Governance and benefits

Responsible use of agentic AI — including building this demonstration site with a Cursor agent under human direction. Clear benefits, hard controls, and a practical assurance checklist.

Demonstration caveat: This is a personal demonstration site using public aggregate and synthetic demonstration data only. It is not an official Dorset HealthCare report and does not contain confidential or patient-identifiable information.

A balanced view

Used well, agentic AI offers real benefits in performance work; used carelessly, it carries real risks. Responsible use is simply about being honest about both and putting controls in place that keep a human accountable for every output. The lists below set the benefits against the controls that make them safe — neither fearful of the technology nor reckless with it.

Benefits

Faster first drafts of reports and briefing notes — humans edit rather than start from blank
More consistent report structure across services and shared targets
Reusable subject matter rules for the Mental Health Services Data Set (MHSDS), Community Services Data Set (CSDS), quality assurance (QA) and Information Governance (IG) — captured once, version controlled
Clearer caveats and data quality warnings — harder to accidentally omit
Better translation from analysis to action — draft narratives tested against operational questions
Better onboarding for analysts and performance managers joining complex portfolios
Improved assurance before reports go to services or Board
Reduced single-person dependency on tacit knowledge held by one individual
Admin and delivery support — project alignment, backlog sync and activity logging with human approval
More time for human judgement, stakeholder engagement and improvement work

Controls

No patient-identifiable data in AI workflows without explicit IG approval
Public and synthetic data only — this site uses public aggregate NHS data and clearly marked synthetic demonstration data; no confidential or patient-identifiable information
Human approval before use — named owner sign-off on every output
Source-linked definitions — metrics traceable to authoritative guidance
Public sources only unless explicitly approved — no unpublished internal documents in AI workflows
Version-controlled agent rules and prompts — auditable change history
Audit log of agent outputs — what was asked, what was produced, who reviewed
Clear confidence levels — confirmed, likely, conditional, needs owner confirmation
Report QA checklist — mandatory before distribution
Escalation where data quality affects interpretation — do not quietly publish uncertain figures

AI assurance checklist

Before any AI-assisted performance output is shared with services, directors or Board, a human owner should confirm:

Done	Check
	Is the data synthetic, public or approved?
	Is the metric definition linked to a source?
	Has the denominator been checked?
	Are date ranges clear?
	Are caveats visible?
	Has a human owner reviewed the output?
	Could the output be misread by a service or Board audience?
	Is there any risk of patient, staff or service confidentiality breach?
	Are specification citations included and checkable at source?

Remember: Building this site with AI is itself an example of governed use — human direction, synthetic/public data only, and no implication of official Trust status. See the Home page and operating model. The Report Analysis worked example shows this checklist applied to a draft brief — that agent is programmed to escalate IG risks, unsupported claims, missing caveats, and any request to approve publication; it produces draft wording only. If any checklist item raises doubt, escalate to the Information Lead before sign-off.

Keeping accountability with named owners, insisting on accurate and validated information, and being transparent about limitations reflect the public service values of transparency and accountability. Setting clear boundaries and knowing when to escalate are exactly the kind of judgement and leadership the Business & Performance Business Partner role calls for — whether the tool is AI or any other part of the performance toolkit.