Sergeant Diane Ortega stood at the briefing-room whiteboard before day shift, a dry-erase marker in one hand and the previous quarter's incident log in the other. Her patrol division had closed 1,847 reports. Of those, AI tools had touched 1,612 in some form: drafting, transcribing, summarizing, flagging. Her lieutenant had asked one deceptively simple question, "We bought the tools, so where in our process do they actually go?", and she could not answer it. So she started drawing boxes, left to right, the whole life of an incident: the call comes in, units respond, the scene is worked, the report is written, the records move, the case closes. Then she uncapped a second marker, a red one, because the real work was not drawing the boxes. It was deciding which steps AI was allowed to touch, which steps demanded a human to verify and sign, and which steps AI must never reach. The workflow existed. The AI was already active inside it. But nobody had drawn that line.

The Call-to-Closure Lifecycle

Every incident in a public-safety agency follows a lifecycle. A call arrives. Units respond. The scene is worked. Evidence is gathered. Reports are filed. The case is classified. The record is closed or forwarded to the next stage: follow-up investigation, prosecution, records release. That lifecycle is the same whether the call is a noise complaint closed in forty minutes or a homicide investigation that runs for eighteen months. The structure does not change. The complexity within each step changes.

What mapping a call-to-closure process means, in the context of AI, is drawing that lifecycle explicitly: identifying each step, naming who owns it, and deciding for each step whether AI assistance is appropriate, limited, or prohibited. This is not an abstract policy exercise. It is a practical workflow document that an officer, a supervisor, a prosecutor, and an oversight board can all read and understand. It is the difference between "we use AI in our reporting" and "here is exactly where AI operates, what it produces, who reviews it, and where the human decision is required and documented."

The reason this matters is not primarily about efficiency. It is about accountability. When a defense attorney files a suppression motion and asks which steps in the investigation and documentation process used AI tools, the agency needs a specific, accurate answer. "We use AI for reporting" is not a specific answer. A mapped process with named steps, named AI tools, and named human review points is a specific answer. It is also a better answer. It demonstrates that the agency thought about where AI belongs and where it does not, rather than letting the technology drift into the workflow without deliberate design.

Five Phases of the Lifecycle

For the purposes of this lesson, the call-to-closure lifecycle has five phases. These phases apply broadly across patrol, dispatch, and investigations. Individual agencies will have their own terminology and their own step sequences, but the underlying structure is consistent.

Phase one is intake and dispatch. This is the moment a call enters the system: the 911 call, the radio notification, the officer-initiated contact. The computer-aided dispatch system (CAD, the software platform that receives, logs, and assigns calls) creates the initial record. Triage decisions are made. Priority is set. Units are assigned.

Phase two is on-scene response. Units arrive. The scene is assessed. Actions are taken: interviews, searches, use of force if necessary, medical response if necessary, evidence collection if applicable. The body-worn camera (BWC) is recording. Officers may be narrating. The CAD entry is being updated as the scene develops.

Phase three is documentation. This is the report-writing phase. The officer returns to the station or uses a mobile terminal to draft the incident narrative, the supplemental reports, the use-of-force forms if applicable, and any other required documentation. This is the phase where AI-assisted drafting tools like Axon Draft One operate: drafting report narratives from the BWC audio, which testing officers reported reduced report-writing time by 82% in testing benchmarks.

Phase four is review and adoption. The draft, whether AI-generated or hand-typed, goes through supervisory review. The officer adopts it as their sworn account. Any corrections are made. Disclosure annotations are added if the report was AI-assisted. The report is finalized and entered into the records management system (RMS, the database where case records are stored and managed).

Phase five is records management and closure. The RMS record is complete. The case is classified, cleared, or referred to the next step. Public-records requests are handled. Discovery packages are assembled. Evidence is retained or returned according to agency policy. The lifecycle closes, or a new one begins if the case advances.

AI-Ready Versus Human-Only: A Real Distinction

The first and most important decision in mapping a call-to-closure process is categorizing each step as AI-ready, AI-assisted with required human review, or human-only. These are not just policy labels. They are accountability structures. Getting the categorization right is the work.

