By James Aspinwall, co-written by Alfred Pennyworth (my trusted AI) — March 7, 2026, 18:55
Abridge turns doctor-patient conversations into billable clinical documentation. WorkingAgents governs AI agents so enterprises can deploy them with access control, audit trails, and compliance guardrails. Healthcare is the highest-stakes environment for AI governance — and Abridge is the fastest-growing clinical AI platform in the market. Together, they address a problem that’s about to become urgent: governing the next generation of autonomous healthcare agents that don’t just document, but act.
What Abridge Brings
Abridge is a $5.3B clinical AI company (Series E, June 2025, led by a16z and Khosla Ventures) with $100M+ ARR deployed across 200+ health systems and projected to support 80 million patient conversations in 2026.
Core product — Ambient Clinical Documentation:
- Listens to clinician-patient conversations and generates structured SOAP notes in real time
- Automates 91% of medical note creation across 55+ specialties and 28 languages
- Linked Evidence — maps every AI-generated summary span back to the source audio, allowing clinicians to verify before signing
- 97% hallucination detection rate (vs. 82% for GPT-4o) via proprietary confabulation elimination
- 73% reduction in after-hours documentation, 61% reduced cognitive burden
Contextual Reasoning Engine:
- Dynamically integrates prior encounter data, health system guidelines, and clinician preferences
- Generates billing-ready documentation with coding alignment
- Powers real-time prior authorization — maps conversation to orders, problem lists, and payer policy during the visit itself
EHR Integration:
- Epic’s first “Pal” — native integration inside Haiku and Hyperspace
- Athenahealth partnership extending reach to smaller/independent practices
- Availity collaboration connecting to 95% of payers and 3M providers for automated prior auth
Proprietary AI:
- Custom speech recognition and note generation models trained on 1.5M+ consented, de-identified medical encounters
- 16% lower word error rate than Google Medical Conversations and OpenAI Whisper v3
- No dependency on third-party LLMs — unlike competitors using GPT-4
Enterprise customers: Kaiser Permanente (24,600 physicians), Mayo Clinic (enterprise-wide), Johns Hopkins, Duke Health, UPMC (12,000 clinicians), Yale New Haven, Emory Healthcare.
What WorkingAgents Brings
WorkingAgents is an AI governance platform with three gateways (AI, Agent, MCP) that puts permission boundaries, guardrails, and audit trails around every agent action. The platform governs 60+ MCP tools across CRM, task management, communication, scheduling, and monitoring — all with capability-based access control and full observability.
What WorkingAgents doesn’t have: clinical AI, healthcare domain expertise, EHR integration, or HIPAA-specific compliance infrastructure. What it does have: the governance architecture that healthcare AI will need as it moves from documentation to autonomous action.
The Strategic Thesis
Abridge has conquered ambient documentation. The next frontier — and the reason they raised $300M — is agentic clinical workflows: AI that doesn’t just write the note, but submits the prior auth, orders the lab, schedules the follow-up, and messages the patient. Abridge’s Contextual Reasoning Engine and Availity partnership are the first steps.
But autonomous clinical agents operating across EHRs, payer systems, scheduling platforms, and patient communication channels need governance that goes beyond HIPAA encryption and access controls. They need:
- Per-agent permission scoping — a prior auth agent shouldn’t access clinical notes; a documentation agent shouldn’t submit orders
- Multi-step workflow governance — retries, timeouts, human-in-the-loop checkpoints for high-risk actions
- Cross-system audit trails — a single log showing the agent read the encounter, queried the payer API, submitted the auth, and notified the provider
- Guardrails at every checkpoint — pre-execution validation, real-time monitoring, post-execution inspection
This is exactly what WorkingAgents builds.
