AI where it helps.
Human authority where it matters.
The force decides where AI runs, on what, and when data may cross its boundary. Every interaction leaves a record that stands up to scrutiny, and no AI output becomes a decision without a person taking responsibility for it.
The force sets the default. Not the vendor.
Local
Sensitive work processed entirely inside the force environment, on locally hosted models. Nothing leaves the boundary.
Approved cloud
Specifically authorised tasks routed to force-approved models under policy, with a human gate on egress and the request logged.
No-AI
Every core workflow remains fully operational with generative AI switched off entirely, for forces or tasks where that is the policy.
Nothing reaches an output without passing the human gate.
Every AI request funnels through one human-authority gate. From there the force's chosen mode runs, and only the approved-cloud route ever crosses the security boundary. Every step is recorded.
The human-authority gate is the one point every AI task must pass. The force chooses the mode; only the approved-cloud route leaves the security boundary; and the full sequence is written to the case record.
The same workflow, step by step.
Request
A user asks the platform to perform a bounded AI task.
Policy check
The task is checked against force policy for what mode, model and data it may use.
Authority
Restricted tasks require an explicit authorisation before proceeding.
Processing
The task runs locally, or via the approved-cloud route if authorised.
Review
Output is presented to the user with its classification and sources.
Decision
A person confirms, amends or rejects the output.
Audit
The full sequence is written to the tamper-evident audit trail.
Every model is registered, versioned and controlled.
Approved model register
Only models on the force's approved register may be used, whether local or approved-cloud.
Model & prompt versions
The exact model version and prompt template behind every output are recorded, not just the model family.
Agent versions
Where a task runs through a specialist AI agent rather than a single prompt, the agent's version is recorded alongside the models it called.
Test & evaluation
Models and agents are evaluated against defined tasks before release, and re-evaluated when a new version is proposed.
Release controls
A named authority approves a model or agent version before it becomes available for operational use.
Rollback
A previous model or agent version can be reinstated if a new version underperforms or is withdrawn.
Restricted tasks
Certain task types can be excluded from AI assistance entirely by force policy, regardless of mode.
Output traces back to the material it came from.
Where an AI output draws on case material, the platform carries a reference back to the source artefact and a confidence signal, so the reviewer can check the basis for the output rather than take it on faith.
Confidence signalling supports review; it is not a substitute for it. Every output remains subject to human confirmation before it is relied upon.
Four labels, one rule.
Suggestion
A prompt for a person to consider; carries no evidential weight on its own.
Draft
Draft text or structure for a human to review, edit and take ownership of.
Assessment support
Analysis intended to inform a human assessment, not to stand as the assessment itself.
Prohibited: autonomous decision
The platform does not allow AI output to stand as an investigative, evidential or disclosure decision without human authorisation.
Nothing leaves the boundary by default.
Default-closed
Local mode is the default for sensitive work. Egress to an external model requires an explicit policy match, not just a user request.
Scoped requests
Where approved-cloud processing is authorised, only the data required for that specific task is sent, not the wider case.
Logged both ways
The outbound request and the returned response are both logged, so what left the boundary and what came back are both reviewable.
Every AI interaction is recorded.
Model & version
Exactly which model, and which version, produced the output.
User & purpose
Who ran the task and the stated purpose.
Records accessed
What case material the task touched.
Egress
Whether data left the force boundary, and under what authorisation.
Human decision
What a person confirmed, rejected or amended.
Confidence & sources
Confidence signalling and references back to the source artefact.
The declaration attaches to the case record and can be included in a disclosure bundle, so the use of AI in an investigation is answerable in the same way as any other investigative step.
Where AI assists today.
Entity extraction
Surfacing candidate people, organisations, accounts and locations from unstructured material for an analyst to confirm.
Collection-plan drafting
Drafting a first-pass collection plan against the case objectives, for the SIO to revise and approve.
Case summarisation
Producing a draft summary of case developments for a supervisor to review before it is relied upon.
Document classification
Suggesting handling classification and relevance tags on incoming material, subject to confirmation.
Network-analysis assistance
Highlighting candidate links and clusters for an analyst to test, not asserting a finished network as fact.
Disclosure triage support
Flagging material that may be relevant to disclosure for a disclosure officer's review, never deciding relevance on its own.
In every case above, the output is a suggestion, draft or assessment support, reviewed and owned by a person before it informs a decision. AI does not replace the authorised decision-maker.
Walk your governance model through with us.
Bring your AI-governance policy and see how it maps onto local, approved-cloud and no-AI modes.