Data Academy · Tutorial 3 of 10 · Healthcare providers

Governed AI in healthcare providers

Hospitals and health systems are moving AI out of note-taking pilots and into the revenue cycle, patient access and patient flow, where agents now assemble prior authorisations, work denials and prepare discharges. Whether this scales depends on governance: clinical decisions stay with clinicians, protected health information is exposed only to the minimum necessary, and every action an agent takes in the EHR or billing system can be traced and explained.

01 What is changing

Where AI in healthcare providers is heading

From ambient documentation and chat assistants to agents that complete multistep administrative work across the EHR, payer portals and the billing system.
From scattered point solutions bought department by department to a shared data and governance layer that many agents reuse.
From pilots judged on enthusiasm to a small number of revenue cycle and patient access domains judged on hard returns: cleaner claims, fewer denials, faster cash.
From payer phone calls and faxes to standards-based electronic prior authorisation (FHIR and Da Vinci CRD, DTR and PAS), which gives agents a structured channel to act through.

02 Use cases

Three use cases on the governed path

Each use case runs the same path: a business question, governed context, a deterministic rule, specialist agents, a policy check, an action and a record. How the path works →

Illustrative: names and figures are invented to show the flow.

Use case 1

Prior authorisation assembly and submission

Imaging and procedures are delayed or cancelled while staff chase payer rules and clinical evidence by hand. An agent that checks coverage, gathers documentation and submits a complete request protects the schedule and reduces avoidable denials.

“Which scheduled procedures next week still need a prior authorisation, and can any be submitted today without clinician input?”Asked by a patient access manager

Context: EHR: MRI lumbar spine booked for patient P-55120 in 6 days at the Riverbend Medical Centre
Payer rules (Northvale Health Plan): authorisation required; at least 6 weeks of documented conservative therapy
EHR notes: 8 weeks of physiotherapy and analgesia documented; neurological exam recorded at last visit
Payer standard decision time 3 business days, so the request must go today to clear before the appointment
Rule: If the payer requires authorisation and every documentation element in its published criteria is present in the record, submit electronically; if any element is missing, route to the ordering clinician.
Decision: Submit prior authorisation Severity: Low
Agents: Coverage agent Reads the payer's coverage requirements for the procedure code and lists the documentation elements they call for.
Evidence agent Finds each element in the EHR (therapy notes, exam findings, dates) and cites the exact source note for each.
Packet agent Assembles the request with only the minimum necessary clinical information and drafts the cover summary.
Policy: Agents may submit only when every criterion is evidenced; they never alter clinical notes or the order. Disclosure follows the HIPAA minimum necessary standard, and any request lacking evidence goes to the ordering clinician.
Action: Prior authorisation request submitted electronically to the payer (Da Vinci PAS / X12 278) and released automatically within policy; status tracked against the appointment in the EHR work queue.
Data: EHR orders and schedulingClinical notes and problem listPayer coverage rulesPatient insurance and eligibilityAuthorisation status history

Use case 2

Discharge readiness and bed flow

When discharges happen late in the day, emergency patients wait for beds and elective lists are cancelled. Spotting patients who are clinically ready earlier, and clearing the non-clinical blockers, frees beds sooner without pressuring clinical judgement.

“Which patients on Ward 7B could go home before noon, and what is stopping them?”Asked by a hospital site and bed manager

Context: Bed management system: Ward 7B has 28 beds, 26 occupied, 9 patients boarding in the emergency department awaiting admission
EHR: 4 patients on 7B carry a medically optimised flag from the morning ward round
Of those 4, 2 await take-home medication from pharmacy and 1 awaits a care-home placement confirmation
Patient transport has 2 slots free before 11:00
Rule: Flag a patient as discharge-ready for clinician review only if a medically optimised flag is documented, no results are pending and medicines reconciliation is complete; no agent may discharge or create a discharge order.
Decision: Expedite discharge Severity: High
Agents: Flow agent Matches ready patients to waiting admissions and shows which beds would free first.
Blocker agent Identifies the non-clinical blocker for each patient (pharmacy, transport, placement) and the owner of each.
Handover agent Drafts a short readiness summary for the consultant, citing the ward round note and pending items.
Policy: Discharge is a clinical decision made by the responsible clinician. Medication-safety checks override flow targets, and agents read only the fields needed for readiness, not the full record.
Action: Discharge-readiness task referred to the attending consultant in the EHR; pharmacy and transport requests prepared but not sent; a provisional bed allocation held in the bed management system until the discharge order is signed.
Data: EHR inpatient record and ward round notesBed management and ADT eventsPharmacy dispensing queuePatient transport bookingsSocial care placement status

Use case 3

Denial management and appeals

Denied claims are often left to age because working them is slow and manual. Sorting denials by fixability and drafting evidence-backed appeals recovers revenue that would otherwise be written off.

