AI governance · review desk

Every go-live is a decision someone can defend

Writing the policy is the easy half. This desk is the other half: the working review function that decides whether an AI system ships at a health system, run end to end in your browser.

Interactive demonstration. All data is synthetic; no information leaves this page. Meridian Health, every person, vendor, model, and date below are fictional. Nothing here makes a network request, writes storage, or sets a cookie — the logic runs entirely in your tab.

The desk opens on the live queue: every open review at Meridian, ranked by a published priority formula, with the capacity math shown and line-jumping possible only through a recorded override. Four of those cases are then worked end to end, from the highest-risk imaging model to routine marketing copy. Change the intake facts, toggle evidence, flip a reviewer’s vote, and run the gate: the tier, the checklist, and the outcome all follow the same written rules, and the whole decision serializes into one auditable package.

0 · The queue

What gets reviewed first is a formula, not a mood

One case at a time is a demo; a governance function works a queue. These are Meridian’s open reviews as of 2026-06-02 (fictional), ranked by a published formula: tier points (HIGH 40 / MEDIUM 20 / LOW 10) + one point per day waiting (capped at 30) + 25 if the review has blown its service-level window. Ties go to the longer-waiting case, then the lower case id. The same queue state always ranks in the same order — jumping the line takes a recorded override, below.

Open cases
16
HIGH tier
5
Past their window
4
Weekly capacity
8
Weeks to clear
2

Capacity assumption, stated rather than implied: 2 review sessions a week, 4 cases per session. Every figure above is computed from the rows below — none of them is typed into the copy. Service windows by tier: HIGH 10 days, MEDIUM 15, LOW 20.

Open AI reviews ranked by the published priority formula; expedite overrides require a reason, an approver, and an expiry
#CaseTierWaitingPriorityExpedite
1MH-AI-2026-008 · Sepsis early-warning model, annual revalidationInpatientHIGH19dpast 10d window8440+19+25
2MH-AI-2026-014 · Chest imaging worklist prioritizationRadiology · HIGH12dpast 10d window7740+12+25
3MH-AI-2026-018 · Prior-authorization letter generationRevenue CycleMEDIUM16dpast 15d window6120+16+25
4MH-AI-2026-028 · Supply-chain demand forecastingSupply ChainLOW22dpast 20d window5710+22+25
5MH-AI-2026-024 · Patient portal message triageAmbulatoryHIGH7d4740+7
6MH-AI-2026-027 · Agentic scheduling assistantPatient Access · HIGH6d4640+6
7MH-AI-2026-031 · Bed-management optimizerInpatient OperationsHIGH3d4340+3
8MH-AI-2026-032 · Clinician formulary question chatbotPharmacyMEDIUM14d3420+14
9MH-AI-2026-025 · Claims coding suggestion serviceRevenue CycleMEDIUM13d3320+13
10MH-AI-2026-016 · ED discharge-summary draftingEmergencyMEDIUM11d3120+11
11MH-AI-2026-021 · Ambient visit-note draftingAmbulatory · MEDIUM9d2920+9
12MH-AI-2026-022 · Radiology report proofreading assistRadiologyMEDIUM8d2820+8
13MH-AI-2026-030 · Interpreter-service request routingPatient AccessMEDIUM5d2520+5
14MH-AI-2026-029 · Recruiting resume screening assistHuman ResourcesLOW10d2010+10
15MH-AI-2026-033 · Service-line marketing copy draftsMarketing · LOW4d1410+4
16MH-AI-2026-034 · Public website FAQ chatbotDigitalLOW2d1210+2

Queue audit log

No overrides this session. The log records every application and revocation, with its reason, approver, and expiry.

MH-AI-2026-014 · Chest imaging worklist prioritization. A critical finding that waits in the queue costs time a patient may not have — and a wrong ranking buries the sickest patient at the bottom. This model reorders the radiology worklist so likely-critical studies are read first.

1 · Intake

Get intake wrong and every later check inherits the error

The desk starts with facts, not the vendor deck. Four of these fields are live — change the decision role, PHI, write-back, or autonomy and watch the tier, the evidence list, and the gate react.

