Test one operational AI workflow against real work.
A bounded sprint around one real safety, security or resilience workflow. Agree what success looks like, test against operational material and collect the evidence needed for a deployment decision.
Replace assumptions with evidence.
A trial is not a demonstration of what a general-purpose AI model can say. It tests whether a complete operational workflow is suitable, useful and controllable in your environment.
| Before the trial | After the trial |
|---|---|
| We think AI may help | Tested against agreed real examples |
| We do not know how reliable it is | Known limitations and review points |
| Potential benefits are unproven | Measured or evidenced outcomes |
| Implementation is uncertain | Defined sources, controls and requirements |
| We are debating what to do | An implement, refine or stop recommendation |
Choose work worth testing.
The best starting point is not the most ambitious AI idea. It is a workflow with clear operational consequences, defined boundaries and a result that can be assessed — whether used daily or only in an emergency.
Where a trial can start
| Procedure and knowledge access | Retrieve the correct approved procedure, with references to its source. |
| Document evaluation | Review RAMS or other submissions against expected requirements, showing the evidence for findings. |
| Intake triage and routing | Sort incoming reports consistently and surface items requiring attention or escalation. |
| Intelligence and investigation | Assemble information across incident reports, access events and other operational records. |
| Actions and retained learning | Track follow-up and make previous decisions and lessons available when relevant. |
Each trial also needs accessible source material and a clear point of human review or responsibility.
The workflow, not just the model.
SIRV brings together the operating task, approved sources, controls and reviewable outputs. Human judgement remains the decision point where it matters.
INPUT
Operational taskAgreed scope and approved source material
CONTROLLED WORKFLOW
SIRV AIDefined checks, boundaries and exceptions
OUTPUT & REVIEW
Evidence for a person to reviewTraceable findings, decision and follow-up
The output is not merely an answer; it should be possible to review the evidence behind it.
From workflow definition to decision.
Define
Set the question, scope, measures and review responsibility.
Prepare
Agree sources, access, sample cases and controls.
Test
Run representative cases and capture outputs.
Review
Check accuracy, usefulness and exceptions with people.
Decide
Assess evidence and agree the appropriate next step.
A typical working sprint is approximately 4–6 weeks, depending on scope and data readiness. Duration and deliverables are agreed before starting.
What you know at the end.
The objective is a defensible next decision, with evidence that the workflow owner and relevant stakeholders can inspect.
| Question | Evidence produced |
|---|---|
| Does it work? | Results from agreed test cases |
| Is it better? | Comparison with the existing method, where measurable |
| Can we rely on it? | Findings on controls, errors, exceptions and review needs |
| Would people use it? | Practical feedback from reviewers and users |
| What does deployment need? | Sources, integration, governance and ownership requirements |
| Should we proceed? | Documented implement, refine or stop recommendation |
Measures are agreed at the outset. Not every potential benefit can be quantified in a short trial; limitations and unresolved questions are part of the result.
Grounded in real operational work.
These examples illustrate the kinds of defined work that can be tested. The actual trial scope depends on the sources available and the operational decision being supported.
RAMS review
Surface omissions, inconsistencies and points requiring competent-person judgement, with findings linked to the document evidence.
Explore document evaluation ↗Unauthorised access
Reconcile access events and supporting information to build a reviewable sequence for an investigator, rather than automating the final conclusion.
Approved procedures
Test whether staff can retrieve current, relevant instructions more quickly while keeping the source visible for verification.
The trial should end with a decision.
Implement
The evidence supports proceeding to deployment.
Refine
The opportunity is credible but scope, sources or controls need work.
Stop
The evidence does not justify proceeding.
Have a workflow in mind?
An initial conversation can establish whether it is a suitable starting point and what evidence a trial would need to produce.
