Chief Sandra Morales stepped to the podium at the National Chiefs' Conference in March 2026 with a single slide behind her: "Three years ago we deployed AI-assisted report writing. Today, 94 percent of our patrol officers use it on every shift, error-rate complaints to the prosecutor's office are down 31 percent, and the King County-style objection has not once been sustained against us." The room was quiet. Her agency had not just adopted a tool. It had built an operating model, and the difference between those two things was, she would tell anyone who asked, everything.

From Pilot to Operating Model

Every meaningful technology transformation in public safety begins with a pilot. A small cohort, a single use case, a finite window of evaluation. The pilot is not the problem. The problem is what happens when the pilot ends and nobody has a plan for what comes next. The agency that ran a three-month AI-assisted report-writing pilot in 2024 and then declared it a success and rolled it out department-wide without changing a single workflow, policy, or training curriculum did not transform. It created exposure at scale.

Transforming an agency with AI is a multi-year arc, and the arc has a recognizable shape: assess, pilot, validate, build the operating model, and then scale within that model. Each stage has specific outputs, and the outputs of one stage are the inputs to the next. Skipping a stage does not save time. It pushes the consequences forward to the stage where they are most expensive to resolve, usually in a courtroom, a city council chamber, or a community meeting where trust has already been damaged.

The operating model is what the agency is actually running when AI is no longer an experiment but an integrated part of daily operations. It includes policy (who can use what tool, under what conditions, with what review requirement), process (the specific workflow steps from tool engagement to adopted output), technology infrastructure (the platforms, integrations, and data governance), people (the trained workforce, the designated accountability holders, the oversight roles), and measurement (the metrics that tell leadership whether the program is working on every axis that matters).

An agency without an operating model has pilots that never stop being pilots, tools that officers use in different ways depending on who trained them, and a governance structure that cannot withstand a serious review because nothing has been formally designed. The operating model is the answer to the question "How does this agency use AI responsibly and at scale?" If the chief cannot answer that question in two minutes, the operating model does not yet exist.

The Six Elements of a Public Safety AI Operating Model

A complete operating model for AI in public safety rests on six elements. All six must be present before the agency can claim it is operating rather than experimenting.

The first element is a written policy that covers every approved use case: what AI tools are authorized, for what purposes, by whom, under what conditions, and with what review and disclosure requirements. Axon Draft One, the product that drafts police report narratives from body-worn camera (BWC, the recording device worn on an officer's uniform) audio and has been reported to reduce report-writing time by 82 percent in officer testing, is a tool. The policy that governs how officers use it, what they must verify before submitting, and how they must disclose its use to prosecutors is the operating model element that makes the tool defensible.

The second element is a trained and competency-verified workforce. Every officer who touches an AI-assisted workflow must be able to articulate what the tool does, what it can get wrong, how to verify the output, and how to disclose AI involvement. A training completion log is not the same as a competency-verified workforce. The competency question is: can this officer describe the verification standard they applied to this specific draft, in a deposition, today?

The third element is a designated accountability structure, meaning specific named roles with specific responsibilities for the AI program. This includes an agency AI lead who owns the program, a disclosure coordinator who tracks disclosure obligations across cases, and a review cadre across patrol, dispatch, and records who maintain the quality standard. Without named owners, accountability is diffuse, and diffuse accountability is no accountability.

The fourth element is an integrated audit trail. Every AI-assisted output must be logged: when the tool was used, what it produced, what the officer reviewed and changed, and what was ultimately submitted. The Criminal Justice Information Services (CJIS) Security Policy obligations, which govern how criminal justice information is handled, stored, and transmitted, stay with the agency, not the vendor. An integrated audit trail is how the agency demonstrates CJIS compliance and how it responds to a subpoena for the original AI draft.

The fifth element is active prosecutor and oversight alignment. The King County, Washington, prosecutor's decision to bar AI-written police reports from their office is the governance line every agency must address proactively. Meeting that line means the agency has spoken with its prosecutors before deployment, not after a case is challenged. It means the disclosure format the agency uses is one the prosecutor's office has reviewed and accepted. It means the verification standard the agency enforces is one the prosecutor's office would call defensible.

The sixth element is a measurement system that tracks efficiency, integrity, and trust together. No single metric tells the full story. An agency that tracks only report-completion time misses the integrity question. An agency that tracks only error rates misses the trust question. The complete operating model produces a dashboard that shows all three axes to leadership on a regular cadence, so the program is managed on its full performance rather than its most flattering dimension.

The difference between a pilot and an operating model is not scale. It is accountability. An operating model tells you who is responsible for what, right down to the name on the door.

