Officer Tanya Reyes finished her last call at 11:47 p.m. on a Tuesday shift in a mid-size Pacific Northwest city. She had responded to a domestic disturbance, a vehicle theft report, two welfare checks, and a minor traffic collision. Five separate incidents. Five separate reports. She clocked out at 1:18 a.m., ninety-one minutes after her last call, having typed every narrative by hand into the records management system (RMS, the agency's central database for case files, incident reports, and evidence logs). The reports were accurate. They were complete. They were also generated during time that could have been spent doing almost anything else, including going home and sleeping before a 6 a.m. call-back for a briefing. That ninety-one minutes is not an outlier. Study after study, and the experience of officers in agencies across the country, confirms that sworn patrol personnel spend somewhere between 30 and 40 percent of every shift on paperwork. For a department running officers ten hours a day, that is three to four hours of every single shift spent in front of a keyboard rather than in the community, on the street, or managing the next call. AI genuinely helps here, and this lesson is about exactly where that help is real, how large it is, and what the evidence behind it actually says.

The Paperwork Problem Is Not a Small Problem

Before mapping where AI helps, it is worth being precise about the scale of the problem it is addressing. The 30 to 40 percent figure is not a rough estimate. It comes from time-on-task analyses of officer activity across multiple departments, and it is consistent enough across jurisdictions that it has become a baseline assumption in public-safety workforce planning. For a department with one hundred sworn officers, that figure means the equivalent of thirty to forty full-time officers are, at any given shift, doing administrative documentation rather than anything requiring a badge.

The paperwork burden is not evenly distributed across incident types. A simple traffic stop with no arrest can produce a short report. A felony arrest with multiple witnesses, body-worn camera (BWC, the recording device officers wear on their uniform to document encounters) footage from three angles, a use-of-force component, and a property seizure can produce a report that takes two hours or more to write correctly. The incidents that matter most legally are also, in most cases, the incidents that create the heaviest documentation burden. The officer who has just handled a complex, high-stakes call is the same officer who then faces the longest report, at the end of a depleting shift, under time pressure to clear the queue.

The result is a set of well-documented problems. Reports written at the end of a long shift contain more recall gaps than reports written closer to the event. Officers who are rushing to finish paperwork so they can respond to the next call sometimes leave out detail that later becomes important in prosecution. Defense attorneys have learned to probe the timing and sequence of report writing. And supervisors tasked with reviewing reports for completeness and accuracy are reviewing documents that reflect, in part, the fatigue and time pressure of the officer who wrote them. This is the context into which AI-assisted report drafting arrived, and understanding that context is necessary for evaluating the claim that it genuinely helps.

Draft One and the 82 Percent Number

In 2023 and 2024, Axon (the body camera and public-safety technology company) began rolling out a product called Draft One. Draft One works inside the Axon body-camera evidence platform. When an officer finishes an encounter recorded on a BWC, the platform transcribes the audio, and Draft One uses that transcription as the primary input to generate a draft narrative for the police report. The officer reviews the draft, corrects and supplements it, and adopts it as the sworn account.

The headline figure from officer testing was an 82 percent reduction in report-writing time. That number deserves careful reading before it is either celebrated or dismissed.

What the 82 percent figure means: officers who participated in the Draft One pilot reported that the time they personally spent writing a report narrative fell by roughly 82 percent compared to writing the same report from scratch. A report that previously took sixty minutes to draft now took approximately eleven minutes to review, correct, and finalize. The draft was already there; the officer's job shifted from author to editor and verifier.

What the 82 percent figure does not mean: it does not mean the reports were 82 percent more accurate, or that the verification step takes 82 percent less time than it should, or that every agency using Draft One experiences exactly that reduction. The figure is a pilot-phase self-report average, not a controlled experimental result. Different call types, different officers, and different agency RMS integrations will produce different time savings. Agencies with complex report requirements, mandatory field-specific entries, or strict supervisor review protocols may see smaller reductions. The honest position, which this program takes throughout, is to treat the figure as a serious and credible benchmark rather than a guarantee.

