Build
One agent. Reads the case, cites the criteria, drafts the determination.
ADK · Agent Studio · Model Garden
BLUEPRINT · HEALTHCARE · PAYER
Prior Authorization
Read the case file, draft the determination, hold the audit trail.
Read the case file, draft the determination, hold the audit trail. Named scope, named timeline, named stack — ADK · Memory Bank · Model Armor · Agent Registry · 6 weeks.
Design ceiling
Hours, not days, on routine prior auth — and a clinician in the loop on every determination.
The agent system is designed to assemble the case file, draft the determination against the medical-necessity criteria, and hand a named reviewer a one-screen decision packet. Routine cases clear in a single review pass; the complex ones reach the clinician already framed.
The problem
The case file is the work. The decision is the easy part.
A regional payer running utilization management is paying nurses and medical directors to chase the same artefacts on every routine request — the EHR notes, the plan benefits, the medical-necessity criteria, the prior history. The reviewers are not the bottleneck on the decision; they are the bottleneck on the assembly. Cycle time stretches into days, members wait, and the UM team burns its scarce clinical hours on retrieval, not judgement.
The job is not to replace the reviewer. The job is to walk into review with the case file already assembled, the criteria already cited, and a draft determination on the page — so the reviewer reads, edits, and signs in a single pass. One agent, Memory Bank carrying the case state, Model Armor at the gateway on every PHI access, Agent Registry tracking every draft against the named reviewer who finalised it.
Agent architecture
Four pillars, this blueprint’s stack.
The platform’s four pillars, mapped to the components this agent system actually exercises.
Scale
Case state lives in Memory Bank across the assembly-and-draft pass.
Agent Runtime · Memory Bank
Govern
Every PHI access policy-checked. Every draft tied to a named reviewer.
Agent Registry · Model Armor · Gateway · Identity
Optimize
Eval set walks the criteria edges and the historical reviewer overrides.
Evals · Observability · Agent Analytics
Engagement · 6 weeks
Week by week.
Fixed scope, fixed price, fixed timeline. Here is what happens when.
Week 1
Discovery and criteria inventory.
Walk the current UM workflow with the medical director and the nurse reviewers. Inventory the medical-necessity criteria sources, the EHR connectors, the plan benefits feed. Shape the eval set against decision-quality, not throughput.
Week 2
Case-file assembler agent.
Stand up the ADK agent that reads the EHR, the plan benefits, and the criteria, and assembles the case file as a single structured object the reviewer can read end-to-end.
Week 3
Drafter agent and rationale.
Layer the determination drafter on top of the case file. Every draft cites the medical-necessity clauses applied. First eval pass against historical decisions sampled across the criteria edges.
Week 4
Governance and registry.
Configure Model Armor policies for PHI handling. Enrol the agent in Agent Registry against the named reviewer pool. Stand up the audit trail so every draft links to the reviewer who finalised it.
Week 5
Staging shadow run.
Run the agent in shadow against live requests for a week. Compare drafts to reviewer determinations with the UM lead. Tune the eval set against any drift before cutover.
Week 6
Production cutover and post-launch support.
Deploy to Agent Runtime with the named reviewer in the loop on every determination. Walk the runbook with the UM lead. Four-week post-launch support window for drift watch and reviewer feedback.
What it looks like in code
ADK Python — the agent definition.
The actual shape of the code your team owns at engagement end. Real ADK, real tools, real instruction copy.
agents/prior_auth/agent.py
python
from google.adk.agents import LlmAgentfrom google.adk.tools import FunctionToolfrom .tools import ( fetch_member_plan, fetch_clinical_history, fetch_criteria, record_determination,)prior_auth = LlmAgent( name="prior_auth_drafter", model="gemini-2.0-pro", instruction=( "You assemble the prior-auth case file and draft the determination " "against the medical-necessity criteria. Cite the clauses you " "applied. Hand the draft to the named reviewer for sign-off; never " "finalise a determination without the reviewer in the loop." ), tools=[ FunctionTool(fetch_member_plan), FunctionTool(fetch_clinical_history), FunctionTool(fetch_criteria), FunctionTool(record_determination), ],)What you walk away with
Code your team owns. Governance baked in.
Every blueprint hands the engineering team a deployed agent and the artefacts to run it themselves. No black box, no lock-in.
- One ADK agent (case-file assembler + determination drafter) deployed on Agent Runtime, in your Google Cloud project.
- The Git repo your team owns, with the agent code, tools, prompts, and tests.
- CI/CD pipeline wired to your repo, deploying on merge.
- Eval harness with a sampled-month corpus across criteria edges and reviewer overrides.
- Governance bundle: Agent Registry entries linking each draft to its named reviewer, Model Armor PHI policies, Gateway access controls.
- Agent Analytics dashboard plus an on-call runbook for the UM lead.
Book this blueprint as a pilot.
Two weeks. Named scope. Working agent on Agent Runtime at the end.
Code
Lives in your Git org, owned from commit one.
Governance
Model Armor and Agent Registry on day one.
Speed
Two weeks to a runnable pilot. Eight to production.
Not ready to talk? Take the 4-min readiness assessment