The Arc: Assess, Pilot, Validate, Scale

The multi-year transformation arc is not a rigid timeline, but it has a logic that experienced agencies respect. Moving too fast through any stage produces the specific failure mode of that stage. Moving too slow produces cost and drift. Understanding what each stage produces, and what failure looks like at each stage, is the executive skill that drives the arc.

The Assessment Stage

Assessment determines where the agency is before anything is deployed. It answers four questions. First, what is the agency's current evidentiary posture: what does the policy governing AI use currently say, and if there is no policy, what does the absence mean? Second, what is the technology infrastructure: what BWC platform, computer-aided dispatch (CAD, the system that logs dispatch events and unit assignments), and records management system (RMS, the database where reports are stored) does the agency run, and what AI integrations are currently active? Third, what is the workforce's AI literacy: can officers and dispatchers describe the basic failure modes of AI-generated output? Fourth, what is the agency's community trust posture: does the community have existing concerns about technology surveillance that will shape how AI is introduced?

An honest assessment takes four to eight weeks with a small internal team. The output is a baseline document that names the gaps, not a readiness score. An agency that believes it is ready before the assessment is an agency that has not assessed honestly.

The Pilot Stage

A pilot is a controlled deployment of a single use case with a defined cohort, defined success metrics, and a defined end date. The use case should be selected for highest value and lowest evidentiary risk. AI-assisted redaction of body-camera footage for public records releases is a strong pilot candidate: it offers real time savings, it has a clear accuracy metric (the missed face or license plate is a measurable error), and it does not sit in the sworn narrative chain that Brady v. Maryland (the 1963 Supreme Court case requiring disclosure of exculpatory evidence to the defense) and Giglio v. United States (the 1972 case requiring disclosure of impeachment evidence including information affecting officer credibility) make constitutionally sensitive.

AI-assisted report writing is a higher-risk pilot because the output becomes evidence. Officers in the pilot must have verified the output against the BWC footage, the CAD entry, and field notes before adoption. The pilot should track not only time savings but error detection: how many gap-fills did officers catch and correct? The gap-fill rate in the pilot is the single most important data point the agency has about the maturity of its verification practice.

A well-run pilot produces three outputs: validated time savings figures that are specific to this agency's call volume and incident mix, a gap-fill and correction log that shows the verification pass is real and active, and a disclosed-use record that shows the agency met its prosecutor and discovery obligations throughout. Without all three, the agency cannot claim the pilot validated anything.

The Validation Stage

Validation is the stage most agencies skip or compress. It is the period between the pilot and the scale decision during which leadership reviews the pilot outputs against the success criteria, the prosecutor's office reviews the disclosure record, the oversight body reviews the error log, and the community is briefed on what was tested and what was found.

Validation takes political will. A pilot that produced good time savings but also showed a high gap-fill rate that officers were not consistently catching is a pilot that is not ready for scale. Calling that a validation success because the speed numbers looked good is how agencies produce the high-profile failure that reverses the whole program. Validation requires honest scoring against all criteria, including the integrity metrics that are harder to celebrate at a council presentation.

The Scale Stage

Scale means deploying the validated operating model across the agency. It is not the same as deploying the tool across the agency. Deploying the tool without the operating model is the failure mode. At the scale stage, the operating model, including policy, training, accountability structure, audit trail, and measurement, must all be in place and operational before the first officer outside the pilot cohort touches the AI-assisted workflow.

The Electronic Frontier Foundation (EFF), the civil liberties organization that has raised specific transparency concerns about AI-generated police reports, is watching how agencies handle scale. The concerns the EFF raises are not hypothetical: opacity about what AI did, how it was reviewed, and what errors occurred in the agency's records. The operating model answers those concerns not by dismissing them but by making the record available. Transparency is legitimacy at scale.

The Evidentiary Spine of the Operating Model

The operating model for a public safety agency using AI has a spine that runs through every use case and every workflow: the evidentiary standard. A police report is not a memo. It is evidence. It is disclosed to the defense under Brady, it is tested at trial, and if an officer's credibility is at issue, it is disclosed under Giglio as impeachment material. That single fact reorders every design decision in the operating model.

The evidentiary spine means that the operating model is designed around the question "How does this workflow produce output that holds up in court?" not the question "How does this workflow save the most time?" Those two questions are not always in conflict. Axon Draft One's 82 percent report-time reduction is real and significant. Officers spending 30 to 40 percent of every shift on paperwork is a documented productivity problem. The time savings are a genuine organizational benefit. But the time savings do not survive a suppression motion that throws out the report they were used to produce. The operating model earns both: the time savings and the evidentiary integrity. The two are compatible, but only if the operating model is designed to produce both from the start.