Even discounted, the number is significant. If the true reduction for a given agency is not 82 percent but 50 percent, and officers were spending three hours per shift on paperwork, that is ninety minutes returned per officer per shift. Across a department of one hundred officers, that is 9,000 officer-hours per month. That is not a rounding error. That is the equivalent of roughly fifty full-time officer-shifts per month, available for patrol, community engagement, follow-up investigations, and anything else the department needs.

How Draft One Actually Works

The mechanism matters because it shapes both the upside and the risk. Draft One is not reading a body-camera video file and visually understanding what happened. It is processing the audio track: transcribing speech, identifying speakers where possible, and generating a narrative that reflects the events described in the audio. The model is drawing on the BWC audio transcript as its primary input, along with any structured data available from the computer-aided dispatch (CAD, the system dispatchers use to log calls, assign units, and track incident details in real time) system, such as the call type, location, and dispatched unit.

This means Draft One can only work with what the microphone captured. If a key interaction happened at a distance from the camera, in a noisy environment where speech was unintelligible, in a language the transcription model did not accurately capture, or during a moment when the officer was not vocalizing observations, that gap in the audio becomes a gap in the draft. The model will not leave a blank; it will either skip the event or, in its more concerning failure mode, fill the gap with a statistically plausible detail that the footage does not actually support. That specific failure mode, the gap-fill detail, is addressed in depth in a later lesson in this chapter. For now, the point is that understanding how the tool works is part of knowing where it genuinely helps and where the officer's verification work begins.

Call Triage and Dispatch Assistance

Report drafting is the most visible AI application in public safety, but it is not the only one that has demonstrated genuine operational value. In the dispatch center, AI is assisting telecommunicators (the professional title for 911 call-takers and dispatchers) in ways that address a different set of pressures.

A busy urban 911 center handles thousands of calls per shift. Each call must be transcribed (to a CAD entry), classified by call type (which determines priority and resource assignment), and in many cases translated if the caller does not speak English. Each of these tasks is currently performed by human telecommunicators under significant time pressure, because a wrong call type means the wrong resources are dispatched, and in the worst cases, the wrong resources arriving too late is the difference between life and death.

AI assists in three specific ways in this environment. First, real-time transcription converts spoken calls to text that can populate the CAD entry with dramatically less manual typing, reducing the transcription burden on the telecommunicator and creating a text record that can be used for quality review later. Second, call classification assistance provides the telecommunicator with a suggested call type and priority based on the content of the call, which the telecommunicator can accept, override, or modify. Third, real-time translation assists with calls placed in languages the telecommunicator does not speak, surfacing the caller's words in the telecommunicator's language with a latency measured in seconds rather than minutes.

Each of these applications is genuinely useful, and each requires the same accountability structure: the human telecommunicator reviews and confirms the AI output before it becomes an operational decision. The AI does not dispatch units. It does not set priority. It suggests, and the trained human decides. That boundary is not just a policy preference; it is a safety-critical design requirement, and it is discussed in depth in the dispatch-focused lessons later in this program.

The Language-Access Dimension

One application that deserves particular emphasis in the dispatch context is language access. Many jurisdictions have legal obligations under Title VI of the Civil Rights Act to provide meaningful access to services for individuals with limited English proficiency (LEP). In practice, providing real-time translation for a 911 call has historically required either a bilingual dispatcher, which is rare in languages beyond Spanish, or a language-line service that adds time and complexity to the call. AI-assisted real-time translation can significantly reduce the latency on language-access calls and can cover a wider range of languages than any reasonable staffing plan can accommodate.

This is a case where AI's genuine help has civil-rights implications in the positive direction: faster, more accurate communication with LEP callers means faster dispatch of the right resources and a better documented record of the call. It is a use case where careful deployment produces both operational and equity benefits.

