AI use case

AI for police incident report drafting from body worn camera audio

An AI tool that turns the transcript of body worn camera audio into a first draft of a police incident report narrative, which the officer who was on the call must review, correct and sign before it becomes part of the official record.

By Len Debets · Last verified 29 September 2026 · 2 public deployments

82%
Reported handling time reduction
Fort Collins Police Services, organization claim.
About 750 hours
Hours saved
Rochester Police Department (Minnesota) (organization claim).
USD 81,000 to USD 1.1 million
Indicative value per year
A police department with 150 sworn officers who write reports. Worked example, see how it is calculated.

What problem does it solve?

Writing the report narrative after a call for service is one of the most time consuming parts of patrol work, and it happens after nearly every incident, not just the serious ones. Axon, whose body worn cameras and records systems are used by police departments across the US, reports that officers can spend a large share of their working week on this kind of paperwork. Every hour spent typing a narrative is an hour not spent on patrol, an investigation, or simply talking with the community, and departments already short of officers feel the cost most.

An AI tool that drafts the narrative from the audio the camera already recorded promises to give that time back, but it raises a real question: who is accountable for what the report says once a machine wrote the first version. That promise does not always hold: a peer reviewed study of the Manchester, New Hampshire police department found no time saved once officers' editing time was included, and Manchester and the Anchorage, Alaska police department both stopped using the tool. The two examples on this page show departments using the tool in practice, and independent reporting on the same product shows why the review step has to be genuine, not a formality.

How does it work?

  1. Capture. The officer's body worn camera records audio during the call for service or incident, as it already would without the tool.
  2. Transcribe. The audio is automatically transcribed once the officer ends or uploads the recording, usually within minutes.
  3. Draft. A language model turns the transcript into a narrative in the department's own report format, using only what the audio contains and leaving a visible placeholder wherever information is missing.
  4. Review against the recording. The officer reads the draft while checking it against the recording itself, corrects anything wrong, and completes every placeholder.
  5. Sign and submit. The officer certifies that the report is accurate before it enters the records management system; only that signed version is the official record.
Audience
Employee facing
Autonomy
Copilot
Adoption
Early adopters
Channels
Internal tools

What is it worth?

Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.

Value benchmarks for AI for police incident report drafting from body worn camera audio
KPIMedianReported rangeData pointsClaimed by
Handling time reductionToo few to pool
82%
11 organization
Hours savedNot pooled
about 750 hours
11 organization

Value drivers: Employee productivity, Speed and cycle time.

Indicative value

A police department with 150 sworn officers who write reports

USD 81,000 to USD 1.1 million

Officer time cost avoided on report writing per year

How this is calculated

Formula: officers * reportsPerOfficerPerWeek * (minutesSavedPerReport / 60) * weeksPerYear * costPerOfficerHour. The low scenario uses every low input, the high scenario every high input.

InputLowHighBasis
Sworn officers who write reports officers, officers150150The reference department, close in size to Rochester Police Department's approximately 160 sworn officers.
Reports written per officer per week reportsPerOfficerPerWeek, reports per officer per week35Editorial assumption for a mid sized US municipal department. Replace with your own reporting volume.
Minutes saved per report minutesSavedPerReport, minutes per report525Axon's customer story on Rochester Police Department states that officers save roughly 20 to 25 minutes per report with Draft One, and Fort Collins Police Services (Colorado) separately reported an 82% reduction in report writing time during its trial. Set the low end well under that: a 2024 peer reviewed study of the Manchester, New Hampshire police department found no improvement in report filing time because officers spent significant time editing the drafts, and Manchester and the Anchorage Police Department in Alaska both stopped using Draft One, Anchorage citing zero time savings. Treat this as an uncertain range and validate it against your own department's experience.
Working weeks per year weeksPerYear, weeks4848Standard allowance for leave.
Fully loaded cost per officer hour costPerOfficerHour, USD per hour4575Editorial assumption for a US municipal police officer. Replace with your own fully loaded cost.

What it leaves out: Gross time value only. It leaves out the cost of the AI service itself, the officer time still needed to review every draft against the recording, and any rework cost from an error that is not caught before the report is signed. It also assumes time is actually saved: at least one department (Manchester, New Hampshire) found none in a peer reviewed study and discontinued the tool, so a department should measure its own result before counting on this range.

Who already uses it?

2 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.

Rochester Police Department (Minnesota)

United States · Government and public sector · 2026

ProductionGrade C

Rochester Police Department in Minnesota, a department of about 160 sworn officers, replaced a 30 year old records system with Axon Records and added Draft One, which Axon's own page describes as generating "a first draft from officer interviews/inputs". In an Axon customer story, the captain who led the rollout said the department's officers saved a total of hundreds of hours after writing 1,800 reports over two months with the tool, alongside faster access to linked video evidence.

  • Hours saved: about 750 hours, over two months (1,800 reports written), as stated by the captain; not an annual figure
    "If we've written 1,800 reports over the last two months, our officers have saved about 750 hours, and then we can reinvest that time into the community."
    Claimed by: organization

Fort Collins Police Services

United States · Government and public sector · 2024

PilotGrade C

Fort Collins Police Services (Colorado) tested Axon's Draft One, which turns the transcript of body worn camera audio into a first draft police report narrative for an officer to review and sign. A department sergeant reported a large drop in the time officers spent writing reports during the trial, along with a substantial improvement in report quality, in Axon's own product launch announcement.