Intended use
Rank incoming chest radiographs by likelihood of a critical finding so the riskiest studies are read first.
Users
Radiologists and radiology residents at Meridian Health (fictional).
Model / vendor / version
TriageRank CXR · Helix Imaging AI (fictional vendor) · v3.2.0
Data sources
DICOM images and order metadata from the radiology PACS. Contains PHI.
Human oversight point
A radiologist reads every study regardless of rank; the model changes order, never disposition.
Owners
Evidence: Dr. Amara Osei (CMIO office). Monitoring: Dr. Lena Vasquez (Data Science). Rollback: Ruth Alvarez (Radiology Operations).
Submitted
2026-02-03 (fictional date)

2 · Risk tier

Scrutiny scales with what the system can break

The tier is computed by a versioned rule table, not by whoever is in the room. Every rule that fired is shown with its contribution, and identical inputs always produce the identical tier and fingerprint below.

HIGHscore 7 · MEDIUM at 3+, HIGH at 7+ · Ruleset v1.2 — informed by CHAI risk-categorization concepts. No CHAI endorsement is implied; the weights are Meridian's own.

Risk rules that fired for this case, with each contribution
RuleWhat firedPoints
R1Decision role is augment — output shapes a human decision+2
R3Data sources include PHI+2
R7Used in a clinical context — errors can reach patient care+3

Decision fingerprint: fnv1a:587cfd8f — computed from the canonical JSON of the case inputs and ruleset v1.2. Change an input above and it changes; change it back and the exact hex returns.

The full rule table, including rules that did not fire
  • R1 (+2) — Decision role is augment — output shapes a human decision
  • R2 (+4) — Decision role is replace — the system acts, review happens after
  • R3 (+2) — Data sources include PHI
  • R4 (+2) — System writes back into a system of record
  • R5 (+2) — Autonomy: supervised agent — acts before a human sees it
  • R6 (+4) — Autonomy: autonomous — no routine human review of actions
  • R7 (+3) — Used in a clinical context — errors can reach patient care
  • F1 (forces HIGH) — Replaces human decisions in a clinical context — always HIGH
  • F2 (forces HIGH) — Autonomous with write-back — always HIGH

3 · Required evidence

Missing evidence stops a launch — it does not merely slow one

The checklist is derived from the tier and the case flags, so it moves when the intake moves. On a HIGH-tier case the gate fails closed: a required item that is not on file blocks go-live outright. Toggle items to watch the gate react.

  • E1 · Local validation on Meridian dataRequired — tier is HIGH

    Vendor numbers were earned on someone else's population. Performance is re-measured here before anyone relies on it.

  • E2 · Subgroup performance reportRequired — clinical use at HIGH tier

    An average that looks fine can hide a group the model fails. Performance is broken out by the subgroups the use case touches.

  • E3 · Privacy and security reviewRequired — data sources include PHI

    Data flows, access controls, and retention reviewed against the PHI handling obligations in the vendor agreement, under HIPAA where PHI is present.

  • E4 · Model card on fileRequired — every case, every tier

    What the model is, what it was trained on, and where it is known to fail — written down before go-live, not reconstructed after an incident.

  • E5 · Monitoring plan with a named ownerRequired — tier is HIGH

    A dashboard nobody owns is decoration. The plan names the person who looks, what they look at, and how often.

  • E6 · Rollback plan with a named ownerRequired — tier is HIGH

    How the system is switched off, who can do it, and what the workflow reverts to — tested, not theoretical.

  • E7 · Regulatory screenRequired — clinical use requires the screen

    A documented determination of whether the use falls under medical-device or other regulatory oversight, with the reasoning attached.

4 · Cross-functional review

One reviewer can stop the launch — that is the design

Six roles record a position and a written reason. A Clinical or Security block can never be outvoted: the aggregate is blocked no matter what the other five say. Change any vote and re-run the gate below.

Clinicalblock cannot be outvoted

Dr. Amara Osei, CMIO office (fictional)

Condition: quarterly subgroup review on portable and pediatric films, and ranking auto-disables if weekly case mix shifts beyond the agreed bounds.