Redaction and the Public Records Backlog

The records unit of a police department lives with a problem that is rarely visible to patrol but is consuming significant staff time in every agency that handles body-camera footage: public-records requests. Under state public-records statutes (the open-government laws that require agencies to release records to the public on request), agencies must respond to requests for body-camera footage, incident reports, and other records within statutory deadlines, typically ten to twenty business days depending on jurisdiction. Before releasing footage, the agency must redact protected third-party information: the faces and voices of witnesses, uninvolved bystanders, juveniles, and in some cases victims. The license plates of vehicles belonging to uninvolved parties may also require redaction depending on agency policy and state law.

Manual redaction of body-camera footage is extraordinarily time-intensive. A video clip that runs twenty-two minutes takes a records specialist somewhere between forty-five minutes and two hours to redact manually, depending on the complexity of the scene and the number of third-party faces visible. For agencies managing thousands of public-records requests per year, and some large urban agencies see tens of thousands, the manual redaction backlog is a compliance risk, a staffing crisis, and in many cases an ongoing source of litigation because agencies routinely miss statutory release deadlines.

AI-assisted video redaction tools, integrated with body-camera evidence platforms, can identify and blur faces, license plates, and other protected information at processing speeds that are dramatically faster than manual review. An officer-involved video that takes ninety minutes to manually redact can often be processed by an AI redaction tool in under five minutes, with the output then reviewed and verified by a records specialist who confirms the redaction was accurate before release.

The time savings here are as dramatic as in report drafting, and the compliance benefit is measurable: agencies that have deployed AI redaction tools have reported significant reductions in late-release violations. The verification requirement remains essential, because an AI redaction failure (a missed face, a missed plate, an unredacted name audible on the audio track) is not just a technical error. It is a privacy violation, a potential civil liability, and in some cases a public-safety concern if the unredacted footage identifies a witness or confidential informant. The genuine help from AI in this context is real and substantial, and it comes with a verification responsibility that mirrors the report-drafting context.

Summarization Across Long Case Files

Detectives investigating complex cases accumulate enormous volumes of material: hours of interview recordings, stacks of witness statements, crime-scene reports, forensic results, and surveillance footage from multiple locations. Organizing and synthesizing this material has always been one of the most time-intensive parts of investigative work. A detective assigned to a complex fraud investigation or a serious violent-crime case may spend days reading through material before they can even begin to identify the key factual threads.

AI summarization tools can process large volumes of interview transcripts, witness statements, and case documents and produce an organized summary of key facts, timeline elements, and identified witnesses. This is genuinely valuable for getting a fast orientation to a complex file, particularly when a detective is assigned to a case mid-investigation or is picking up a case from a retired colleague. The summary is not the investigation; it is the starting point that allows the investigator to focus their reading on the most important documents rather than working through everything linearly.

The key requirement, and it cannot be overstated, is that every quote, fact, and conclusion in the AI summary must be verified against the source document before it is used in any way that becomes part of the official record. An AI that summarizes twelve hours of interview audio may produce a useful orientation, but an AI-generated summary that contains a fabricated witness quote, an incorrect timeline, or a conclusion not supported by the underlying material is a case file contamination that can affect prosecution, defense disclosure obligations under Brady v. Maryland (the 1963 Supreme Court case holding that prosecutors must disclose evidence favorable to the defense, pronounced Brady), and the integrity of the investigation. The tools are genuinely useful precisely because the detective remains the author and verifier of every conclusion that enters the record.

Evidence Organization and Cross-Referencing

Related to summarization is the AI-assisted task of evidence organization: cataloging and cross-referencing large evidence collections so that an investigator can quickly find the footage, statement, or document they need. In a complex case with dozens of evidence items, finding a specific segment of surveillance video or locating a statement from a particular witness can itself take significant time. AI tools that can tag, index, and search across evidence collections allow investigators to navigate large evidence files in a fraction of the previous time.