The Gap Analysis
| Abridge Gap | WorkingAgents Solution |
|---|---|
| No multi-agent orchestration framework | Agent Gateway manages complex workflows with retries, timeouts, fallbacks |
| No cross-system permission scoping | Virtual MCP Servers define exactly what each agent can access per role |
| No unified audit trail across external systems | Every tool call logged with full context, cost, and guardrail evaluation |
| No agent-to-agent coordination protocol | MCP + Google A2A protocol for cross-platform agent discovery |
| No cost attribution across agent workflows | Token-level cost tracking by user, team, model, and workflow |
| No framework-agnostic agent governance | Works with any agent via HTTPS and Secure WebSocket APIs |
| Documentation-focused — not yet governing autonomous actions | Governance architecture designed for agents that act, not just observe |
| WorkingAgents Gap | Abridge Solution |
|---|---|
| No healthcare domain AI | 55+ specialty clinical documentation with proprietary models |
| No medical speech recognition | ASR tuned for clinical conversations, 28 languages, cross-talk handling |
| No EHR integration | Epic Pal, Athenahealth, deep clinical workflow embedding |
| No clinical safety framework | 97% hallucination detection, Linked Evidence verification |
| No healthcare compliance infrastructure | HIPAA-compliant, SOC 2, 256-bit encryption, US-based data centers |
| No payer/revenue cycle integration | Availity partnership, real-time prior auth, billing-ready documentation |
| No clinical outcome data | 73% less after-hours documentation, 81% workflow satisfaction improvement |
Synergy Areas
1. Governed Agentic Clinical Workflows
The highest-value opportunity. Abridge’s Contextual Reasoning Engine already generates orders, problem lists, and prior authorizations from conversations. The next step is agents that execute these actions across systems:
Prior Authorization Agent
✓ Abridge: read encounter summary, extract diagnosis codes
✓ Availity: query payer policy, submit authorization request
✓ EHR: update order status, attach approval
✓ Notification: alert provider of approval/denial
× EHR: modify clinical notes (documentation agent's scope)
× Billing: adjust charges (revenue cycle agent's scope)
× Patient portal: send messages (communication agent's scope)Each agent has a defined scope. The prior auth agent can read clinical data but cannot modify notes. The documentation agent can write notes but cannot submit orders. WorkingAgents’ Virtual MCP Servers enforce these boundaries — configuration, not code.
The workflow runs through WorkingAgents’ Agent Gateway with human-in-the-loop checkpoints: “Agent wants to submit prior auth for MRI — approve or deny?” High-risk actions (medication orders, procedure authorizations) require explicit clinician approval. Low-risk actions (scheduling follow-ups, sending appointment reminders) execute autonomously within defined guardrails.
2. Cross-System Clinical Audit Trails
Healthcare compliance requires knowing exactly what happened, when, and why. Today, audit trails are fragmented across EHRs, payer portals, scheduling systems, and communication platforms. Investigating an incident means stitching logs from five different systems.
WorkingAgents provides a unified audit trail across every system the agent touches:
{
"workflow": "post-encounter-processing",
"encounter_id": "ENC-2026-03-07-1423",
"clinician": "[email protected]",
"steps": [
{
"agent": "documentation",
"tool": "abridge.generate_note",
"result": "SOAP note generated, 94% confidence",
"guardrails": { "hallucination_check": "passed", "pii_scope": "within_encounter" }
},
{
"agent": "coding",
"tool": "abridge.extract_codes",
"result": "ICD-10: M54.5, CPT: 99214",
"guardrails": { "code_validation": "passed" }
},
{
"agent": "prior_auth",
"tool": "availity.submit_auth",
"result": "approved_instant",
"guardrails": { "policy_match": "passed", "human_approval": "not_required" }
}
],
"total_latency_ms": 3200,
"total_cost_usd": 0.42
}One log. Every agent, every tool call, every guardrail evaluation, every cost. Regulators, compliance officers, and risk managers get a single source of truth.
3. Role-Based Agent Access for Health Systems
A 24,600-physician deployment like Kaiser Permanente has dozens of roles with different needs and risk profiles. WorkingAgents’ permission model maps naturally to healthcare hierarchies:
Attending Physician Server
✓ Abridge: full documentation (generate, edit, sign)
✓ Abridge: prior auth submission
✓ Abridge: order generation
✓ Patient messaging (with guardrails)
Resident Server
✓ Abridge: documentation (generate, edit — cannot sign)
✓ Abridge: prior auth (draft — requires attending approval)
× Order generation
× Patient messaging
Medical Scribe Server
✓ Abridge: documentation (generate, edit — cannot sign)
× Prior auth
× Order generation
× Patient messaging
Billing/Coding Server
✓ Abridge: read encounter summaries
✓ Abridge: extract and validate codes
× Clinical note editing
× Patient communication
× Order generationEach role gets a Virtual MCP Server with exactly the tools and permissions it needs. A scribe can generate and edit notes but cannot sign them. A resident can draft prior auths but needs attending approval. Billing staff can read encounter summaries and validate codes but cannot modify clinical content.