Data Science

Dr. Lena Vasquez (fictional)

Local validation on 41,200 Meridian studies holds up against the vendor claim. Monitoring thresholds are set and owned.

Privacy

Priya Nair (fictional)

PHI stays inside the PACS boundary; no data leaves Meridian infrastructure. Retention unchanged from current radiology policy.

Securityblock cannot be outvoted

Marcus Webb, CISO office (fictional)

Read-only integration, no new external egress, vendor access reviewed. No objection.

Legal / Compliance

Daniel Cho (fictional)

Vendor holds the regulatory clearance for this indication; our use matches the cleared labeling.

Operations

Ruth Alvarez (fictional)

Rollback is a worklist-config flag, tested in staging on 2026-01-28 (fictional date). Reading-room workflow unchanged.

5 · Go-live gate

The gate records the whole decision or nothing at all

One button, one atomic outcome. Either the complete decision lands — state, conditions, owners, expiry, and an audit event — or nothing is written and the gate names exactly which rule or item stopped it. The rules, in the order they are checked:

  1. G1. A Clinical or Security block ends the review. It cannot be outvoted, averaged, or escalated around.
  2. G2. A block from any other role also blocks. Blocks are resolved by fixing the objection, not by counting votes.
  3. G3. A HIGH-tier case with any required evidence missing is blocked. The gate does not open on a promise to supply it later.
  4. G4. Missing required evidence below HIGH tier, or any reviewer requesting evidence, returns the case for evidence.
  5. G5. Any approval carrying conditions makes the whole decision conditional, with the conditions recorded on it.
  6. G6. Only when none of the above applies is the case approved — with owners and a reassessment date attached.

Audit log — the trail a regulator reads first

  1. 01 · 2026-02-03 (fictional) Intake received from Radiology service line.
  2. 02 · 2026-02-04 (fictional) Risk tier computed under ruleset v1.2.
  3. 03 · 2026-02-10 (fictional) Cross-functional review convened; all six roles recorded a position.
  4. 04 · 2026-02-11 (fictional) Subgroup performance report (E2) requested from vendor; not yet on file.

6 · Audit package

The decision has to outlive the meeting

Everything above serializes into one canonical JSON package — the artifact you hand an auditor eighteen months later. Keys are sorted, dates are fixed, and the same case state always produces the same bytes.

