AgentOS — governed AI workforce

AI that works within healthcare, not around it.

AgentOS is a governed AI workforce with six capability tiers, twelve named workers, review gates where they matter, and measurable outcomes. Every action traces back to evidence. Every boundary is explicit.

6 capability tiers12 named workersProvenance by defaultHuman-in-the-loop
Capability tiers

Six tiers. Clear rules at every level.

Not all AI is equal. AgentOS classifies every capability into one of six tiers, and each tier comes with explicit, enforced rules about autonomy and review.

Tier 0
Utility & productivity
Spell-check, template suggestions, readability hints, translation hints
No PHI leaves regulated boundary. Safe to use without review.
Tier I
Retrieval & summarisation
Chart prep, case summaries, document extraction
Human can use output directly. Provenance mandatory.
Tier II
Drafting & extraction
Notes, outreach drafts, coding hints, intake extraction
Human approval required before record mutation or submission.
Tier III
Prioritisation & routing
Queue priority, staffing, inbox triage, task routing
Autonomy within explicit tenant policy. Operator can override.
Tier IV
Clinically influential
Risk scores, deterioration signals, triage recommendations
Clinical review, safety case and evaluation gates mandatory.
Tier V
Potentially regulated
Diagnostic or treatment suggestions
Separate regulatory workstream. Not released by default.
Named AI workers

Twelve agents. Named jobs. Explicit boundaries.

AgentOS ships with twelve pre-defined agents. Each has a scoped job, a Care Graph grounding and a hard boundary it cannot cross — regardless of prompt.

01
Referral Intake Agent

Reads inbound referrals across forms, email, fax and APIs. Extracts facts, flags missing items and suggests triage.

Cannot accept or reject referrals without human approval.
02
Patient Concierge Agent

Handles booking, reminders, check-ins and low-risk administrative tasks across patient channels.

Must hand off on uncertainty or safety triggers.
03
Documentation Agent

Drafts clinical notes, visit summaries, discharge packs and letters grounded in the Care Graph.

Cannot silently write to the legal clinical record.
04
Care Coordination Agent

Watches pathways for overdue tasks, breach risks and blocked work. Nudges owners and escalates.

Cannot close high-risk items without confirmation.
05
Medication Adherence Agent

Monitors refill gaps, missed doses and outreach effectiveness across medication cohorts.

Cannot change treatment plans or suppress incidents.
06
Prior-Auth Agent

Assembles supporting documents, maps payer checklists and drafts rationale for prior authorisation.

Final submission remains human-accountable.
07
Quality Auditor Agent

Reviews completeness, protocol deviations and audit gaps across episodes and documentation.

May open review items. Cannot alter evidence.
08
Coding & Revenue Agent

Suggests codes, checks completeness and surfaces leakage flags across billable activity.

Assistive only. No autonomous submission.
09
Inbox Triage Agent

Clusters, prioritises and routes tasks, messages and alerts across operator inboxes.

Operator must be able to inspect reasoning and override.
10
Executive Analyst Agent

Generates weekly performance narratives, anomaly explanations and cohort comparisons.

Narrative must trace to metric sources in the Care Graph.
11
Assessment Agent

Drafts risk assessments, intake assessments and review summaries from structured observations and history.

Cannot finalise assessments or change risk classifications without clinician sign-off.
12
Handover Agent

Compiles shift handovers, transfer-of-care notes and between-team summaries across open episodes and alerts.

Cannot close outstanding actions or suppress unresolved alerts.
What’s different

Six things other AI platforms don’t do.

AgentOS is not a chat widget bolted onto an EHR. It is a first-class AI runtime inside careos — graph-grounded, evidence-first, governed end-to-end.

Care Graph-grounded reasoning

AI operates over a structured graph of patients, tasks, medications and messages — not free-text RAG over PDFs. Every answer traces to graph nodes.

Agent Studio

Admins configure agent goals, approval rules, prompts and escalation paths without writing code. Policy is owned by the operations team, not engineering.

Simulation & replay

Test policy changes against historical events before release. See exactly how a new routing rule or approval threshold would have behaved last quarter.

Evidence by default

Every AI action stores provenance — model, prompt, inputs, sources, reviewer and outcome. Audit is free, not an afterthought.

Outcome-linked learning

Agents are measured on closed-loop outcomes: contact success, overdue reduction, time saved, override rates. Not vanity metrics.

Multi-channel patient AI

One governed layer across portal, SMS, email, voice and app. Consistent policy, consistent provenance, consistent safety posture.

AI governance

Nine non-negotiable governance requirements.

Provenance, minimisation, registries, evaluations, review, classification, tenant overrides and customer-facing analytics. Each is specified and enforced.

AI-001
Provenance mandatory

Model, provider, version, prompt, inputs, timestamp, actor and decision path stored for every AI action.

AI-002
PHI minimisation

Redaction, routing and retention controls applied before any model invocation, per tenant and market.

AI-003
Registries & flags

Model registry, prompt registry, feature flags, market restrictions and tenant controls managed centrally.

AI-004
Evaluation harnesses

Quality, hallucination rate, safety, override rate, drift and latency tracked against golden sets.

AI-005
Review & rollback

Review queues, rollback paths, feedback capture, incident handling and bad-output suppression.

AI-006
Risk classification

AI use cases classified by regulatory and safety risk before enablement. Tier gates enforced.

AI-007
Tenant policy overrides

Tenants can tighten autonomy, review thresholds and permitted providers below the platform default.

AI-008
Customer AI analytics

Time saved, approval rates, override rates and outcome deltas exposed to customers, not hidden.

AI-009
Cost tracking & attribution

Per-tenant, per-agent, per-use-case metering of AI spend, throughput and incident rates — tied to commercial metering and review.

Hard safety boundaries

Four rules AgentOS will never break.

These are enforced at the platform level, below tenant policy. No prompt, no configuration, no override can cross them.

Non-negotiablePlatform-level safety
  • 01No AI autonomously prescribes, discontinues, or alters medication.
  • 02No AI silently writes to the legal clinical record.
  • 03Clinically influential outputs require human review before action.
  • 04All submissions to external parties remain human-accountable.
Brief your team

See AgentOS working on real care workflows.

We’ll walk through the tiers, workers, governance controls and evidence exports on your own service lines — not a generic demo dataset.