Commander Elena Vargas ran the operations division of a 220-officer department and had been handed a decision that would have made her predecessor grateful it was her problem and not his: four different AI tool categories were available for deployment, the budget covered two, and the city council wanted an answer by the end of the quarter. Reports, dispatch, records, analysis. Each vendor had a case for why their category was the right starting point. Vargas had learned from the department's readiness assessment that starting with the wrong category could produce a Brady challenge before the program was six months old. She needed a framework that did not just pick the shiniest tool but ranked the options honestly, on the merits that matter in a courthouse as much as in a budget briefing.

Why Prioritization Is a Governance Decision

The choice between deploying AI in report writing, dispatch, records, or analysis is not primarily a technology choice or a budget choice. It is a governance choice, because each use case carries a different accountability structure, a different evidentiary exposure, and a different community-trust profile. Getting the order wrong does not produce a failed pilot; it can produce a case that gets thrown out, a moratorium demanded by the prosecutor's office, or an oversight investigation triggered by community concerns about a high-risk tool deployed without adequate safeguards.

Commander Vargas's decision framework had to answer three questions for each use case: What is the value, measured in time returned, cost avoided, or capability gained? What is the evidentiary risk, measured in what happens if the AI output is wrong and reaches a sworn record, a dispatch decision, or a court filing? And what is the community-trust profile, measured in how the public and oversight bodies would view this use case if they read about it in a public-records request?

The impact-and-risk matrix answers all three. It does not produce a single right answer for every agency, because every agency has a different starting posture. An agency that has already built strong verification discipline from an existing body-worn camera (BWC) program has a different starting point than one deploying cameras for the first time alongside AI. But the matrix produces the right reasoning process: an honest ranking of value against risk that a chief can defend to the prosecutor, to city council, and to the community advisory board in the same conversation.

Prioritization is not about picking the most impressive tool. It is about picking the tool whose failure mode the agency can manage before picking the one whose failure mode is a suppression motion.

The Records and Redaction Case: The Defensible Start

AI-assisted redaction of body-worn camera footage and records for public-records responses is the strongest starting point for most agencies, and the reasoning is worth laying out in full because it is counterintuitive. Redaction does not feel like the high-value use case. Report writing feels like the high-value use case, because officers spend 30 to 40 percent of every shift on paperwork and report drafting returns the most dramatic time savings.

But the impact-and-risk matrix is not just about value. It is about value weighted against risk and compared against the agency's current ability to manage that risk. On that combined measure, records and redaction comes out ahead as a starting point for most agencies.

The Value Side

The value case for records and redaction is real and often underestimated. Public-records requests involving body-camera footage are among the most time-consuming work in a records unit. A single contested case can involve dozens of hours of footage, each frame of which potentially contains a third-party face, a license plate, a juvenile, or protected victim information that must be manually reviewed and redacted before release. In high-volume agencies, records request backlogs run to hundreds of pending cases and weeks or months of delay. AI-assisted redaction can reduce per-request processing time by 60 to 80 percent for large-footage requests, turning a two-week backlog item into a two-day task. That is not a headline number like the 82-percent reduction in report-writing time associated with tools like Axon Draft One, but it is a sustainable and defensible number with a direct impact on the records unit's capacity, the department's compliance with open-government statutes, and the community's ability to access information about incidents that involve them.

The Risk Side

The consequences of a redaction error are serious. A missed face or license plate in a released video is a privacy violation. It can result in a complaint, a civil claim, and a records-correction obligation. In a high-profile case, it can create significant media and community concern. All of this is real and must be managed with a verified human review step before release.

But the critical distinction is what a redaction error does not do. It does not go into a sworn report. It does not become an officer's testimony under oath. It does not create a Brady (Brady v. Maryland, requiring exculpatory evidence disclosure) or Giglio (Giglio v. United States, requiring disclosure of impeachment evidence) issue in a criminal case. The correction pathway is: identify the missed element, issue a corrected release, document the error, and implement a tighter verification standard going forward. That is a serious and workable correction pathway. The correction pathway for an AI gap-fill in a use-of-force sworn report is deposition testimony, suppression motions, and potential case dismissal. The difference in consequence is the difference in phase priority.

The Dispatch and Call Summarization Case: High Value, Human-Owned Priority

AI assistance in computer-aided dispatch (CAD) and call handling occupies an interesting position in the prioritization matrix. The value is significant and the stakes are unique: in dispatch, an AI error is not a courtroom problem, it is a life-safety problem. A misclassified call type or a delayed priority assignment can mean responders arrive with the wrong resources for a medical emergency, or that a high-priority domestic violence call is routed as a lower-priority welfare check. These are not evidentiary failures. They are operational failures with potentially catastrophic consequences.