  • Handling time reduction: 82%, during the agency's trial of Draft One
    "Our agency has been testing Draft One, and we have seen an 82% decrease in time spent writing reports."
    Claimed by: organization

How do you implement it?

A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.

Data you need

  • Body worn camera audio with reliable automatic transcription
  • The department's own report narrative template and writing conventions
  • A written policy defining which incident types and charge levels the tool may draft for

Systems to integrate

  • Body worn camera and digital evidence platform
  • Records management system that holds the official report
  • Case and charge management system, so charge levels stay linked to what the tool may draft

Complexity: Medium

Transcribing audio and drafting a narrative is the easy part. The real work is restricting the tool to a safe set of incident types and charge levels, integrating it with the department's records management system, and building a review process that catches errors before a report is signed, not after.

  1. 1

    Start with minor, low risk incident types

    Axon's own default configuration excludes arrests and felony charges from Draft One at launch. Start the same way and expand only once officers and supervisors have reviewed enough drafts to know the pattern of errors the tool makes.

  2. 2

    Keep the model tied to the recording

    Draft only from the transcript of the audio actually captured, with no invented detail, and leave a visible placeholder wherever the audio does not cover something the report needs.

  3. 3

    Require a genuine edit, not a rubber stamp

    Make officers open, read and complete every placeholder before they can sign, and keep the AI drafted version and the officer's edited version as separate records, so the department can show what was actually reviewed if a report is challenged in court.

  4. 4

    Run your own quality comparison before rollout

    Axon compared Draft One narratives with officer only narratives across several quality dimensions using independent reviewers before making claims about quality. Run a similar comparison on your own reports before and after rollout instead of relying only on a vendor's study.

  5. 5

    Train supervisors to sample, not just approve

    Have a supervisor review a random sample of signed reports each week against the underlying recording, specifically checking for facts the AI added or missed that the officer did not catch.

Guardrails

  • Restricted to a defined list of low risk incident types and charge levels at launch
  • Officer sign off required before a report enters the records management system
  • No draft text generated from anything other than the audio transcript
  • The AI drafted version is kept separate from the officer's edited version, for discovery and audit

KPIs to instrument

  • Minutes saved per report compared with the same report type before the tool
  • Rate of factual corrections found in supervisor sampling
  • Share of reports where a placeholder was left unedited before signing, as a warning sign
  • Prosecutor and defense feedback on report clarity and completeness

Human in the loop

The officer who was on the call reviews, completes and signs the report; the tool never submits a report on its own, and the department's written policy states which incident types are eligible for it.

Common failure modes

A review that does not really review
Reporting on Draft One has found that the product does not store the original AI draft at all, by design, which makes it hard to show what was genuinely reviewed once the officer's edited version is the only copy left. Keep the original AI draft and the edited version as separate, retained records.
A draft that still needs real correction, not a rubber stamp
A 2024 peer reviewed study of the Manchester, New Hampshire police department found officers had to remove irrelevant information, fix inaccuracies and add facts the AI missed on Draft One narratives, and King County prosecutor Daniel Clark has said he has seen AI written reports get the names of witnesses and officers wrong or place an officer at a scene they only heard over the radio. Require the officer to check the draft against the recording itself, not just read the text for tone and grammar.

What are the risks and rules?

EU AI Act

Limited risk (transparency)

A tool that only drafts a narrative for a named officer to review, correct and sign is not listed in Annex III: it does not evaluate the reliability of evidence or a person's risk of offending or reoffending, the law enforcement uses Annex III point 6 covers. Article 50(4)'s disclosure duty for AI generated text does not apply either way, because that duty covers only text published to inform the public on matters of public interest, and an internal police report is not published for that purpose; officer review is not what exempts it. Article 50(2) is the provision that does apply: a provider whose system generates synthetic text from a transcript, rather than only lightly editing text the deployer already supplied, must mark its output in a machine readable, detectable format. That places the tier at limited, not minimal. It would move toward Annex III point 6, and a high tier, if the same kind of system were extended to judge witness credibility or predict reoffending rather than to draft a narrative.