An AI-ready step is one where AI assistance can operate with high confidence and low legal consequence from an error. Transcription of a recorded interview is a strong candidate: the AI transcribes the audio, the human reviews the transcript for accuracy, and an error in the transcript is caught in review without causing downstream harm. Generating a chronological index of timestamps from BWC footage is another: the AI lists what occurred when, the human confirms the sequence, and the worst-case error from the AI is a missed timestamp that the human review catches. These steps benefit from AI speed and scale without carrying the consequence that comes from a factual error in a sworn document.

An AI-assisted-with-required-review step is one where AI assistance provides significant value but where the output is high-stakes enough that human review is not optional, it is mandatory and documented. Report drafting is the canonical example. Axon Draft One produces a narrative from BWC audio. That narrative is evidence. It will be disclosed to the defense under Brady v. Maryland (the 1963 Supreme Court decision requiring prosecutors to disclose exculpatory evidence to the defense) and Giglio v. United States (the 1972 decision extending disclosure to include impeachment evidence about witnesses, including the testifying officer). The AI draft is not the sworn report. It becomes the sworn report only after the officer reviews it, verifies every factual claim against the footage, corrects any errors, and adopts it as their own account. The review is not a nicety. It is the event that transforms a machine-generated text into sworn testimony.

A human-only step is one where AI may not participate in the decision, even in an advisory capacity, because the consequence of an error is too severe, the legal accountability is too direct, or the judgment required is irreducibly human. Setting call priority in dispatch is a human-only decision: a telecommunicator classifying a call as a priority 3 when it is a priority 1 because AI suggested 3 is an error that can cost a life. The arrest decision is human-only: no AI tool may decide whether to place someone under arrest, because that decision carries Fourth Amendment implications, requires probable cause assessment grounded in the officer's training and the totality of the circumstances, and the accountability for it is the officer's, sworn and personal. The decision to use force is human-only. The decision to charge is human-only. The decision to close a case without charges is human-only. These decisions may be informed by AI-generated analysis. They may not be made by AI.

Draw the line before the tool is deployed, not after it has produced something that ends up in a suppression motion.

Mapping the Process in Practice

A process map for a call-to-closure workflow does not have to be complicated. It has to be specific. The following structure works for most patrol agencies and can be adapted for dispatch-heavy environments or investigation-heavy environments.

Step One: List Every Step in the Workflow

Start with what actually happens, not with what the policy manual says should happen. Walk through a typical incident from the moment the call arrives in the CAD to the moment the record is closed in the RMS. Write down every discrete step. Include the steps that feel obvious: "officer activates BWC," "officer arrives on scene," "officer completes report," "supervisor reviews report." Include the steps that are easy to skip in a description but matter: "officer reviews AI draft against footage," "disclosure annotation added," "report adopted by officer as sworn account."

A typical patrol incident workflow has between fifteen and thirty discrete steps when described at this level of granularity. That may feel like a lot. It is not. Each step is where something happens, and each step is where an error can enter the record. Knowing which steps exist is the prerequisite for deciding which of them AI can touch.

Step Two: For Each Step, Decide the Category

Once the steps are listed, categorize each one: AI-ready, AI-assisted-with-required-review, or human-only. Apply these tests:

First, what is the worst-case consequence of an AI error at this step? If the answer involves evidence admissibility, a constitutional right, a life-safety decision, or a sworn account, the step requires either required-review categorization or human-only status. If the answer involves administrative inefficiency or a correctable record error, AI-ready may be appropriate.

Second, who is legally accountable for the output of this step? If the output is evidence, the accountable party is the officer who adopts it as their sworn account. That officer needs meaningful review authority, which means AI-assisted-with-required-review at minimum, with a documented and logged review. If the output is a discretionary decision (priority, arrest, charge), it is human-only: the accountability cannot be delegated to a machine.

Third, can the error be caught before it causes harm? Transcription errors can be caught in review. A misclassified call type may not be caught until units arrive and find a different scene than the call described. The catchability of an error determines how much risk the AI-ready or AI-assisted categorization carries.

Step Three: Name the Tools and the Review Points

Once the categorization is done, name the specific AI tools in use at each AI-ready or AI-assisted step. "AI-assisted drafting: Axon Draft One, BWC audio to narrative draft." "AI-assisted transcription: [platform name], recorded interview audio to text transcript." "AI-assisted call classification support: [CAD AI module name], classification suggestion reviewed and confirmed by telecommunicator before entry."