The prioritization guidance for dispatch is therefore different from the records-versus-reports comparison. Dispatch AI should be prioritized for the functions where AI assistance has the clearest value and the most recoverable failure modes: call transcription, translation support for non-English callers, and after-the-call summarization for CAD entry. These functions return time to the telecommunicator without placing AI output in the critical path of a real-time priority decision. The telecommunicator remains in command of the classification and the dispatch.

The specific dispatch functions where AI should not be in the priority chain include real-time call prioritization recommendations and automated dispatch routing. The lesson from dispatch AI governance is clear: the classification of whether a call is a life-threatening emergency is a human decision, always, because an AI system that gets it wrong is not producing a correctable report error, it is producing a delayed response that may have a permanent consequence for someone waiting on the line.

For prioritization purposes: dispatch transcription and summarization belong in phase one alongside records and redaction because they are high value and have recoverable failure modes. Real-time dispatch AI assistance belongs in a later phase, after extensive testing under conditions that do not place unevaluated AI output in the priority chain of a live emergency.

The Report Writing Case: The Highest Value, The Highest Stakes

AI-assisted report drafting is the use case that produces the most dramatic efficiency numbers and the most significant governance questions. Both facts are true and both must be presented to command staff, city council, and the community at the same time.

The value case is compelling. Officers spend 30 to 40 percent of every shift on paperwork. A tool that reduces report-writing time by 60 percent in operational settings (even if the controlled-testing benchmark is 82 percent) returns roughly 18 to 24 percent of every shift to sworn duties: patrol, investigations, community engagement, and crisis response. In a 220-officer department, that is the equivalent of adding approximately 40 officer-equivalents of capacity without a single new hire. The budget impact over a five-year contract is potentially larger than the contract cost.

The risk case is equally compelling and must be stated with equal prominence. A sworn police report is evidence. It is disclosed to the defense under Brady. An AI gap-fill in a narrative, adopted without verification, is the officer's sworn account. When defense counsel plays the body-camera footage and the sworn account differs from the footage, the question in court is not whether the AI was wrong. The question is why the officer signed a report they had not verified. That question does not have a good answer, and a series of those questions becomes the kind of systemic discovery pattern that led the King County, Washington prosecutor's office to bar AI-written police reports entirely.

The Axon Draft One 82-percent time reduction figure deserves its own honest analysis. That figure comes from controlled testing with officers who were trained specifically for the test environment. Operational results in Commander Vargas's peer agencies ranged from 45 percent to 65 percent reduction, with the verification pass adding back 12 to 20 minutes per report. The net benefit is still significant. But presenting the 82-percent figure to city council without the operational-context caveat is a governance failure: it sets an expectation the program cannot reliably meet, and when the first performance review finds 55 percent rather than 82 percent, the program's credibility suffers in a way that affects every other AI initiative the department is trying to advance.

Prioritization Position

AI-assisted sworn report drafting belongs after the records and dispatch summarization phase because: the training program required to do it safely takes time to build and verify; the prosecutor's sign-off on the disclosure standard must be in place before the first case using AI-assisted reports is charged; and the command culture that protects the time required for a thorough verification pass must be demonstrated at lower stakes before it is tested at sworn-report stakes. None of those prerequisites are impossible. All of them take time. The prioritization order exists to ensure that time is invested before the exposure, not after.

The Analysis Case: Highest Stakes for Civil Liberties

AI-assisted analysis covers a spectrum of capabilities: summarizing case documentation and evidence for investigators, identifying patterns in crime data, generating risk assessments for individuals or locations, and automated surveillance analysis. The spectrum ranges from high value with manageable risk (case summarization) to contested value with significant civil-liberties exposure (individual risk scoring).

Case summarization for investigators is the clearest starting point within this category. An investigator working a complex case can have hundreds of hours of recorded interviews, thousands of pages of documents, and evidence from multiple sources that must be synthesized into a working theory and a prosecution summary. AI-assisted summarization can reduce the time required to identify key statements, flag inconsistencies, and produce an initial working synthesis from weeks to days. The risk is real: a fabricated quote attributed to a witness in an AI-generated case summary is a case-ending error. But the risk is manageable with the same verification discipline the report-writing phase establishes: every quote verified verbatim against the recording, every key factual claim traced to a specific source document.

