is that the agents can read it the way a clinician does. A language model can take the fax, the note, the schedule and the payer rule together and act on all of them at once, which is exactly what a good medical assistant does and exactly what no rules engine could.
It is also what makes the agents trustworthy. When something looks wrong you can see what the agent read and what it did, because there is only one place it could have read it from.
Nothing gets re-asked. Nothing gets re-typed. Nothing is lost between two systems that were never introduced to each other.
05What changed
The spreadsheet went away. The clinic stopped being the integration layer between its own tools. Reminders know about reschedules, authorizations know about the note, claims know about the authorization, and the physician’s question about a patient gets an answer that includes this morning’s fax.
This is the part of the product that is hardest to demo and easiest to feel, which is why we wrote it down.