Agent-assisted brief
NHS Oversight Framework
First-draft performance brief from public NHS Oversight Framework data. This uses published peer median and rank information. The figures are not recalculated.
View worked exampleDraft reports
Six worked examples of agent-assisted analytical briefs using public aggregate NHS data for Dorset HealthCare.
Each brief is a first draft for human review. They are not official Dorset HealthCare reports and should not be used for operational decision-making.
The purpose of this page is to show how agentic AI could support business and performance work when it is used with clear sources, reproducible scripts, explicit caveats and human sign-off.
Demonstration caveat: These reports use public aggregate data only. They are not official Dorset HealthCare reports. They have not been operationally validated. Human review and local owner confirmation would be required before any operational use.
What this page demonstrates
A Business & Performance Business Partner needs to turn data into clear, useful performance intelligence.
That means more than producing numbers. It means explaining:
- what the figure is
- what it should be, where a target or comparator exists
- whether it is improving, worsening or stable
- what caveats affect the interpretation
- what questions need human review
- what action or escalation may be needed
These examples show how an AI agent could help produce a structured first draft, while keeping the human responsible for validation, judgement and sign-off.
Public-data agent workflow examples
Each brief follows the same structure:
- Question — what the brief is trying to answer
- Data used — which public source was used
- What this report can and cannot tell us — scope and limits
- Headline reading — plain-English takeaways before the detail
- Key findings by review area — grouped narrative for local owners
- Main metric table — figures, comparators and initial readings
- Draft interpretation — first-draft narrative for review
- Human validation checklist — what a person must confirm
- Audit trail and source checks — prompt excerpt, CSV paths and technical trace
This creates a clear audit trail from public data to draft narrative.
Worked examples
Agent-assisted brief
MHSDS mental health access
Access and activity measures using public Mental Health Services Data Set data. This includes monthly trends where the public time series supports them.
View worked exampleAgent-assisted brief
CSDS community services
Community services activity briefing using public Community Services Data Set data. This shows how historic public data can support trend commentary.
View worked exampleAgent-assisted brief
Talking Therapies
Referrals, access and waiting measures from public NHS Talking Therapies data. This avoids recovery or outcome claims where the public extract does not support them.
View worked exampleAgent-assisted brief
Assurance and statutory returns
A source map for selected assurance and statutory reporting areas. This includes KO41a, ERIC, DSPT, FFT gaps and CQC context. It is not a scorecard.
View worked exampleAgent-assisted brief
Urgent care and diagnostics check
A source-validation brief covering RDY presence in A&E, DM01 and KH03 public data. This is deliberately cautious and focuses on what the public sources can and cannot support.
View worked exampleHow these briefs were created
The briefs were created using a human-directed agent workflow.
Public NHS sources were downloaded and catalogued in DATA_SOURCE_REGISTER.csv.
R scripts filtered the relevant Dorset HealthCare rows and created processed public-data extracts.
The report rendering script then produced each brief using the same structure: question, data used, key findings, agent summary, human checks and verification notes.
This means the reports are not just manually written text. They are part of a repeatable workflow that links source data, processing and narrative.
Every brief ends with caveats and a human review gate.
Supporting evidence: regenerate briefs with Rscript site/R/03_render_public_reports.R after updating processed extracts in public-data/processed/.
Why this matters
Performance reports are most useful when they are clear, accurate and honest about uncertainty.
Agentic AI can help with the first draft: organising the data, structuring the findings, drafting the narrative and highlighting caveats.
But the value comes from the governed workflow around it.
In a live NHS setting, an accountable person would still need to confirm the data source, check the definitions, validate the figures, agree the interpretation and decide what action is needed.
This page shows how AI could support that process without replacing professional judgement.