Naming the tools matters for three reasons. First, it creates an audit trail: if a tool produces an error that ends up in a report, the agency can trace it. Second, it enables vendor accountability: if a specific tool is generating systematic errors (softened use-of-force language, gap-fill scene descriptions, misclassified call types), the pattern can be identified and addressed. Third, it satisfies the disclosure obligation that the Electronic Frontier Foundation (EFF) and other transparency advocates have raised: the public and the defense have a legitimate interest in knowing which AI tools were used in a case, not just that "AI was involved."

At each AI-assisted step, name the review point: who reviews, what they check, and how the review is documented. "Officer reviews AI draft against BWC footage, CAD entry, and field notes. Officer corrects any unverified claims. Officer adds disclosure annotation. Officer adopts report as sworn account. Supervisory review confirms disclosure annotation is present." That is a specific, auditable process. It is also a defensible one.

Where the Map Breaks Down: Common Failure Modes

Agencies that have mapped their call-to-closure processes have encountered consistent failure modes. Understanding them before building the map is more efficient than discovering them after deployment.

The Drift Problem

The drift problem occurs when AI assistance gradually expands beyond its designated scope without a deliberate decision to expand it. An agency maps its process, designates report drafting as AI-assisted-with-required-review, and deploys. Six months later, under pressure to clear a report backlog, officers are adopting AI drafts with a lighter review pass than the process specifies. A year later, the review pass has become perfunctory on routine calls. The process map says required-review. The actual practice is closer to AI-ready without the formal categorization. The map and the practice have drifted.

The drift problem is not a technology problem. It is a supervision and governance problem. The fix is audit: periodic review of a sample of AI-assisted reports against the footage record to confirm that the review pass is being performed as specified. This is the same quality-assurance function that supervisors apply to hand-typed reports, applied to AI-assisted ones. If the review reveals systematic drift, the agency has a training and supervision gap to close, not a technology problem to solve by tightening the AI's parameters.

The cost of undetected drift is significant. Officers spend 30 to 40% of a shift on paperwork. If AI-assisted drafting saves a substantial portion of that time, the time saving is real and valuable. But if the verification pass that justifies the AI-assisted categorization is being skipped, the agency is not running an AI-assisted-with-required-review workflow. It is running an AI-decides workflow, and the legal accountability for a sworn report has not changed because the review was skipped. The Brady and Giglio obligations do not become optional because the shift was busy.

The Handoff Gap

The handoff gap occurs at the boundaries between phases. The classic example is the boundary between phase three (documentation) and phase four (review and adoption). The AI draft exists. The officer has it. The question is: at what moment does the officer's review and correction of the draft happen, and is that moment formally logged? If the workflow does not specify when the verification pass occurs, it may not occur at all, or it may occur in a rushed, incomplete form at the moment of submission when the officer is moving to the next call.

The handoff gap is also common at the boundary between phase four (review and adoption) and phase five (records management and closure). The report is adopted. It goes into the RMS. The disclosure annotation specifying that AI was used and that the officer verified and adopted the draft is supposed to accompany it. If the RMS workflow does not have a mandatory field for disclosure annotation, the annotation may be forgotten. If it is forgotten, the agency is not meeting the disclosure standard that the King County (Washington State) prosecutor's office applied when it barred AI-written police reports: the concern was precisely that AI involvement was not being disclosed, tracked, or annotated in the record.

The Scope Creep Problem

Scope creep is the inverse of drift. Where drift is the review becoming less rigorous than specified, scope creep is AI assistance being applied to steps that were categorized as human-only. An AI module that was deployed to assist with report drafting starts being used to suggest call priorities. A summarization tool deployed for case documentation starts being used to summarize interview content in ways that inform charge recommendations. The tool was approved for one step. It is being used at a different step with a different risk profile.

Scope creep is often driven by vendor upsell or by genuine user convenience. The tool is present, it produces useful-looking output, and the line between "assisting" and "deciding" is not always obvious in practice. The map is the instrument that makes the line obvious. When a supervisor sees that a call-priority suggestion from an AI module is being entered into the CAD without human review, the map provides the reference: "call priority: human-only." The deviation is named. The correction is specific.

