{"resourceId":"manchester-draft-one-randomized-trial","versions":[{"version":"external-85f042b2f8a4861420234196a1b9b011a046285b9cfc8df113dd30172713d1e0","resource":{"id":"manchester-draft-one-randomized-trial","title":"Police report-writing trial finds no significant time savings","organization":"Ian T. Adams and colleagues; Manchester Police Department","sector":"Law enforcement","geography":"Manchester, New Hampshire, United States","publishedAt":"First published October 2, 2024; assigned to 2026 journal volume","publicationDate":"2024-10-02","eventDate":null,"sourceName":"Journal of Experimental Criminology","sourceLabel":"Peer-reviewed randomized field trial","sourceUrl":"https://link.springer.com/article/10.1007/s11292-024-09644-7","evidenceClass":"academic-research","outcomeClass":"cautionary","topics":["knowledge-work","infrastructure","governance-procurement","operating-model"],"finding":"Access to Draft One did not significantly reduce report completion time in a preregistered Manchester trial.","sledRelevance":"Baseline evidence newly added to this archive for police departments considering narrative copilots; this is not a newly conducted trial.","evidence":"85 officers were assigned to usual reporting (43) or AI access (42). Mixed-effects analysis covered 755 reports using RMS creation-to-submission timestamps; a 6,084-report annual analysis also found no significant effect.","architectureImplications":"Interpretation: Instrument the full body-camera-to-RMS workflow, including uploads, transcription, structured entry and reviewer corrections. Compare approved cloud processing against local data-handling requirements; the paper does not establish an on-premises deployment option.","governanceImplications":"Interpretation: Separate a pilot decision from a staffing reduction; require independently scored quality alongside elapsed time.","securityPrivacyImplications":"Interpretation: Restrict access to witness and victim recordings, retain provenance under approved records rules, and test export and deletion.","caveats":"Single-agency intent-to-treat study: access did not mean use on every report. Elapsed timestamps are not active labor time. Report quality and downstream benefits were not established; a null result does not prove universal ineffectiveness.","streamIds":["public-safety"],"roles":{"sales":"Interpretation: Police command, records managers, finance, prosecutors and defenders need to know whether narrative assistance relieves their actual bottleneck. Ask how much work lies in structured fields, how often reports return for correction, and whether claimed savings include review. A bounded engagement could baseline one incident category and compare approved workflows. The value hypothesis is better allocation of effort if local results justify it. Do not promise headcount reductions or transfer a vendor's advertised savings to a department. This study supports testing the business case; it does not establish that a current product version will fail, or that report quality improves.","engineering":"Interpretation: Fit is limited to assisted documentation under officer control. Map camera ingestion, transcripts, narrative generation, RMS fields and final approval before integrating. Prerequisites include usable timestamps, comparable incident categories, authorized test data and reviewer availability. Test transcript errors, missing visual observations and interrupted uploads. Apply case permissions, encryption and controlled retention across intermediate artifacts. A proof of value should compare total handling effort and blinded factual accuracy against the existing workflow, recording actual tool use as well as assignment. Agents should prepare drafts without independent filing authority. Revalidate after model changes; this historical trial is not a benchmark of every current deployment.","delivery":"Interpretation: The records unit should own implementation with IT, training staff and legal reviewers. Establish a baseline, train a limited cohort, record upload and correction effort, and review exceptions weekly. Dependencies include camera availability, RMS interfaces and a lawful retention plan. Proposed acceptance criteria are a preregistered improvement target for total effort, no deterioration in independently checked factual accuracy, and successful permissions and export tests. These are future criteria, not observed findings. Adoption reviews should identify who stops using the tool and why. Pause expansion if review work grows or critical facts are omitted, and retain a conventional reporting path."},"retrievedAt":"2026-09-07T03:01:36Z","enrichedAt":"2026-09-07T03:01:50Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Interpretation: Include officers with varied writing and language needs in validation; assess narration and correction burdens.","procurementImplications":"Interpretation: Price the complete evidence and records integration, reviewer labor and exit costs; avoid purchasing against assumed savings.","operatingModelImplications":"Interpretation: A records supervisor should own report acceptance, with evaluation support separate from vendor success reporting.","sourceVerification":{"openedUrl":"https://link.springer.com/article/10.1007/s11292-024-09644-7","referenceExcerpt":"AI assistance did not significantly affect the duration of writing police reports.","promptVersion":"sled-research-v3.1","model":null,"basis":"agent-reported inspection"}}}]}