Healthcare already wrote everything down. The chart note, the medication administered, the procedure performed, the code submitted, the payer's answer. The record exists — what's missing is a governed way to read it. We build that layer, and we point it at two deliberately separate questions.
Every finding cites the document it came from.
An enterprise AI system in a regulated setting has three layers. The language model is the least differentiated of them — everyone can rent the same one. The data underneath is the customer's, and it is theirs alone. What determines whether the system is trustworthy is the layer in the middle.
What the system is permitted to conclude on its own, what it must route to a human, and what happens when it is uncertain. Written down as rules, not left to a prompt.
Records are de-identified and tokenized before inference. The token preserves the patient's longitudinal journey without carrying their identity. Access is logged; authentication is multi-factor.
A finding that cannot be traced to a source document is not a finding. Every output carries the citation and the version of the rule set that produced it.
A coding finding you cannot defend to a payer is worthless. An outcome signal you cannot defend to a manufacturer's medical affairs team, or to a reviewer, is worse than worthless. In both of our divisions the deliverable is not an answer — it is an answer with its evidence attached.
These two programs share an ingest and extraction layer and nothing else. They are not two features of one product, and we do not sell them as a bundle. A customer of one need never encounter the other.
Optimizing insurance billing workflow against the treatment that was actually provided.
Medical billing companies, physician groups, and health systems — the people responsible for getting a correct claim paid the first time.
Analyzing treatment data — procedures, medications, side effects, outcomes — to inform how patients are treated and diagnosed, and to feed real-world usage signal back to manufacturers.
Health systems and quality teams, clinical research organizations, and pharmaceutical medical affairs and post-market surveillance groups.
Everything that goes wrong downstream in a claim was already decided upstream in the record. The code is derived from what was documented; the denial is derived from the code. So the useful place to stand is not at the claim — it is at the note.
Steps 1 and 2 are common to both divisions. Steps 3 through 5 diverge completely: the billing corpus and the clinical corpus have nothing to do with each other.
We would rather be believed than impressive. The platform is a working build, one industry design partner is engaged and validating output against real workflow, and we are actively onboarding new customers now.
Working software: ingest, de-identification, retrieval-grounded per-claim analysis with cited findings, and an authenticated interface. One design partner from the medical billing industry is reviewing output against real-world workflow, and we are onboarding additional customers.
Service-date-scoped rule versioning. Payer-specific rule sourcing. Documentation-to-code derivation from the note. Denial-history feedback so the system learns from actual adjudication outcomes.
The Clinical Evidence division: longitudinal patient assembly, outcome and side-effect signal, and de-identified aggregate reporting for manufacturers. Design work is underway; this division is not yet in production.
Antecedent Health is built on Amazon Web Services and runs on Amazon Bedrock. We built it because the engineering underneath it is the engineering healthcare organizations keep needing: PHI boundaries, retrieval you can audit, and governance that holds up when someone asks how the system reached a conclusion.