Deputy Chief Marcus Reyes had the readiness assessment on his desk, two blocking gaps closed, and a command staff that had been waiting four months for a roadmap. The vendor was ready to turn on the AI drafting platform for all 340 officers on the day the contract activated. Reyes knew that was the wrong move. An 82-percent reduction in report-writing time was the headline; a suppression motion in month three because nobody had trained the night shift on the verification standard was the story he was trying not to write. He needed a roadmap that sequenced the wins in the right order, earned trust with the prosecutor's office before the first contested case arrived, and kept the agency ahead of its own risk rather than chasing it.

What a Roadmap Is Not

An agency AI roadmap is not a deployment calendar. A deployment calendar answers the question "when does each unit go live?" A roadmap answers the harder question: "in what order should we take these capabilities, and why, given what we know about value and risk?"

The distinction matters because AI use cases in public safety vary enormously in their evidentiary stakes. A tool that assists with redacting third-party faces from body-camera footage before a public-records release carries a different risk profile than a tool that drafts use-of-force report narratives that will be disclosed to the defense, tested in depositions, and potentially used to challenge an officer's credibility. Both save time. Both have verification requirements. But they do not belong on the same phase of a roadmap, because getting the redaction tool wrong produces a privacy complaint and a records correction, while getting the use-of-force drafting tool wrong produces a suppression motion, a Brady (Brady v. Maryland, requiring exculpatory evidence disclosure) challenge, and potentially a wrongful outcome.

A roadmap that does not account for this difference is a deployment calendar dressed up in strategy language. It answers the budget question but not the accountability question. Deputy Chief Reyes needed both answered.

Sequence your AI wins by value and evidentiary risk, not by vendor readiness. Take the wins you can fully defend before the ones that invite a case challenge.

The Value-Risk Framework for Sequencing

The core tool for building a public safety AI roadmap is a two-axis framework that plots each proposed use case against two dimensions: the value it delivers (measured in time returned, costs avoided, or capability gained) and the evidentiary risk it carries (measured in the consequence if the AI output is wrong and gets into a sworn record, a charging decision, or a courtroom).

This is not an abstract academic exercise. Each use case lands in one of four quadrants, and the quadrant determines the roadmap phase.

Quadrant One: High Value, Lower Risk

These are the first-phase wins. They deliver real, measurable benefit and carry consequences for error that are serious but recoverable without courtroom stakes.

AI-assisted redaction of body-worn camera (BWC) footage for public-records releases is the clearest example. Officers spend significant time on records requests: a single complex request involving hours of footage can take a records clerk an entire week of manual review. AI-assisted redaction can reduce that to hours. The consequences of an error are real: a missed face or license plate in a released video is a privacy violation that can result in a complaint, a lawsuit, or a records correction. But it does not typically put someone in prison or set them free. It does not go into a sworn report. It does not test the verification standard under cross-examination.

AI-assisted call summarization for computer-aided dispatch (CAD) entry is in the same quadrant. The telecommunicator reviews and adopts the summary. A wrong CAD entry has consequences, but the correction pathway is fast and the human review at the dispatch console is robust. The time savings are real: a busy dispatch center handles hundreds of calls per shift, and even modest reductions in documentation time per call compound significantly over a month.

AI assistance with routine administrative documentation, training records, and non-evidentiary correspondence also belongs in this quadrant. These uses save time with minimal stakes.

Phase one of the roadmap should consist entirely of quadrant-one use cases. They build the agency's AI competency, produce measurable results the chief can present to council, and create the verification culture the higher-stakes phases will require, without creating the exposure that comes from deploying use-of-force drafting before the training program is mature.

Quadrant Two: Lower Value, Lower Risk

These use cases belong in the roadmap eventually but are not first-phase priorities. They include AI-assisted scheduling, AI-generated training materials, and AI summarization of non-evidentiary research. They should be considered after phase one is stable and producing the culture and competency the roadmap needs.

Quadrant Three: High Value, Higher Risk

This is where the most important AI use cases in public safety sit: AI-assisted sworn report drafting from body-camera audio and footage. The value is large and real. Officers spend 30 to 40 percent of every shift on paperwork. A tool that returns even half of that time to patrol, community engagement, and investigations has a significant operational impact. Testing of tools like Axon Draft One has shown up to an 82-percent reduction in report-writing time under controlled conditions. In an agency of 340 officers, that is thousands of hours per month redirected to sworn duties.

But the evidentiary risk is also at its highest here. A sworn police report is evidence. It is disclosed to the defense under Brady. It is tested in depositions. The narrative of a use-of-force incident, drafted with AI assistance and not verified against the footage, can contain gap-fills that are plausible, professionally worded, and wrong. Those gap-fills, if adopted without correction, become the officer's sworn account. When a defense attorney plays the body-camera footage and the sworn account differs from the footage, the gap is not an AI problem in the courtroom: it is the officer's problem. The King County, Washington prosecutor's bar on AI-written reports was a direct response to cases where the controls were not yet in place to prevent exactly this scenario.

