Over the last five months, Grata has built what looks like an unreasonable amount of software.
We built recovery companions and the shared infrastructure beneath them. We built intake, eligibility, telehealth, prescribing, clinical documentation, drug-screening workflows, scheduling, and follow-up. We built systems to find, interview, contract, credential, train, and activate clinicians. We built our own referral network, partner portal, CRM, and clinical tools for recovery professionals. We built software for claims, denials, pharmacy fulfillment, at-home testing, and the internal work of the company itself.
The orthodox startup answer is that we have poor discipline. Software companies should build what differentiates them and rent everything else. Clinics, especially, should deliver care and leave the software to vendors.
I used to explain our decision economically. Healthcare is full of administrative labor; software compresses it; the clinic becomes more profitable. That is true. It is also the least interesting consequence of what we are doing.
The deeper reason is epistemic. Software is not downstream of a healthcare institution. It determines what kind of institution can exist.
The normal build-versus-buy calculation assumes that the end is already known and the tool is merely a means of reaching it. John Dewey thought this separation was false. Means and ends arise together inside activity. The route we take changes the destination we can imagine.
This is obvious in a clinic. Intake is not paperwork that happens before care; it determines who enters care and what the clinician is able to know. Recruiting is not an HR function; it determines the clinical judgment the institution can exercise. Billing is not an administrative epilogue; a denial determines whether the care model is economically real. The companion is not a marketing channel; it changes how early we can know someone and what continuity after a visit can mean.
Generic software arrives with different assumptions. The EHR sees a patient as a sequence of encounters. The applicant system sees a clinician until the day they are hired. The CRM sees a referral until the moment it converts. The billing system sees a claim detached from the care that produced it. Each product may perform its assigned function perfectly. Together, they partition the institution.
The handoffs that follow are not implementation mistakes. They are consequences of the ontology. Buying software means buying its account of what exists, which relationships matter, and where an event begins and ends.
An API can move fields across those boundaries. It cannot make the underlying accounts of the world agree. Interoperability is not continuity.
This matters because medicine is a practical art before it is an information problem. Aristotle distinguished practical wisdom from abstract knowledge because action is concerned with particulars. Michael Polanyi made the adjacent point that “we can know more than we can tell.” A clinic’s most valuable knowledge often first appears as a judgment someone cannot yet formalize: the phrase that makes a frightened patient finish intake; the detail that tells a clinician to ask one more question; the pattern that tells an operator this missed visit is different from the others.
Off-the-shelf software begins after those judgments have been generalized. We build while they are still being formed. The distance between the person who encounters reality and the person who changes the system can be a few feet. That is how tacit knowledge becomes explicit without pretending that judgment can be eliminated.
Gilbert Simondon offers a useful distinction. An immature technical object is abstract: each component performs an isolated function, and something outside the system must continually reconcile the pieces. As the object becomes concrete, its functions become integrated. A consequence produced in one part becomes useful to another. The parts begin to condition one another’s operation.
Most clinics run on an abstract machine. One vendor schedules the visit, another hosts it, another records it, another writes the claim, and another manages the clinician. People become the corrective mechanism between them. The apparent savings from buying software return as administrative labor, delayed decisions, and lost knowledge.
We are trying to build the concrete version.
The audio from a Grata visit can become a transcript, a clinical note, a plan, a patient summary, and the basis for follow-up. A referral does not end when a name enters a queue; the referring clinician can see whether the patient entered care and what happened next. Demand from a recovery companion can reveal where we should recruit clinicians and pursue payer contracts. Human review of clinician interviews improves the next round of screening. A denied claim can become a rule applied before the next claim leaves the system.
The important property is not that Grata possesses all the data. Large healthcare institutions already possess extraordinary amounts of data. It is whether a lesson learned in one part of the institution can travel far enough to alter another.
Charles Peirce wrote that “the essence of belief is the establishment of a habit.” Different beliefs are real, on his account, insofar as they produce different conduct. The same test can be applied to an institution. A clinic has not learned something because a clinician wrote it in a note, an operator added it to a spreadsheet, or a billing team discussed it in a meeting. It has learned only when it reliably acts differently.
That is the standard behind our software. Can an observation become a changed habit of action? Can the change survive the person who first understood it? Can it reach the next patient, clinician, claim, referral, and state?
This does not mean building every primitive. We use cloud infrastructure, payment rails, pharmacies, laboratories, and clearinghouses. The boundary is not ownership for its own sake. We rent primitives. We build the layer in which Grata represents the world, makes decisions, and remembers their consequences.
Nor does learning mean allowing a medical system to modify itself invisibly. In healthcare, judgment must remain accountable. The point is to make every consequential decision traceable, every operational rule testable, and every lesson capable of becoming durable under human and clinical governance.
The real object is not a suite of applications. It is continuity between perception and action.
This is what I mean by an AI-native clinic. Not a conventional clinic with a model attached, and not an attempt to automate medicine into the absence of people. It is an institution whose experience can alter its future conduct: whose model of the world becomes less wrong because reality is allowed to answer back.
The goal is simple: the next patient should inherit everything the last patient taught us.
That sentence is both a moral claim and a technical specification. To make it true, Grata must own the path by which experience becomes memory, memory becomes judgment, and judgment becomes action.
That path is the software. So we build it.