Deputy Chief Marcus Tran pulled up the slide at the all-hands and drew a line down the middle. On the left he wrote: "AI does this." On the right he wrote: "You do this." Officers had been using Axon Draft One, which generates police report narratives from body-worn camera (BWC, the recording device an officer clips to their uniform) audio, for three months. Report-writing time had dropped, roughly consistent with the 82 percent reduction Axon reported in testing. The problem was nobody had agreed on where the line was. Some officers were reading the draft once and submitting. Some were running it back against footage paragraph by paragraph. Some were adding it to the computer-aided dispatch (CAD, the system tracking calls and units) entry without reviewing it at all. Tran did not call it a training failure. He called it what it was: a workflow design failure. The agency had handed officers a powerful tool and left them to guess at the operating model. Some guessed right. Some guessed in ways that were going to produce a Brady v. Maryland problem, the constitutional obligation requiring prosecution disclosure of evidence favorable to the defendant, before the year was out.

Why Workflow Design Is an Accountability Decision

The word "workflow" sounds administrative. It sounds like process improvement, like something that belongs on a consultant's slideshow with color-coded swimlanes. In public safety AI, workflow design is an accountability decision, because the workflow is the mechanism that determines whether a human being reviewed, verified, and adopted an AI output before it became a sworn account, a dispatch priority, or an official record. A workflow that contains no human checkpoint between AI output and official submission is a workflow that has removed accountability from the loop. That is not an efficiency gain. It is an evidentiary risk with a specific legal name.

Every AI workflow in a public-safety context has a judgment boundary: a specific point at which the AI's output is either accepted by a human and adopted as that human's professional product, or flagged, corrected, and then adopted. The placement of that boundary is the central design decision. It cannot be left to individual officers to determine organically. It must be designed explicitly, documented in policy, trained into practice, and reviewed periodically to ensure the practice matches the policy.

This lesson is about how to make those design decisions. It is addressed to agency executives, HR and organizational design leads, and supervisors who are responsible for how AI tools integrate into daily operations. The goal is a workflow design where humans remain on judgment, where AI handles the throughput tasks it is well suited for, and where the accountability for every official product is traceable to a specific human decision.

The workflow is where policy becomes practice. A verification standard that exists in a written policy but is not built into the workflow is a standard that will be followed inconsistently, which is another way of saying it will not be followed when the pressure is highest.

Mapping the AI-Ready and Human-Only Steps

The first task in workflow redesign is a clean map of which steps in each process are appropriate for AI assistance and which are not. This mapping exercise should be done for each AI use case the agency is deploying: report drafting, dispatch support, records redaction, case summarization, and any others. The mapping should be done by people who actually do the work, not by executives who observe it, because the judgment boundary is visible in the operational detail, not in the description of the process at a high level.

The Report Drafting Workflow

In the report-drafting workflow, the AI-ready steps are clear: the tool ingests the BWC audio, transcribes it, and generates a draft narrative. It is doing what it is well suited for, producing a structured text from an audio source at a speed no human can match. The 82 percent reduction in report-writing time that testing with Axon Draft One produced comes from this step. Officers who spent 30 to 40 percent of a shift on paperwork get that time back when the draft is generated automatically.

The human-only steps begin the moment the draft exists. The officer's footage-grounded verification pass (checking every factual claim in the draft against the BWC recording and the CAD entry at the relevant timestamp) is not AI-assisted. It is human judgment applied to an AI product. The officer's corrections to the draft are human decisions. The officer's adoption of the draft as their sworn account, which means signing or certifying the report as their own professional record, is explicitly a human act. No workflow design can automate that step without removing the accountability that makes the report evidentiary.

The workflow map for report drafting should show exactly these steps in sequence: BWC records incident (automated), officer completes field notes at scene (human), AI draft is generated from BWC audio (automated), officer runs footage-grounded verification pass (human), officer makes corrections and notes discrepancies (human), officer adopts and submits report with AI disclosure notation (human), supervisor spot-check for disclosure notation and verification documentation (human), records management system (RMS, the agency's case documentation platform) archives both the AI draft and the adopted report (automated). Each human step should have a named time standard (not "as soon as possible," but a specific window within which the step must be completed) and a named accountability (not "the officer," but the specific supervisory chain that will check the step).

The Dispatch Support Workflow

In dispatch, the AI-ready and human-only distinction is more urgent because the stakes of a misclassification are immediate. In report writing, an unverified gap-fill is an evidentiary problem for a future case. In dispatch, a misclassified call is a response problem for a current emergency. The workflow must reflect that asymmetry.