The Map as a Disclosure Document

A well-built call-to-closure process map is not just an internal management tool. It is a disclosure document. When a defense attorney, a prosecutor, a civilian oversight board, or a journalist asks "what AI tools were used in this case and how," the answer is the map plus the case-specific log.

The King County, Washington prosecutor who barred AI-written police reports was responding to a disclosure gap: AI was being used, and the use was not being documented or disclosed consistently. The EFF's transparency concerns are the same gap from the civil-liberties direction: the public has a legitimate interest in knowing when AI tools are participating in law enforcement processes that produce evidence used against defendants. A process map that names the tools, the steps, and the review requirements is a direct response to both concerns. It does not eliminate the concerns. It gives a documented, auditable answer.

The CJIS (Criminal Justice Information Services) Security Policy, which governs how criminal justice data must be stored, transmitted, and accessed, establishes obligations that stay with the agency regardless of what vendor tools are in use. If an AI vendor processes BWC audio or CAD data, the agency's CJIS obligations apply to that processing. The process map should include a CJIS compliance checkpoint at any step where an AI tool touches data governed by CJIS: who confirmed that the tool's data handling meets the CJIS Security Policy, and when was that confirmed.

Agencies signing bundled multi-year contracts with sole-vendor AI providers (contracts that have reached approximately $45 million and terms of up to ten years in some jurisdictions) need this map even more urgently than agencies using individual tools. A long-term sole-vendor contract creates a situation where the vendor's full ecosystem of tools, cameras, drones, AI reporting, cloud storage, AI analytics, may be active across multiple workflow steps simultaneously. Without a map, the agency cannot tell which steps are AI-touched, cannot audit the review process, and cannot produce a disclosure document that accurately describes AI involvement in a specific case.

Reading the Map Back to the Room

When Sergeant Ortega finished her board, she walked her lieutenant through it the way she would later walk a prosecutor or an oversight board. She did not start with the tools. She started with the human-only boxes, marked red: "These are the decisions AI never makes, because the law and the Constitution put them in a person's hands, call priority, arrest, use of force, charging, case closure." Then the AI-assisted boxes: "These are where the tools save us time, but a named human reviews, corrects, signs, and the review is logged, here is exactly how." Then, last and briefly, the AI-ready boxes: "These are the routine, low-consequence steps where the tools just run with a light check." She closed on the disclosure log, because that was the part that let her give any questioner the same true answer about exactly which tool touched which step and which human owned the result. That walk, from red to assisted to ready to the audit trail, is the deliverable. The map is the skill.

Key Takeaways

• A call-to-closure process map names every step in the incident lifecycle, categorizes each step as AI-ready, AI-assisted-with-required-review, or human-only, and identifies the specific tools and review requirements at each step. Without this map, AI drifts into the workflow without accountability.
• The five lifecycle phases are intake and dispatch, on-scene response, documentation, review and adoption, and records management and closure. AI tools operate differently at each phase, and the risk profile of an error differs at each phase.
• AI-ready steps are those with low legal consequence from an error and strong catchability before harm occurs. AI-assisted-with-required-review steps, like report drafting, require a documented human review before the output becomes a sworn account. Human-only steps include call priority, arrest, use-of-force, and charging decisions.
• The drift problem occurs when the review pass becomes less rigorous than the map specifies, turning an AI-assisted-with-required-review step into an effective AI-decides step. Periodic audit of a sample of AI-assisted reports against the footage record is the corrective.
• The handoff gap occurs at phase boundaries, most critically between draft and adoption and between adoption and RMS entry with disclosure annotation. Mandatory fields and logged review points close the gap.
• Scope creep, when AI assistance expands to steps categorized as human-only, is named and corrected by reference to the map. The map is what makes the line visible.
• The call-to-closure process map is a disclosure document as well as an internal management tool. It is the specific, auditable answer to questions from defense counsel, prosecutors, oversight boards, and the public about where AI operated in a given case.
• CJIS Security Policy obligations and Brady and Giglio disclosure obligations stay with the agency regardless of vendor involvement. The process map is how the agency demonstrates it has operationalized those obligations across the workflow, not just acknowledged them in policy.