Quadrant-three use cases belong in phase two of the roadmap, after the agency has demonstrated in phase one that its officers understand verification, that the command culture supports the time required for a thorough pass, and that the prosecutor's office has signed off on the disclosure standard. They should launch with a scoped cohort, active monitoring, and a defined error-escalation pathway before expanding to the full agency.

Quadrant Four: Lower Value, Higher Risk

These use cases should not appear in an early roadmap. They carry high evidentiary or civil-liberties risk without the commensurate value to justify the exposure. Predictive risk scoring for individuals falls here: the value is contested, the algorithmic bias risks are substantial, the transparency requirements are demanding, and the civil-liberties objections from groups like the Electronic Frontier Foundation (EFF) are serious and documented. A chief who deploys a quadrant-four use case before the agency has demonstrated mature AI governance on quadrant-three use cases is inviting the kind of scrutiny that can produce a moratorium on all AI use, including the high-value tools already delivering results.

Building the Phases: A Three-Phase Structure

Most agencies benefit from a three-phase roadmap structure: Foundation, Report Drafting, and Expansion. The specific timelines depend on agency size, existing readiness posture, and procurement timeline, but the sequence is consistent.

Phase One: Foundation

Phase one covers the first six to nine months and focuses on quadrant-one use cases with the parallel infrastructure work that makes phase two possible.

The use-case layer includes AI-assisted BWC redaction for public-records releases, AI-assisted CAD entry summarization, and AI assistance with non-evidentiary documentation. These run with a defined officer and records cohort, active monitoring, and a weekly review of output quality.

The infrastructure layer runs in parallel: completing the Criminal Justice Information Services (CJIS) security review for the AI platforms involved, finalizing the written AI use policy with legal review and prosecutor coordination, standing up the training program that will certify officers for phase two, and conducting the community engagement that establishes a transparent public record of the agency's AI use.

Phase one ends when the agency can demonstrate: measurable results from quadrant-one use cases, a trained and certified officer cohort for phase two, documented prosecutor sign-off on the disclosure standard, and a completed community engagement record. These are not bureaucratic checkboxes. They are the evidentiary base that makes a contested use-of-force case defensible.

Phase Two: Report Drafting

Phase two introduces AI-assisted sworn report drafting to a scoped cohort of officers who have completed the phase-one training program. The scoped cohort is important: starting with officers who volunteered, who have the highest verification discipline, and who carry lighter caseloads than the most senior investigators, produces the cleanest learning environment.

Phase two monitoring is intensive. The agency should review a random sample of AI-assisted reports each week against the footage, looking for gap-fills that the officer corrected, gap-fills that the officer missed, and patterns in the types of errors the AI system produces in this agency's specific operational environment. The monitoring data feeds three outputs: a weekly quality review for the AI program manager, a quarterly report to command staff, and a documented feedback loop to the vendor when platform-level error patterns emerge.

The disclosure standard is active in phase two. Every report drafted with AI assistance carries a disclosure notation. The format should be coordinated in advance with the prosecutor's office, because a disclosure notation that satisfies the prosecutor is evidence of transparency; one that was created without coordination may not satisfy Brady or Giglio requirements on its face.

Phase two also includes the first genuine test of the command culture. Officers in the cohort should be observed, not just surveyed, for their verification behavior. A supervisor who spots an officer skimming the draft rather than running the pass has a choice: discipline the behavior in a way that sends a message, or ignore it in a way that sets a standard. The roadmap must be explicit that the expected behavior is the former. The lesson of the first verification failure in phase two is far cheaper than the lesson of the first verification failure in court.

Phase two expands to the full agency when the scoped cohort has produced at least sixty days of clean monitoring data, when the error rate and correction rate are within acceptable ranges, and when the command staff can articulate the verification standard in the same language the policy uses. These are the tests. An expansion driven by budget pressure rather than monitoring results is a deployment calendar, not a roadmap.

Phase Three: Expansion

Phase three extends AI assistance to new use cases, new units, or deeper integration within existing ones. Investigations summarization, where AI assists in summarizing hours of recorded interviews for case documentation, is a typical phase-three addition. The stakes are high: a fabricated quote attributed to a witness in an AI-generated case summary is a case-ending error. But by phase three, the agency has demonstrated that its officers understand verification, that its command culture supports it, and that its disclosure and documentation practices meet the prosecutor's standard.