AI-ready steps in dispatch include: transcribing incoming call audio in real time, suggesting a call type from the transcript, surfacing relevant prior-call information for the address, and generating the initial CAD entry text. Human-only steps include: assigning the final call priority (the telecommunicator owns this decision, not the AI), dispatching units (human decision), escalating a call from one priority to another based on caller behavior or new information (human decision), and making any decision where a misclassification would risk delayed response to a life-safety emergency.

The workflow design for dispatch must include a specific protocol for situations where the AI's suggested call type conflicts with the telecommunicator's assessment of the caller's behavior. The AI reads the words. The telecommunicator hears the voice, the background, the fear or calm in the caller's tone. Those are human-judgment inputs the AI does not have, and the workflow must make clear that the telecommunicator's overriding judgment is not just permitted but expected when the inputs conflict.

The Records and Redaction Workflow

Records redaction is the use case where AI assistance most nearly approaches full automation in practice, and therefore the use case where workflow design must be most explicit about where the human checkpoint sits. AI-assisted redaction tools can process BWC footage and case documents at a speed and scale impossible for a human reviewer working alone. They can flag faces for blurring, license plates for redaction, and protected health information for removal across thousands of hours of footage. That speed is the value. The risk is the missed face, the unblurred plate on a witness in a domestic violence case, the health information that passes through a release because the AI classified it as non-protected. Under open-records statutes (public-records laws that require agencies to release certain materials in response to requests, while protecting certain classes of information from release), an under-redaction failure is a privacy violation and potentially a lawsuit. An over-redaction failure withholds public information and may trigger a legal challenge of its own.

The workflow must specify that AI-assisted redaction is a first pass, not a final product. A trained human reviewer checks the AI's output before release. The review should be risk-stratified: footage from use-of-force incidents, footage involving minors, footage involving protected witness information, and footage from sensitive investigations gets a more intensive human review than routine patrol footage. The workflow should define those categories explicitly, not leave them to the reviewer's judgment in the moment.

Designing the Explicit Human Checkpoint

The human checkpoint is the specific moment in the workflow where a human being reviews, approves, or adopts the AI's output. Designing that checkpoint well is the most important single decision in workflow redesign, because the checkpoint is where accountability is either preserved or lost.

What Makes a Checkpoint Real

A checkpoint is real when three conditions are met. First, it is a required step, not an optional one. The workflow does not allow the AI's output to proceed to official submission without passing through the checkpoint. The RMS or the evidence platform should enforce this technically if possible: the report cannot be submitted without the officer's verification certification, the CAD entry cannot be finalized without the telecommunicator's priority confirmation. Where technical enforcement is not possible, supervisory enforcement must substitute.

Second, the checkpoint is documented. Not in the officer's memory, not in a verbal understanding with a supervisor, but in the record. The RMS entry for every AI-assisted report should show: that AI assistance was used, the date and time the officer's verification pass was completed, any corrections made, and the officer's adoption of the final report. That documentation is what allows the agency to answer a Brady disclosure request, a defense discovery motion, or an oversight audit with actual data rather than assertions.

Third, the checkpoint is meaningful. A checkbox that says "I reviewed the AI output" is not a checkpoint. It is a liability waiver. A meaningful checkpoint requires the officer to do something that demonstrates review: time-stamped access to the footage during the verification window, a logged comparison of the AI draft against the CAD entry, documented corrections when discrepancies are found. The mechanical trace of a real verification pass is different from the mechanical trace of someone clicking through a required screen. The workflow should be designed to capture the former, not accommodate the latter.

Supervisor Spot-Checks as Workflow Infrastructure

Supervisor spot-checks are not an add-on to a well-designed workflow. They are infrastructure. They are the mechanism that tells the organization whether the checkpoint is actually functioning as designed, or whether it has become a box-checking exercise. They are also the signal to the officer that the checkpoint matters, that someone in the chain of command will actually look at whether the verification pass was real.

A spot-check protocol should specify a percentage of AI-assisted reports reviewed per unit per month (a realistic number is 10 to 15 percent, enough to establish a pattern without consuming supervisor bandwidth for its own sake), the specific elements the supervisor checks (disclosure notation present? Verification documentation present? Any corrections made and documented? Any significant discrepancy between the AI draft and the adopted report?), and the escalation path when a report fails the spot-check. Failure in a spot-check is not automatically a disciplinary matter. It is first a training and workflow signal. If multiple officers in a unit are failing spot-checks in the same way, the workflow has a design problem, not a personnel problem, and the AI Lead and supervisor need to address it as such.

