How Agnodice is designed to work
Agnodice is designed around one idea: software agents may propose work, but deterministic code decides what is allowed. Each claim moves through a fixed lifecycle, each agent action is checked against written limits, and every decision is recorded. Autonomy is designed to be earned in four levels. This page describes the design, not a product that is operating today.
The claim lifecycle
Every claim is designed to move through the same stages, recorded as events that are never edited.
Intake
- Received
- Coded
- Scrubbed
- Submitted
- Acknowledged
- Adjudicated
Outcome
- Paid
- Denied
- Partly paid
Resolution
- Posted
- Appealed
- Written off
Four steps for every task
Each task is designed to follow four steps.
Read a minimized case
The agent is designed to see only the data the task needs, never more.
Propose through typed tools
The agent is designed to ask for an action using tools with strict input and output shapes, not free text.
Check against written limits
Deterministic code is designed to compare the proposal with written limits and to allow or refuse it, with reasons.
Record the decision
The agent, its version, the model, the tools used and the policy applied are designed to be written to an append-only log.
Autonomy is designed to be earned in four levels
- L0
Shadow
The agent records what it would do. Nothing is acted on.
- L1
Suggest
Staff review and approve each proposal.
- L2
Act, with audit
The agent acts on reversible work while a sample of its decisions is reviewed.
- L3
Autonomous, monitored
The agent acts alone within its limits while outcomes are monitored.
An agent is designed to move up a level only after measured outcomes support it, and to drop back automatically after a serious error.
What we are not doing
- We are not making clinical decisions.
- We are not replacing a clinician's judgment on care or documentation.
- We are not asking any model to sign or attest on a practice's behalf.
Tell us about your practice
If denials and unpaid claims take real time at your practice, we would like to hear how. Please leave out patient information.