“Which of last week's denials from Harbourline Mutual are worth appealing, and which can simply be corrected and resubmitted?”Asked by a revenue cycle director

Context: Billing system: 42 claims denied by Harbourline Mutual last week, $186,400 in total
30 denials for missing or invalid information, $65,400 in total, each with the missing field identifiable
12 denials for medical necessity, $121,000 in total, each above $5,000
EHR: supporting documentation found for 10 of the 12 medical-necessity denials
Rule: Missing-information denials with an identifiable fix are corrected and resubmitted; medical-necessity denials above $5,000 with supporting documentation are appealed; any write-off requires approval.
Decision: Appeal medical-necessity denials Severity: High
Agents: Denial triage agent Groups denials by reason code and payer and separates fixable errors from clinical disputes.
Appeal drafting agent Drafts each appeal letter citing the payer's own policy and the dated clinical evidence.
Root cause agent Traces repeat denial reasons back to registration, coding or documentation steps so they can be fixed upstream.
Policy: Coders and billing staff may approve corrected claims; appeals and any adjustment or write-off need the revenue integrity lead's approval. Coding changes must be supported by the clinical record.
Action: 10 appeal packets created in the denials work queue pending approval by the revenue integrity lead; 2 cases without documentation referred to clinical documentation improvement.
Data: Billing and claims (837/835 remittance)Denial reason codesEHR clinical documentationPayer contracts and medical policiesCoding and charge master

03 The foundation

What the agents need to understand

Core entities in the ontology

PatientEncounterOrderBedClinicianPayerCoverageAuthorisationClaimDenial

Systems they come from

EHR: patients, encounters, orders, notes, medications and results
ADT and bed management: admissions, transfers, discharges and live bed status
Revenue cycle / patient accounting: charges, claims, remittances, denials and balances
Payer gateway and eligibility: coverage, authorisation requests and responses
Pharmacy information system: dispensing, medicines reconciliation and stock
Scheduling and patient access: appointments, referrals and pre-registration

04 Guardrails

The controls that let it scale

Clinician owns clinical decisions

Agents prepare and explain; diagnosis, treatment and discharge orders are signed by a licensed clinician.

Minimum necessary access

Each agent sees only the protected health information its task requires, consistent with HIPAA and GDPR where it applies.

Safety overrides efficiency

When a flow or revenue goal conflicts with a medication or clinical safety check, the safety check wins.

Every action traceable

Each submission, draft and task records the rule, evidence and approver, so it can be reviewed by compliance and payers.

Named owner per agent

Every agent has an accountable owner, defined permissions and a retirement date, so none run unattended.

05 Rollout

From the first use case to many

1
Pick one revenue cycle domain
Start with prior authorisation or denials, where rules are explicit and value shows up in claims and cash.
2
Model the business context
Define patients, encounters, payers, coverage and claims once, with field-level sensitivity tags.
3
Codify rules with the owners
Write the payer, coding and discharge rules with the people accountable for them, and test against past cases.
4
Run with approvals on
Begin with every action pending approval, then release low-risk actions automatically once accuracy is shown.
5
Reuse for the next domain
Carry the same ontology, guardrails and audit trail into patient flow and scheduling rather than buying another point tool.

06 What to measure

Outcomes, not activity

Initial denial ratePrior authorisation turnaround timeDays in accounts receivableDischarges before noonEmergency department boarding timeCost to collect

07 Pitfalls

What usually goes wrong

Buying a tool per problem. Departments buy separate AI products that each need their own data feeds and controls; a shared data and governance layer avoids rework.
Automating a broken process. An agent that speeds up bad registration or documentation just produces denials faster; fix the upstream step the root cause points to.
Letting agents drift into clinical judgement. Discharge and coding agents can start to look like decision makers; keep the clinician sign-off explicit in the workflow.
Proving value once. A pilot that showed results is not the same as value captured every month; measure the same outcomes after go-live.

08 Diagnostics

Questions to ask your team

1
Which two or three domains have rules clear enough and returns large enough to scale first?
2
Where does each agent read protected health information, and is it limited to the minimum necessary?
3
Who approves an appeal, a write-off or a discharge, and is that recorded with the action?
4
Are our payer and clinical rules written down and versioned, or held in people's heads?

09 Keep going

Frequently asked

Will agents make clinical decisions?

No. Agents gather evidence, draft and route; clinical decisions such as discharge, diagnosis and treatment are made and signed by clinicians. The rule layer prevents agents from creating clinical orders.

Why start with the revenue cycle?

Its work is repetitive and rules-based, the outcomes are measurable in claims and cash, and errors are reversible. That makes it a safer place to prove the governed path before patient flow.

How is patient data protected?

Each agent reads only the fields its task needs, every access is logged, and disclosures to payers follow the minimum necessary standard. Sensitive actions wait for a named approver.