Executive summary
Launch an RSLPF Evidence Intelligence Pilot.
Police investigations are information problems before they are artificial-intelligence problems. A single case may require officers to reconcile interview notes, incident reports, CCTV references, receipts, phone records, property lists, forensic returns, vehicle descriptions and unanswered evidence requests. When those records are fragmented, investigative time is consumed reconstructing the file rather than testing the facts.
Saint Lucia is not starting from zero. A customised police records-management system is already reported as active through the regional CariSECURE work; officers have received recent training in intelligence analysis, crime-scene work, cybercrime and cryptocurrency investigations; and the justice system is adding digital case-management capacity. The right next move is therefore not a free-standing chatbot. It is a controlled support layer connected to approved records and built around the existing legal case file.
The first useful functions are deliberately modest: transcribe approved material, build a source-linked chronology, extract people, places, vehicles and objects, flag factual conflicts for human review, show missing documents, search authorised prior reports and draft routine summaries with every proposition linked back to its source. The system may suggest questions. It must never decide guilt, manufacture evidence, hide uncertainty or turn an unverified association into a fact.
SLPA proposes a staged Evidence Intelligence Pilot: test the system first on closed, quality-reviewed property-crime files; measure accuracy and time saved against a human baseline; move to live shadow mode only after legal, security and disclosure checks; and permit operational use only when an accountable investigator verifies every output. Predictive person-scoring, automated arrest recommendations, emotion recognition and uncontrolled facial recognition should sit outside the pilot.
Developed from an SLPA field observation about the information burden inside a routine Saint Lucia investigation. The design was checked against Saint Lucia’s digital-policing foundations, justice reforms, data-protection law and the INTERPOL–UNICRI responsible-AI framework in July 2026.
Key findings
Better case records are the foundation for useful AI.
Hiring and training matter, but investigators also lose time when facts must be repeatedly copied, searched and reconciled across paper, messages and separate digital systems. Standard records, consistent identifiers and a complete audit trail are prerequisites for useful AI.
Evidence-led findingStart with timelines, extraction and document comparison.
Chronologies, entity extraction, document comparison, request tracking and source-linked drafting are easier to verify than forecasts about who will offend. They can reduce administrative burden without converting probabilistic output into coercive state action.
Evidence-led findingEvery AI output needs a link to its source.
Every generated sentence, conflict flag and suggested lead should open the exact report passage, image, transcript segment or record that produced it. Unsupported output must be visibly labelled and excluded from the evidential record until independently established.
Evidence-led findingBuild disclosure and chain-of-custody rules into the system.
The system should preserve originals, access history, model and prompt versions, officer corrections, export history and the boundary between evidence and analysis. Prosecutors, courts and defence counsel need an intelligible route to test how a material output was produced.
Evidence-led findingSmall-state scale makes privacy controls more important.
Saint Lucia can run a bounded pilot, train a small user group and learn quickly. But a small population makes re-identification and informal access especially consequential. Role-based access, purpose limits, retention rules, breach response and independent audit are core operating controls.
Evidence-led findingCurrent public-safety signal
RSLPF reported declines in five major offence categories.
% reduction reportedPercentage change for 1 January–10 July 2026 against the corresponding 2025 period. These administrative figures describe reported offence counts; they do not by themselves measure investigative quality, detection, case attrition or the effect of any technology.
SLPA policy proposal
RSLPF Evidence Intelligence Pilot
Build a secure, human-accountable evidence assistant that makes case material easier to organise, search, test and disclose—without delegating suspicion, guilt or coercive decisions to a model.
One secure evidence system
Operate inside an approved environment connected to the police records system, with role-based access, encryption, retention rules, immutable originals and no use of public consumer AI services for identifiable case material.
Source-linked case summaries
Generate timelines, entity tables, contradiction matrices, missing-evidence lists and draft summaries only when every output can be traced to an authorised source and its confidence is visible.
Human approval for every decision
Keep an accountable officer responsible for verification and every investigative decision. Ban automated guilt scores, arrest recommendations, witness-credibility ratings and unsupported person-risk rankings.
Full audit access across the justice system
Preserve model versions, queries, outputs, corrections and exports so prosecutors, courts, defence counsel, auditors and authorised oversight bodies can examine material uses.
Test results before wider use
Benchmark closed files, then run a live shadow pilot in a bounded case class. Expand only if accuracy, security, timeliness, workload, fairness and user discipline meet published thresholds.
Delivery sequence
Prepare the records and benchmark the tool
- Map the current police-record, evidence, forensic, disclosure and court handoffs; identify the authoritative system and every uncontrolled copy.
- Complete legal, privacy, cyber-security, records, procurement and human-rights impact assessments using the INTERPOL–UNICRI toolkit.
- Create a de-identified benchmark of quality-reviewed closed property-crime files and score timeline, extraction, contradiction and citation accuracy against trained investigators.
Run a live shadow test with independent review
- Select a small trained unit and a bounded lower-risk case class; keep model output outside the official file until an investigator verifies and adopts it.
- Test source linking, request tracking, disclosure export, chain-of-custody logs, access alerts, failure reporting and supervisor review under real operating pressure.
- Publish aggregate pilot results, material error types, security incidents and the decision to stop, redesign or advance—without exposing victims, witnesses or investigative methods.
Expand the functions that pass every test
- Allow only the functions that passed accuracy, legality, security and workflow tests; retrain users before adding a new evidence type or case class.
- Create standing judicial, prosecutorial, defence, privacy, technical and community review routes for material system changes.
- Join verified caseflow measures to the wider justice dashboard so speed, disclosure quality, fairness and case outcomes are governed together.
Public accountability
Measures for public accountability
Recommended publication: quarterly operating signals and one independently reviewed annual outcome report.Tests whether the system returns investigator time without rewarding rushed conclusions.
Makes accuracy and human correction visible by function and evidence type.
Tests whether every consequential output remains inspectable across the justice chain.
Treats privacy and evidence integrity as operating outcomes, not policy language.
Checks whether administrative savings translate into stronger, fairer case movement.
Limits, uncertainty & sources
Limits of this analysis
- The reported 2026 offence reductions are RSLPF administrative comparisons and do not establish causation, detection rates or investigative performance.
- AI can amplify incomplete reports, historic bias and data-entry errors. Human review does not make a weak system safe unless reviewers have time, training, authority and recorded accountability.
- This brief proposes a pilot architecture, not a claim that a specific model or vendor is ready for evidential use. Procurement should follow testing requirements rather than determine them.
- Some investigative techniques and security controls cannot be published in operational detail. Public accountability can still cover purpose, prohibited uses, governance, aggregate performance and rights protections.
Primary and institutional sources
01