The Time Standard for the Verification Pass

One of the most consequential workflow decisions an agency makes is the time standard for the verification pass. How long after a call does the officer have to complete the verification and submit the AI-assisted report? The temptation is to set a long window to give officers flexibility. The operational reality is that a longer window means the officer is reviewing footage and correcting a draft hours or days after the incident, when memory has faded and the footage is doing more of the verification work without the supplement of the officer's contemporaneous recall.

The best practice is to complete the verification pass and adopt the report at the end of the same shift or within 24 hours of the incident, whichever is shorter for the call type. Use-of-force incidents should have the shortest window, because they carry the highest evidentiary stakes and because supervisory review of use-of-force reports is typically required within a short timeframe by department policy and sometimes by law. The time standard should be written into the workflow policy, not left as a general expectation.

There is a counterintuitive point worth making about time standards and AI assistance. One concern officers and supervisors sometimes raise is that AI assistance creates pressure to submit reports faster, which may shorten the verification window. The discipline here is sequential: AI assistance speeds the drafting step, not the verification step. The 82 percent reduction in report-writing time that Axon Draft One testing produced is a reduction in the drafting time, from blank page to draft. The verification pass takes the time it takes, and that time does not compress as a result of faster drafting. What compresses is the total time from incident to submission, because the drafting step is faster. The verification step is not faster, and it should not be designed as though it is.

Accountability Traps to Design Out

Several patterns in AI workflow adoption create accountability problems. They are common enough that they warrant explicit attention in any workflow redesign effort. The goal is to design the workflow so these patterns cannot emerge organically.

The Blank Adoption Trap

The blank adoption trap is the pattern where an officer reads through the AI draft, finds nothing obviously wrong, and certifies it as their sworn account without cross-referencing the footage. This is the most common accountability failure in AI-assisted report writing, and it is also the least visible, because the adopted report looks right. The gap-fill that the footage would have caught is not a flagrant error. It is a plausible description of a plausible event. It is only wrong when compared to the footage, and if nobody compared it to the footage, the comparison never happens.

The blank adoption trap cannot be designed out by telling officers not to do it. It has to be designed out structurally. The workflow must require a specific, logged action that demonstrates the footage comparison was made. The RMS should require the officer to enter the timestamp of the footage reviewed before the verification certification is accepted. That is not foolproof, but it is a structural friction that distinguishes between officers who made the comparison and officers who clicked through the screen.

The Supervisor Approval Substitution Trap

In some departments, supervisor approval of a report functions as a substitute for the officer's verification pass, in the sense that supervisors are checking content quality rather than independently verifying AI-specific disclosure and verification documentation. This substitution defeats the purpose of both steps. The supervisor's approval is not a fresh footage-grounded verification of every factual claim. The supervisor was not at the scene. They cannot verify the footage against the draft. What they can do is check whether the officer's verification process is documented, whether the disclosure notation is present, and whether significant discrepancies were flagged. That check requires supervisors to know what they are looking for, which requires training specifically calibrated to AI-assisted reports.

The workflow should make explicit what supervisors are checking in their approval of AI-assisted reports, and that list should be different from (and supplementary to) what supervisors check in manual reports. The supervisor is not re-doing the officer's verification pass. They are verifying that the officer's verification pass was documented, complete, and real.

The Tool Upgrade Gap

AI tools change. Vendors update models, change output formats, modify the tool's behavior in ways that may affect the types and frequency of errors the tool produces. A workflow designed around one version of a tool may not be adequate for the next version. The governance triangle (the Agency AI Lead, the Disclosure Coordinator, and the Review Cadre) must include a defined process for assessing workflow implications whenever a vendor updates the tool. That process does not need to be elaborate, but it must exist. At minimum, the AI Lead should receive advance notice of significant tool updates, the Review Cadre should review a sample of outputs from the first month after an update, and the Disclosure Coordinator should be briefed on any changes that affect the disclosure documentation.

The bundled, multi-year contracts now common in public-safety AI procurement, some running $45 million or more over ten years, make the tool-upgrade gap a genuine risk. An agency that signs a ten-year contract with a sole vendor is in a relationship where the tool will change significantly over the contract term. Building the workflow-review trigger into the contract is the right approach: require the vendor to provide advance notice of model updates, and require the agency's AI governance structure to review the workflow implications before the update goes live in production.