This is a lower-risk AI application in terms of the hallucination problem, because the task is organizational rather than generative: the AI is helping find things that exist rather than generating new text about them. The primary risk is missed items (the AI fails to tag a relevant piece of evidence) rather than invented ones. But even here, the investigator's verification role remains essential, because a missed piece of evidence is a disclosure problem just as a fabricated one is.

Reading the Real Evidence: A Framework for Evaluating AI Claims

The applications described in this lesson are supported by real evidence: pilot data, operational testing, and field reports from agencies that have deployed these tools. The 82 percent figure from Axon Draft One is the most prominent single number, and this lesson has treated it with appropriate care. But the broader landscape of AI claims in public safety contains a large number of vendor assertions, conference presentations, and press releases that do not have the same evidentiary foundation.

Every public-safety officer and command staff member evaluating AI tools needs a basic framework for reading AI claims. Four questions are worth asking of every new AI application:

First: What exactly was measured? The 82 percent reduction in report-writing time is a specific claim about a specific metric: the time an officer personally spends on a report narrative. It says nothing about the time spent on verification, supervisor review, or downstream corrections. A claim that is precise about what was measured is more useful than a claim that is vague.

Second: Who measured it and under what conditions? Vendor-sponsored pilot studies conducted with motivated early adopters in controlled conditions will often show better results than the same tool deployed at scale in a department with varied motivation and different workflow constraints. That does not make the vendor data useless, but it does mean the numbers should be treated as a ceiling, not a floor.

Third: What are the failure modes, and how were they measured? A vendor presentation that shows the accuracy rate of correct outputs without equally measuring and disclosing the rate and type of errors is not giving a complete picture. In public safety, the failure modes are not inconveniences; they are Brady problems, impeachment opportunities, and potential wrongful outcomes. Any AI tool evaluation in this context should include a systematic review of how the tool fails, not just how it succeeds.

Fourth: What is the full cost of ownership, including review time? The time savings from AI drafting are real only if the review step does not consume all of the savings and then some. An agency whose verification protocol requires an officer to spend forty-five minutes verifying a draft that took forty minutes to write from scratch has not saved time; it has added a step. The workflow math needs to work, which means the verification step needs to be scoped appropriately to the risk of the specific document type.

These four questions apply to every AI claim in public safety, from dispatch assistance to predictive analytics to surveillance tools. The skill being built in this program is not just the ability to use AI tools, but the professional judgment to evaluate them honestly, deploy them appropriately, and hold them to the same evidentiary standard as everything else in the work.

Key Takeaways

• Officers spend 30 to 40 percent of every shift on documentation. That is the problem AI-assisted report drafting is addressing, and the scale of the savings, even discounted, is operationally significant.
• Axon Draft One reported an 82 percent reduction in officer report-writing time during pilot testing. That figure is a credible benchmark, not a guarantee; actual results vary by agency, call type, and workflow design.
• Draft One works by transcribing BWC audio and generating a draft narrative from the transcript. The officer's role shifts from author to editor and verifier. The officer remains the author of record and the accountable party for the sworn report.
• AI-assisted dispatch tools (real-time transcription, call classification support, and real-time translation) address genuine pressure points in the 911 center. The telecommunicator retains decision authority over priority and resource assignment.
• AI redaction tools can process body-camera footage for public-records release dramatically faster than manual redaction, reducing compliance risk from late-release violations. A verification pass by a records specialist is required before any footage is released.
• AI summarization of large case files is genuinely useful for investigation management. Every quote, fact, and conclusion in an AI summary must be verified against source documents before entering the official record, because fabricated or incorrect summaries create Brady disclosure problems.
• A four-question framework for evaluating AI claims (what was measured, who measured it, what are the failure modes, and what is the full workflow cost) is the professional discipline that distinguishes evidence-based adoption from vendor-influenced enthusiasm.
• All AI applications in public safety share a common accountability structure: the tool assists, the trained human decides, and the professional who reviews and adopts the output is accountable for its accuracy regardless of what generated the first draft.