{
  "auditTrail": [
    {
      "date": "2026-02-03 (fictional)",
      "event": "Intake received from Radiology service line."
    },
    {
      "date": "2026-02-04 (fictional)",
      "event": "Risk tier computed under ruleset v1.2."
    },
    {
      "date": "2026-02-10 (fictional)",
      "event": "Cross-functional review convened; all six roles recorded a position."
    },
    {
      "date": "2026-02-11 (fictional)",
      "event": "Subgroup performance report (E2) requested from vendor; not yet on file."
    }
  ],
  "caseId": "MH-AI-2026-014",
  "evidenceManifest": [
    {
      "because": "tier is HIGH",
      "id": "E1",
      "label": "Local validation on Meridian data",
      "onFile": true,
      "status": "required"
    },
    {
      "because": "clinical use at HIGH tier",
      "id": "E2",
      "label": "Subgroup performance report",
      "onFile": false,
      "status": "required"
    },
    {
      "because": "data sources include PHI",
      "id": "E3",
      "label": "Privacy and security review",
      "onFile": true,
      "status": "required"
    },
    {
      "because": "every case, every tier",
      "id": "E4",
      "label": "Model card on file",
      "onFile": true,
      "status": "required"
    },
    {
      "because": "tier is HIGH",
      "id": "E5",
      "label": "Monitoring plan with a named owner",
      "onFile": true,
      "status": "required"
    },
    {
      "because": "tier is HIGH",
      "id": "E6",
      "label": "Rollback plan with a named owner",
      "onFile": true,
      "status": "required"
    },
    {
      "because": "clinical use requires the screen",
      "id": "E7",
      "label": "Regulatory screen",
      "onFile": true,
      "status": "required"
    }
  ],
  "gate": {
    "outcome": "not yet run",
    "recorded": false
  },
  "intake": {
    "autonomy": "assistive",
    "clinicalContext": true,
    "dataSources": "DICOM images and order metadata from the radiology PACS. Contains PHI.",
    "decisionRole": "augment",
    "intendedUse": "Rank incoming chest radiographs by likelihood of a critical finding so the riskiest studies are read first.",
    "model": "TriageRank CXR",
    "oversight": "A radiologist reads every study regardless of rank; the model changes order, never disposition.",
    "owners": {
      "evidence": "Dr. Amara Osei (CMIO office)",
      "monitoring": "Dr. Lena Vasquez (Data Science)",
      "rollback": "Ruth Alvarez (Radiology Operations)"
    },
    "phi": "yes",
    "submitted": "2026-02-03 (fictional date)",
    "system": "Chest imaging worklist prioritization",
    "users": "Radiologists and radiology residents at Meridian Health (fictional).",
    "vendor": "Helix Imaging AI (fictional vendor)",
    "version": "3.2.0",
    "writeBack": "no"
  },
  "reviews": [
    {
      "decision": "Approve with conditions",
      "rationale": "Condition: quarterly subgroup review on portable and pediatric films, and ranking auto-disables if weekly case mix shifts beyond the agreed bounds.",
      "reviewer": "Dr. Amara Osei, CMIO office",
      "role": "Clinical"
    },
    {
      "decision": "Approve",
      "rationale": "Local validation on 41,200 Meridian studies holds up against the vendor claim. Monitoring thresholds are set and owned.",
      "reviewer": "Dr. Lena Vasquez",
      "role": "Data Science"
    },
    {
      "decision": "Approve",
      "rationale": "PHI stays inside the PACS boundary; no data leaves Meridian infrastructure. Retention unchanged from current radiology policy.",
      "reviewer": "Priya Nair",
      "role": "Privacy"
    },
    {
      "decision": "Approve",
      "rationale": "Read-only integration, no new external egress, vendor access reviewed. No objection.",
      "reviewer": "Marcus Webb, CISO office",
      "role": "Security"
    },
    {
      "decision": "Approve",
      "rationale": "Vendor holds the regulatory clearance for this indication; our use matches the cleared labeling.",
      "reviewer": "Daniel Cho",
      "role": "Legal / Compliance"
    },
    {
      "decision": "Approve",
      "rationale": "Rollback is a worklist-config flag, tested in staging on 2026-01-28 (fictional date). Reading-room workflow unchanged.",
      "reviewer": "Ruth Alvarez",
      "role": "Operations"
    }
  ],
  "risk": {
    "decisionHash": "587cfd8f",
    "floors": [],
    "rulesFired": [
      {
        "id": "R1",
        "label": "Decision role is augment — output shapes a human decision",
        "points": 2
      },
      {
        "id": "R3",
        "label": "Data sources include PHI",
        "points": 2
      },
      {
        "id": "R7",
        "label": "Used in a clinical context — errors can reach patient care",
        "points": 3
      }
    ],
    "ruleset": "v1.2",
    "score": 7,
    "tier": "HIGH"
  },
  "synthetic": true,
  "system": "Meridian Health (fictional) — AI governance review desk demonstration"
}

Where each field lands when someone asks about frameworks

Crosswalk from decision-package fields to NIST AI RMF functions and ISO/IEC 42001 clause themes
Package fieldNIST AI RMF functionISO/IEC 42001 clause theme
intakeMapUnderstanding the system, its context, and impacts
risk.rulesFired, risk.tierMap / MeasureAI risk assessment
evidenceManifestMeasurePerformance evaluation and documented information
reviewsGovernRoles, responsibilities, and accountability
gateManageOperational planning and control of AI systems
risk.decisionHash, auditTrailGovernTraceability and record-keeping

The ISO/IEC 42001 column names clause themes, not clause numbers — quoting numbers this page cannot cite would be exactly the fabricated precision this desk exists to catch.