Making the Redesigned Workflow Stick

A workflow redesign that exists on paper but is not practiced in the field is not a workflow redesign. It is a document. Making the redesign stick requires three things: training that connects the workflow design to the specific consequences of not following it; supervision that enforces the workflow in real time rather than after-the-fact; and feedback loops that update the workflow when practice reveals design gaps.

Training for Workflow, Not Just Tool Use

Most AI tool training in law enforcement focuses on tool use: how to generate a draft, how to navigate the platform, how to submit the output. That is necessary but not sufficient. Workflow training teaches something different: what the tool is authorized to do in this workflow, where the human checkpoint is, what a real verification pass looks like compared to a blank adoption, and what the consequence is, in terms of Brady exposure and Giglio implications, when the checkpoint is not honored.

The most effective format for workflow training in this context is the case study: a scenario where a real or hypothetical gap-fill was not caught, and the officer is walked through what happened in the case, how the verification pass would have caught it, and what the evidentiary consequence was. The scenario works at a departmental all-hands, at roll call, and as a written module in the agency's training record management system. It connects the abstract workflow design to the concrete stakes that motivate real practice.

Building AI Disclosure into the RMS

The records management system is a powerful workflow enforcement tool if it is configured correctly. The RMS should have a field for AI-assistance flag, required for any report that used AI drafting. That field, once populated, should trigger an automatic disclosure notation in the report's header and an entry in the AI use log that the Disclosure Coordinator maintains. Where the RMS cannot enforce the disclosure flag technically, the supervisor spot-check process is the substitute. But every agency should be working with their RMS vendor to build the AI-assistance flag into the system as a required field, not a voluntary one.

The CJIS Security Policy governs how criminal justice information is stored, transmitted, and accessed. CJIS obligations stay with the agency, not the vendor. The AI use log, the AI draft archive (if the agency is preserving original AI drafts alongside adopted reports), and the verification documentation are criminal justice records subject to CJIS security standards. The workflow should specify how these records are stored, who has access to them, and how long they are retained. Those decisions should be made before a defense attorney asks for them in discovery, not after.

The Quarterly Workflow Review

Every workflow redesign needs a scheduled review cadence. Practice reveals design gaps that no desk-based redesign exercise will catch: the edge case the workflow did not account for, the call type where the verification step is consistently skipped because the time standard is not realistic, the RMS configuration that makes the disclosure flag harder to use than to ignore. The Agency AI Lead should convene the Review Cadre and the Disclosure Coordinator for a quarterly workflow review, using the cadre's error data and the Disclosure Coordinator's incident log as the primary inputs.

The output of the quarterly review is a workflow update memo: a brief, specific list of changes to the workflow design based on what the data showed. That memo goes to command for approval and then to training for incorporation into the next training cycle. The workflow is not a fixed document. It is a living operating model that improves with operational experience, as long as the improvement loop is formalized and functions as designed.

Key Takeaways

• Workflow design is an accountability decision, not just a process improvement exercise. The placement of the human judgment boundary in each AI-assisted workflow determines whether accountability is preserved or lost when the AI's output becomes an official record.
• Every AI workflow in public safety must have a defined AI-ready section and a defined human-only section. Report drafting, dispatch support, and records redaction each require explicit mapping of which steps the AI performs and which steps a human must perform and document.
• The footage-grounded verification pass is a human-only step in the report-drafting workflow. It does not compress because drafting is faster. The 82 percent reduction in report-writing time from Axon Draft One testing applies to the drafting step, not the verification step.
• A human checkpoint is real when it is required, documented, and meaningful. A checkbox is not a checkpoint. The workflow should require a logged action that demonstrates the verification pass was made, not merely that a screen was clicked through.
• Three accountability traps must be designed out: blank adoption (reading the draft without footage comparison), supervisor approval substitution (treating supervisor sign-off as a substitute for officer verification), and the tool upgrade gap (failing to review workflow implications when the vendor updates the AI model).
• Supervisor spot-checks are workflow infrastructure, not optional oversight. A sample of AI-assisted reports should be reviewed by a supervisor each month, with a defined checklist and a documented escalation path when reports fail the check.
• The RMS should enforce AI disclosure technically through a required AI-assistance flag, not rely on officer memory or voluntary compliance. CJIS obligations cover the AI use log, the draft archive, and the verification documentation: those are criminal-justice records, not administrative ones.
• Workflow redesigns require quarterly review cadences, using error data from the Review Cadre and incident data from the Disclosure Coordinator. A workflow that does not improve from operational experience is a workflow that is creating undetected problems.