Individual risk scoring is categorically different. The Electronic Frontier Foundation (EFF) and other civil-liberties organizations have documented the specific problem with AI systems that assign risk scores to individuals based on historical data: those systems encode the historical patterns of policing, which are not neutral with respect to race, neighborhood, or socioeconomic status. A model trained on arrest data will reflect who was arrested under the policing practices of the past, which is not the same as who poses a public safety risk in the future. The consequential errors are not random; they are likely to fall disproportionately on communities that were already subject to more intensive policing. That is not a small qualification to manage away with a disclosure notation. It is a structural problem that requires genuine engagement with community oversight before the tool is deployed at all.

For prioritization purposes: case documentation summarization for investigators belongs in a later phase, after the verification culture and disclosure infrastructure are proven in the report-writing phase. Individual risk scoring belongs in a dedicated evaluation process that precedes any deployment decision, conducted with community oversight involvement and an independent bias audit of the specific model being considered.

Building the Priority Matrix: Commander Vargas's Decision

When Commander Vargas applied the impact-and-risk matrix to her four categories, the ranking looked like this.

Records and redaction: Phase one. High operational value, manageable failure mode, recoverable error pathway. Builds the verification discipline needed for phase two without sworn-report stakes. Delivers visible results to the community (faster public-records responses) while building internal competency.

Dispatch transcription and summarization: Phase one, alongside records. High operational value, high volume, recoverable failure mode. Human-owned priority decision at all times. Delivers measurable time savings to the dispatch center without placing AI in the critical path of emergency priority.

Sworn report drafting: Phase two. Highest operational value, highest evidentiary risk. Requires phase-one infrastructure: CJIS compliance documentation, written AI use policy, trained officers with verified competency, prosecutor sign-off on disclosure, community engagement record. When phase-one conditions are met and phase-two is introduced to a scoped cohort with active monitoring, the efficiency gains are real, the accountability is maintained, and the program is defensible.

Investigative case summarization: Phase three. High investigative value, high verification requirement, lower community-trust concern than individual risk scoring. Requires phase-two verification culture and disclosure practices to be mature before investigators apply those habits to case-summary AI output, which carries equally high stakes in a different context.

Individual risk scoring: Not in the roadmap until a dedicated bias audit and community oversight engagement process have been completed for the specific model being considered. The EFF's concerns are not a reason never to evaluate such tools. They are a reason to evaluate them with the seriousness the civil-liberties implications require.

Commander Vargas presented this matrix to her city council with the value figures, the risk figures, and the sequencing rationale in the same briefing document. She did not present only the efficiency numbers. She presented the accountability framework alongside them. The council approved the phase-one budget and asked to receive the monitoring results before approving phase two. That is exactly the right governance relationship between a command staff and a city council on AI deployment.

Key Takeaways

• Prioritization of AI use cases is a governance decision, not a technology or budget decision. The order determines the accountability exposure before the efficiency gain is realized.
• The impact-and-risk matrix evaluates each use case on three dimensions: value delivered, evidentiary risk if the output is wrong, and community-trust profile. All three belong in the same briefing to council.
• Records and redaction is the strongest starting point for most agencies: high operational value, a manageable and recoverable failure mode, and a verification discipline that builds the competency needed for report drafting. A missed redaction is a privacy violation and a serious problem; it is not a Brady challenge.
• Dispatch transcription and CAD summarization belong in phase one alongside records: high value, human-owned priority, recoverable failure mode. Real-time AI dispatch priority recommendations belong in a much later and more carefully evaluated phase.
• AI-assisted sworn report drafting carries the highest evidentiary stakes because a sworn report is evidence. The 82-percent time-reduction benchmark is from controlled testing; operational results have ranged from 45 to 65 percent, with the verification pass adding 12 to 20 minutes per report. Both figures belong in any briefing to council.
• Individual risk scoring requires a dedicated bias audit and community oversight engagement before any deployment decision, because the algorithmic failure mode is not random: it tends to fall on communities already subject to more intensive policing, making the civil-liberties exposure structural, not incidental.
• The King County prosecutor's bar on AI-written police reports is a prioritization lesson: high-evidentiary-risk capabilities were activated before the governance infrastructure (trained officers, verified disclosure, prosecutor coordination) was ready.
• Brady v. Maryland and Giglio v. United States frame the evidentiary-risk dimension of the matrix: any AI output that could touch a sworn record, a charging decision, or a disclosure package carries constitutional disclosure implications that must be resolved before deployment, not discovered after the first defense motion.