Phase three is also when the agency assesses quadrant-four use cases with the governance structure that makes a responsible evaluation possible. With a mature AI program, a functioning governance board, and a track record of transparent operation, the agency is in a position to evaluate high-risk capabilities on their merits rather than in the reactive posture of a chief who deployed without governance and is now defending the program from a community moratorium.

The Procurement Trap: Bundled Contracts and the Roadmap

Deputy Chief Reyes's roadmap had to survive a procurement reality that many command staff encounter: the vendor's preferred contract structure did not match the roadmap's preferred deployment sequence.

The vendor wanted a bundled agreement covering cameras, cloud storage, AI drafting, and analytics on a single seven-year contract. The roadmap called for phase-one tools in year one and phase-two tools only after phase-one conditions were met. The vendor's bundled pricing assumed full deployment of all capabilities within the first year. The roadmap's phased deployment meant paying for capabilities the agency was not yet ready to use.

This is the procurement trap that command staff must understand before signing. Bundled, multi-year, sole-vendor contracts on the order of $45 million and up to ten years in length are in active procurement across the country. The vendor's interest is in activating all capabilities as quickly as possible, because active use drives renewal and upsell. The agency's interest is in deploying capabilities in the sequence that the roadmap, the readiness assessment, and the evidentiary standard require.

The negotiating position is straightforward: the contract activates all capabilities on the agreed timeline, but the agency's deployment of each capability is conditioned on meeting the readmap's phase conditions. The vendor cannot contractually compel the agency to use a tool the agency's own readiness assessment has not yet cleared. That right must be explicit in the contract. If the vendor resists, that resistance is itself data about whether the vendor's interest in the agency's responsible deployment is genuine.

Exit rights also belong in this conversation. A ten-year sole-vendor contract on a technology that is evolving as rapidly as AI represents a genuine lock-in risk. If the vendor's platform introduces a systematic error that the agency cannot mitigate, or if the regulatory environment shifts in a way that makes the current disclosure approach inadequate, the agency needs the contractual ability to respond. A contract that does not include defined exit conditions and data portability rights is a contract that transfers leverage from the agency to the vendor over the life of the agreement.

Governing the Roadmap: Who Owns Each Phase

A roadmap without ownership is a document. Ownership requires a named individual accountable for each phase, a defined review cadence for roadmap progress, and a decision-making structure that can handle the phase-gate questions when they arise.

The phase-gate questions are the moments when the roadmap requires a decision: does the agency advance from phase one to phase two, or does it extend phase one because the monitoring data shows the verification culture is not yet stable? These decisions should be made by command staff on the basis of the monitoring data, not by the vendor on the basis of contract timeline, and not by the chief in isolation without the legal and prosecutor input that makes the decision defensible.

The ownership structure for a mid-sized agency typically includes: an AI program manager who owns day-to-day monitoring and quality review; a command-staff sponsor who owns the roadmap and represents it to the chief; a legal advisor who owns the policy and disclosure questions; a prosecutor liaison who maintains the ongoing relationship and flags any disclosure concerns before they become case challenges; and a community liaison who maintains the engagement record and handles incoming stakeholder questions. None of these is a full-time role in most agencies. All of them are accountability assignments that ensure the roadmap has a human face at each decision point.

Key Takeaways

• A roadmap sequences AI use cases by value and evidentiary risk, not by vendor deployment preference. The sequence determines what the agency can defend, not just what it can activate.
• The value-risk framework places each use case in one of four quadrants. High-value, lower-risk cases (BWC redaction, CAD summarization) belong in phase one. High-value, higher-risk cases (sworn report drafting) belong in phase two, after phase-one conditions are met.
• Officers spend 30 to 40 percent of a shift on paperwork. AI-assisted report drafting returns real time, but only to agencies that have built the verification culture to use the tool safely. The time savings and the accountability standard must be pursued together.
• Phase-gate conditions are not bureaucratic obstacles. They are the evidence base that makes a contested report defensible: trained officers, documented prosecutor sign-off, active monitoring, and a stable error rate.
• Bundled, multi-year, sole-vendor contracts of up to $45 million and ten years are in active procurement. The agency's deployment sequence must be protected contractually, and exit rights and data portability belong in the contract before signing.
• The King County prosecutor's bar on AI-written reports was a readiness and sequencing failure: high-risk capabilities deployed without the governance infrastructure that makes their use defensible. The roadmap is the structural answer to that failure.
• Roadmap ownership requires named individuals at each accountability point: AI program manager, command sponsor, legal advisor, prosecutor liaison, and community liaison. A roadmap without owners is a document without a future.
• Quadrant-four use cases (high risk, lower or contested value) should not appear in early phases. An agency that establishes mature AI governance through phases one and two is in a position to evaluate those capabilities on their merits rather than under crisis conditions.