Guidance

Controls to put in place

  • Written agency policy on which report types and charge levels the tool may draft
  • Disclaimer kept with every report noting it was drafted with AI assistance
  • Audit trail that keeps the AI draft separate from the officer's edits
  • Supervisor sampling of signed reports against the underlying recording

When it went wrong elsewhere

  • Axon's Draft One Is Designed to Defy Transparency. The Electronic Frontier Foundation reports that Draft One does not store the original AI draft at all: the officer copies it into the report and it disappears as soon as the browser window closes, by design, which makes it hard for defense lawyers, prosecutors and the public to audit an AI drafted report.
  • Axon Says AI Police Reports Save Time. Public Records Show They Get Facts Wrong. Forbes reported public records showing errors in Axon's separate Form One product, which fills in names, plates and ID details, at the Lafayette, Indiana police department. On Draft One narratives specifically, Forbes cites a 2024 peer reviewed study of the Manchester, New Hampshire police department that found officers had to remove irrelevant information, fix inaccuracies and add missed facts, and quotes King County prosecutor Daniel Clark saying he has seen AI written reports get witness and officer names wrong, which is why a genuine officer check against the recording matters more than the tool itself. forbes.com blocks automated fetches (403); this was checked against the Wayback Machine copy at https://web.archive.org/web/2026/https://www.forbes.com/sites/thomasbrewster/2026/07/22/axon-says-ai-police-reports-save-time-public-records-show-they-get-facts-wrong/.
  • Guardrail bug disclosed to Frederick PD and King County's refusal to accept AI drafted reports. The Electronic Frontier Foundation reports that Axon disclosed to the Frederick Police Department in Colorado that engineers had found a bug that let officers circumvent Draft One's review guardrails on at least three occasions, and that the King County Prosecuting Attorney's Office in Washington has told police "our office has made the decision not to accept any police narratives that were produced with the assistance of AI."

Frequently asked questions

Does an AI write the final police report?
No. Every example on this page requires the officer who was on the call to review, correct and sign the draft before it becomes the official report. Axon built Draft One so that every report must be reviewed and approved by a human officer. Axon's customer story on Rochester Police Department describes Draft One as generating "a first draft from officer interviews/inputs", and Axon's own product design is built around the officer completing that draft, not a finished report.
How much time does this actually save?
Public figures vary sharply by department. Fort Collins Police Services (Colorado) reported an 82% decrease in time spent writing reports during its trial, and Rochester Police Department in Minnesota reported saving about 750 officer hours after writing 1,800 reports over two months with Draft One. Neither figure is an independently audited, company wide result: the Fort Collins number is from Axon's own product launch press release and the Rochester number from an Axon customer story. Independent evidence points the other way at some departments: a 2024 peer reviewed study found no improvement in report filing time at the Manchester, New Hampshire police department because officers spent significant time editing the drafts, and Manchester and the Anchorage Police Department in Alaska both stopped using Draft One, Anchorage citing zero time savings.
Are AI drafted police reports accurate?
Not automatically. Axon's own double blind study found Draft One narratives performed as well as or better than officer only narratives on several quality measures, but independent evidence points the other way on some points: a 2024 peer reviewed study of the Manchester, New Hampshire police department found officers had to remove irrelevant information, fix inaccuracies and add missed facts on Draft One drafts, and King County prosecutor Daniel Clark has said he has seen AI written reports get the names of witnesses and officers wrong. A genuine officer review against the recording matters more than the technology itself.
Is this high risk under the EU AI Act?
Usually not under Annex III, as long as the tool only drafts a narrative: Annex III's law enforcement category covers systems that evaluate the reliability of evidence or a person's risk of offending or reoffending, not drafting assistance. Separately, the tool's provider likely has to mark the generated text as artificially produced under Article 50(2) of the EU AI Act, which is what puts the system at limited risk rather than minimal, regardless of officer review. It would need Annex III's full high risk governance if the same system judged evidence reliability or predicted reoffending instead of drafting a narrative.

How to cite this page

Blits.ai AI Use Case Library, "AI for police incident report drafting from body worn camera audio", last verified 29 September 2026, https://www.blits.ai/ai-use-cases/police-incident-report-drafting. Licensed under CC BY 4.0. Method: how we verify use cases.

Changelog
  • 29 September 2026: First published

Related use cases

Government and public sectorHealthcare

AI support for emergency call triage (112 and 911)

AI that supports emergency call takers and dispatchers during 112 and 911 calls, with live transcription, translation, summaries, location cues and alerts for critical conditions such as cardiac arrest, while the call taker keeps every triage and dispatch decision.

Deployments
3 public, best grade B
Autonomy
Assist
Government and public sector

AI for non emergency service requests and 311 routing

An AI agent on a city's 311 style phone, chat and messaging channels that answers routine municipal questions, takes service requests such as potholes, missed collections or broken street lights with the right location and details, creates the case in the work order system and routes anything urgent or complex to the right team.

Deployments
7 public, best grade B
Reported accuracy
80%
City of Kelowna, vendor claim
InsuranceHealthcare

AI for health insurance prior authorization and claims adjudication support

AI that reads prior authorization requests, medical claims and appeals with their clinical and billing documents, extracts diagnoses, treatments and costs, checks them against the policy and published clinical criteria, and prepares a summary and recommendation for a clinician or adjudicator, who makes every adverse decision.

Deployments
5 public, best grade B
Reported handling time reduction
about 50%
Acentra Health, vendor claim
Cross industryBanking

AI for inbound correspondence triage and routing

AI that sorts inbound correspondence before anyone answers it: it takes every inbound letter, email, upload and secure message into one intake, identifies what it is, extracts the key fields, links it to the right customer and account, sets priority and routes it to the right team or workflow, replacing the manual sorting desk.

Deployments
6 public, best grade B
Reported accuracy
91%
Travelers, vendor claim