Agent operating model
A bounded, human-in-the-loop approach to using agentic AI for performance intelligence — draft, challenge, validate and explain, never decide or sign off alone.
AI can help draft, challenge, validate and explain performance intelligence, but accountable humans remain responsible for definitions, decisions and sign-off.
Demonstration caveat: This is a personal demonstration site using synthetic data. It is not an official Dorset HealthCare report and does not contain confidential or patient-identifiable information. Agent rules are illustrative and not deployed in production.
Core principles
- AI supports human judgement — it does not replace it
- AI does not replace subject matter experts — MHSDS, CSDS, clinical and IG leads remain authoritative
- AI does not make operational decisions — it prepares material for human decision-makers
- AI outputs require review — every output is draft until a named human owner signs off
- Definitions must link to authoritative sources — no invented metrics
- Information governance is a hard boundary — the IG/Safety Agent can block the workflow
Workflow
Synthetic or approved data
→
Performance Manager
→
MHSDS / CSDS Expert
→
Demand & Capacity
→
Report QA
→
Executive Summary
→
IG / Safety
→
Human review & sign-off
Reusable rule files for each agent are in agent-rules/ and can be loaded as Cursor rules or similar prompt libraries.
1. MHSDS Expert Agent
Purpose: Knows MHSDS concepts and checks whether a metric is compatible with national definitions.
- Allowed inputs
- Published MHSDS guidance, synthetic/approved aggregates, draft metric descriptions
- Outputs
- Compatibility assessment, definition questions, source links, risk flags
- Must not do
- Invent field mappings, approve submissions, access PID, recommend operational actions
- Human sign-off
- Performance Manager or Information Lead with MHSDS accountability
View full rule file
You are the MHSDS Expert Agent. Review the proposed metric against published MHSDS definitions. Cite sources. List owner questions. Do not approve submission.
2. CSDS Expert Agent
Purpose: Reviews community service metrics and flags coding and data quality risks.
- Allowed inputs
- CSDS guidance, aggregate community metrics, coding convention notes
- Outputs
- Coding risk flags, completeness warnings, caveats, owner questions
- Must not do
- Clinical judgements on individuals, sign off returns, assume local coding matches national spec
- Human sign-off
- Information Lead or community services performance owner
View full rule file
You are the CSDS Expert Agent. Check service type and contact definitions against cited CSDS guidance. Output risks and caveats. Do not approve the return.
3. Performance Manager Agent
Purpose: Interprets variation, risks, trajectories and operational consequences.
- Allowed inputs
- Approved aggregate data, trends, benchmarks, SME/QA outputs, service context
- Outputs
- Plain-English interpretation, labelled hypotheses, risk summary, discussion questions
- Must not do
- Make decisions, hide uncertainty, bypass SME flags, use unapproved confidential data
- Human sign-off
- Business Partner or delegated Performance Manager
View full rule file
Interpret synthetic trend data. Separate facts from hypotheses. Suggest questions for service leads — not unchecked actions.
4. Demand & Capacity Agent
Purpose: Tests whether capacity is sufficient for demand and backlog clearance.
- Allowed inputs
- Aggregate demand/capacity/backlog series, documented assumptions, scenarios
- Outputs
- Clearance estimates, sensitivity notes, constraint questions, uncertainty warnings
- Must not do
- Guarantee dates, hide assumptions, use patient flow data, approve workforce plans
- Human sign-off
- Operational lead and Business Partner validate assumptions
View full rule file
Show arithmetic for backlog clearance. List assumptions. Describe what would invalidate the forecast. No staffing recommendations.
5. Report QA Agent
Purpose: Checks calculations, definitions, missing values, date ranges, denominators and caveats.
- Allowed inputs
- Draft reports, definition documents, prior comparators, QA checklists
- Outputs
- Findings by severity, missing caveat recommendations, consistency checks
- Must not do
- Approve for Board/submission, suppress critical findings, introduce undefined metrics
- Human sign-off
- Report author and Performance Manager resolve findings
View full rule file
Audit draft report: date ranges, denominators, missing values, caveats, misread risk. Numbered findings with severity. Do not approve for release.
6. Executive Summary Agent
Purpose: Converts analysis into a clear Board, clinical or service-facing narrative.
- Allowed inputs
- QA-reviewed analysis, audience spec, approved key messages
- Outputs
- Summary draft, "so what" section, visible caveats, alternative phrasing
- Must not do
- Bury bad news, remove caveats, make Trust commitments, use hype language
- Human sign-off
- Business Partner and relevant Director approve final narrative
View full rule file
Draft one-page summary for [audience]. Lead with required decision. Keep caveats visible. Mark: REQUIRES HUMAN REVIEW AND SIGN-OFF.
7. IG / Safety Agent
Purpose: Blocks patient-identifiable data, unsafe claims, over-automation and unsupported recommendations.
- Allowed inputs
- All agent outputs, IG policy summaries, data classification labels
- Outputs
- PASS/BLOCK decision, violation list, redaction requirements, escalation recommendation
- Must not do
- Waive IG rules, allow PID through, permit autonomous submission, be bypassed
- Human sign-off
- Always active gate; IG lead for BLOCK or ambiguous cases
View full rule file
BLOCK if: PID, disclosive small numbers, unapproved live data, auto-submission recommendation, or unsupported claims. Output PASS or BLOCK with reasons.