How to govern AI agents in a hospital like new hires

A practical governance model for AI in Indian hospitals: scope, permissions, KPIs, escalation and audit — and where Ospia fits.

Most AI pilots in hospitals don’t die in the data lab. They die in finance, audit or operations. A bot books a charge the CFO didn’t approve. A script edits a claim with no trail. An assistant nudges discharge without an escalation path. Nobody signed off the scope; nobody owns the outcome.

The pattern is predictable: AI is treated as a feature, not a governed employee. In a hospital, features don’t have KPIs. Employees do.

The mechanism: why AI efforts stall

• No scope of work: Agents are asked to "help" rather than given a bounded job description with in-scope and out-of-scope tasks.
• Loose permissions: The system can propose and execute in the same breath. There’s no clear separation between suggestion and action.
• No autonomy dial: Everything is either manual or fully automated. Nothing sits in the middle where confidence is high but human oversight still holds the switch.
• Weak audit: If you can’t replay what happened, you can’t defend it in audit — or reverse it cleanly.
• No escalation path: When the model is uncertain, it should know who to ask and by when. Without this, it creates shadow processes.

What to put in place before you buy anything

• Write a one-page job description for each proposed agent: scope, objectives, in-scope actions, out-of-scope actions.
• Define permissions explicitly: what it may propose versus what it may execute, and at what thresholds.
• Set an autonomy dial per task: L0 (observe), L1 (propose), L2 (propose with evidence), L3 (execute within limits), L4 (execute within hospital-written boundaries with full audit). Start low; earn trust.
• Attach KPIs that a CFO/CEO recognises: leading indicators (proposal quality, exception rate) and lagging ones (reversals, approval-to-execution cycle time). Make the learning loop explicit: approvals, edits and rejections tune thresholds.
• Design escalation rules: who gets asked, by what channel, in how much time, and what to do if nobody responds.
• Demand replayability: every run should leave a ledger entry and event trail so you can answer "what happened and why" months later, with a reversal path.

Govern the way you would govern a new hire. Identity, scope, permissions, KPIs, schedule, escalation — and an audit trail.

Where Ospia fits

Ospia HOS is an AI-native Hospital Operating System that ships a workforce of governed digital employees, not just features. The twenty-one agents — Revenue Watch, Insurance Watch, Procurement Watch, Inventory Watch, Pharmacy Expiry Watch, Operations Watch, Front Office, Discharge Agent, Medical Superintendent, Nursing Supervisor, Diagnostics Watch, Quality Watch, Infection Control, NMC Compliance Watch, Compliance Watch, Biomedical Watch, HR Watch, IT Support Watch, Reliability Watch, KPI Watch and Chief Executive Digest — each have a job description, permissions, a schedule, escalation rules, an audit trail and a named human they report to.

• Autonomy as a dial: Hospitals set autonomy from L0 to L4 per task. Even at L4, every run leaves a replayable ledger entry, event trail and a reversal path.
• Safety model: Agents propose; only whitelisted executors act; schema is re-validated and audited. Separation of proposing from doing is designed in.
• Learning loop: Agents track outcomes of proposals — approved, rejected or edited — to tune thresholds over time within hospital-governed limits.
• Event-driven: Agents communicate over a durable event bus by subscribing to domain and agent events rather than calling each other directly. This keeps coupling low and audit high.
• Starting points: Illustrative starting autonomy levels are published per agent so hospitals can begin conservatively and move up with evidence.
• Cross-hospital learning, without sharing data: Only the shape of a configuration — a rule structure or workflow pattern — can travel, and only by opt-in contract. No patient data, no commercial terms, nothing identifiable.
• Seeing it, safely: The executive demonstration runs on synthetic data shaped like your hospital, with zero real patients.

If you already have automation, use the checklist above to put guardrails around it. If you are evaluating AI in core operations, insist on autonomy controls, audit trails and escalation that a CFO can govern. That is the difference between a clever feature and a deployable workforce.