An investigation is a judgement exercise with legal weight: its findings assign no blame but drive real change, and its report may end up in front of a regulator. That makes AI in investigation software simultaneously valuable and dangerous: valuable because investigators drown in assembly work, dangerous the moment a model's guess becomes a finding nobody consciously made.
The doctrine: AI proposes, an investigator decides
The line that keeps AI safe in this domain is simple to state and must be engineered deliberately: every AI output is a candidate (labelled as AI-generated, carrying its rationale and evidence citations) that a human explicitly accepts, edits or rejects. Nothing auto-applies to the investigation record. In SafetyPulse this is enforced structurally: suggested contributing factors, timeline entries and recommendations arrive in a review drawer where the investigator selects what enters the record; drafted report sections are loaded, edited and approved by a person; accepted AI-assisted wording stays permanently badged as such.
Where AI genuinely earns its keep
- Suggesting candidate contributing factors from the evidence, with citations, for the investigator to test against the ICAM families
- Drafting timeline entries from witness statements and evidence, for confirmation against the record
- Transcribing interviews with speaker separation, so investigators listen instead of scribbling
- Reading handwritten field notes and whiteboards into text
- Assembling the report draft from the structured record: the analysis is the investigator's; the typing isn't
Where it must never decide
Severity and SIF-potential calls, factor classification, the basic cause, and recommendation sign-off are judgement calls the method assigns to accountable people. A platform that auto-classifies factors or auto-writes conclusions into the record has replaced the investigator's judgement with a model's, and no one will be able to say, under scrutiny, who actually made the finding.
The governance questions to ask any vendor
- Can any AI output enter the investigation record without explicit human acceptance?
- Is AI-assisted content labelled, at decision time and permanently?
- Where does inference physically run? (For Australian operators: is it onshore, and can the vendor prove the region per-request?)
- Is customer data used to train models? Get it in writing.
- Are AI features rate-limited and auditable, so usage is governed rather than open-ended?
SafetyPulse's answers are published in our Trust Centre: assistive-only by design, every touchpoint human-gated, inference on AWS Bedrock confined to Australia (Sydney and Melbourne, auditable via CloudTrail), and customer data never used for training. AI should give investigators their hours back. The judgement was never its to take.