4. Patient Communication Governance
Abridge already generates patient-facing visit summaries at 8th-grade reading level. As health systems add AI-driven patient communication (follow-up reminders, medication adherence, appointment scheduling), governance becomes critical:
- Content guardrails — block AI from providing medical advice, generating diagnoses, or making treatment recommendations in patient messages
- PII controls — ensure patient communications contain only that patient’s information
- Tone and language validation — enforce health literacy standards, cultural sensitivity
- Approval workflows — high-sensitivity communications (test results, medication changes) require clinician sign-off
WorkingAgents’ three-checkpoint guardrail system (pre-execution, real-time, post-execution) applies directly: validate the message content before sending, monitor for scope violations during generation, inspect the output for PII leakage before delivery.
5. Multi-Health-System Agent Orchestration
Large health systems like Kaiser Permanente and UPMC operate across dozens of hospitals and hundreds of clinics, each with different EHR configurations, payer contracts, and clinical protocols. An agent that works at one facility may violate policy at another.
WorkingAgents’ gateway architecture handles this with facility-scoped configurations:
- Different guardrail thresholds per facility (academic medical center vs. community clinic)
- Different payer routing per region (different insurance markets, different prior auth requirements)
- Different approval workflows per department (ED auto-approves certain orders; surgery requires committee review)
- Unified observability across all facilities — system-wide dashboards with facility-level drill-down
This is the governance layer that lets a health system deploy Abridge’s agentic capabilities at enterprise scale without facility-by-facility custom code.
Partnership Model
Phase 1: Architecture Alignment (Weeks 1-6)
- Map Abridge’s Contextual Reasoning Engine outputs to WorkingAgents MCP tool schemas
- Define permission scopes for clinical agent roles (documentation, coding, prior auth, communication)
- Build healthcare-specific guardrails: medical advice detection, PHI scope validation, clinical safety checks
- Prototype a governed prior authorization workflow: Abridge encounter → code extraction → Availity submission → EHR update, with WorkingAgents orchestrating permissions and audit trail
Phase 2: Pilot with Partner Health System (Weeks 7-16)
- Deploy with a single Abridge customer (mid-size health system, 500-1,000 clinicians)
- Start with governed prior authorization — the workflow Abridge is already building with Availity
- Measure: authorization turnaround time, clinician approval burden, compliance audit completeness
- Validate role-based access model with real clinical hierarchies (attending, resident, scribe, billing)
Phase 3: Enterprise Governance Product (Weeks 17-24)
- Package “Abridge + WorkingAgents” as a governed clinical AI platform
- Joint compliance documentation: how the combined platform satisfies HIPAA, HITECH, and emerging AI regulations
- Publish reference architectures for governed clinical workflows across specialties
- Expand to additional agentic workflows: automated referrals, medication reconciliation, care gap closure
Why This Partnership Works
Abridge has solved the hardest part of healthcare AI: turning unstructured clinical conversations into structured, accurate, billable documentation at scale. Their 200+ health system deployments, $100M+ ARR, and KLAS #1 ranking prove the product works.
But Abridge is crossing from observation to action. The Contextual Reasoning Engine, Availity partnership, and prior auth automation signal a platform that’s moving from “AI that documents” to “AI that acts.” Prior auth today. Orders tomorrow. Referrals next quarter. Patient outreach next year.
Every one of those actions crosses system boundaries, touches patient data, and carries regulatory risk. The governance challenge scales faster than the capability.
WorkingAgents exists to solve exactly this problem: making AI agents auditable, permissioned, and compliant across every system they touch. Healthcare — where the stakes are patient safety, the regulators are active, and the compliance requirements are the strictest in any industry — is the highest-value vertical for agent governance.
Abridge makes clinical AI intelligent. WorkingAgents makes it governable. The partnership makes it deployable at the scale Abridge is already operating — 200+ health systems, 80 million conversations, 24,600 physicians at Kaiser alone — with the trust infrastructure that autonomous clinical agents demand.
James Aspinwall is the founder of WorkingAgents, an AI governance platform specializing in agent access control, security, and integration services for enterprises deploying AI at scale.