The evidentiary spine shows up in specific design choices. Every AI-generated draft must have a human review and adopt step, and the adopt step is explicit and logged. Every disclosure must document AI involvement and how it was reviewed, in a format the prosecutor has accepted. Every quality metric must include an accuracy and error-detection component, not just a speed component. And the verification standard, the footage-grounded pass that checks every factual claim against the BWC recording and the CAD entry, must be enforced by the operating model, not left to individual officer judgment on a per-shift basis.

The Brady and Giglio Integration

Brady v. Maryland requires disclosure of exculpatory evidence. Giglio v. United States requires disclosure of impeachment evidence. When AI was used to produce or assist a report, and there is a reasonable question about whether the AI introduced an error, suppressed a detail, or softened a fact in ways that affected the narrative, those concerns are constitutionally relevant under Brady and Giglio. The operating model must include a disclosure standard that specifies exactly what must be disclosed, in what format, and to whom, when AI touched the evidence chain.

The King County, Washington, prosecutor's bar on AI-written reports is the clearest signal in the field that prosecutors are treating this as a discovery and constitutional issue, not a technology preference. An agency that meets King County head-on, with a disclosure standard the prosecutor's office has reviewed and accepted, has solved the King County problem. An agency that ignores the King County signal until a case is challenged has created it for themselves.

The Contract and Vendor Dimension

The transformation arc is not just an internal organizational journey. It intersects with a procurement reality that every executive must understand before signing anything: the bundled, multi-year, sole-vendor contract.

Contracts for AI-integrated public safety technology have grown to include cameras, drones, cloud storage, AI processing, and evidence management in a single bundled agreement. Contracts in this space have reached approximately 45 million dollars with terms of up to ten years. A ten-year contract locks the agency into a single vendor's technology, pricing, and data-handling practices through the full duration of its AI transformation. If the vendor's AI model proves to have a bias problem, the agency cannot easily change vendors. If the vendor's disclosure practices create discovery exposure, the agency is contractually bound to a technology it may want to exit. If the policy environment changes (as it has, with King County) the agency may need capabilities the vendor's platform does not support, with no easy way out.

The CJIS Security Policy obligations stay with the agency no matter what any contract says. The vendor can agree to handle data in a CJIS-compliant way. The vendor cannot assume the agency's CJIS obligations. If the vendor has a breach, the agency is still accountable for the criminal justice information that was in the vendor's system. This is not a legal technicality. It is a structural accountability reality that must inform every vendor negotiation and every contract term.

The operating model approach to contracts is to negotiate specifically against the failure modes of long-term lock-in. That means data portability terms that specify how the agency gets its data back if the contract ends. It means performance standards and audit rights that allow the agency to independently verify that the vendor is meeting its obligations. It means termination clauses that allow the agency to exit if the vendor's AI model is found to have systemic accuracy or bias problems. And it means modular contract structures where possible, so the agency can renegotiate the AI component without renegotiating the camera contract.

Governance and the Three Audiences

Running an AI transformation in public safety means running it simultaneously for three audiences: command staff, the city council (or county commissioners, or elected governing body), and the community. Each audience needs a different level of detail, a different vocabulary, and a different emphasis. But they all need the same factual account. The transformation fails when the story told to command, the story told to council, and the story told to the community diverge. If they diverge, eventually someone will notice, and the credibility of the program will be damaged in ways that take years to repair.

Command staff needs the operational story: the operating model, the verification standard, the error metrics, the disclosure compliance rate, and the gap-fill data from the pilot. They need to understand what officers are doing differently, what supervisors are checking, and where the accountability structure sits. The command audience can absorb technical detail if it is presented in operational terms, and they need it, because they are the ones who will face the deposition question when a case is challenged.

The city council needs the governance story: the policy that governs use, the accountability structure, the community input process, the measurement framework, and the constitutional obligations the agency is meeting. Council members are not operations experts. They are stewards of public resources and community trust. They need to understand the safeguards, the investment, and the accountability, and they need to be able to answer their constituents' questions with honest, factual responses. The executive who walks into a council presentation with only the efficiency numbers and none of the safeguard details has not prepared the council; they have set them up to be blindsided.

The community needs the transparency story: what AI tools the agency is using, what they do, how errors are caught, what the disclosure obligations are, and how the community can raise concerns. The EFF transparency concern is that this information is not available. The agency that makes it available, proactively and in plain language, has turned the EFF's concern into a strength. Community oversight bodies, civil liberties advocates, and engaged residents become partners rather than adversaries when they have access to real information rather than vendor marketing.

Key Takeaways

• The transformation from pilot to operating model requires six elements: written policy, a competency-verified workforce, named accountability roles, an integrated audit trail, active prosecutor and oversight alignment, and a measurement system that tracks efficiency, integrity, and trust together.
• The transformation arc has four stages: assess, pilot, validate, and scale. Skipping or compressing any stage pushes its failure mode to the most expensive point in the arc to resolve.
• The evidentiary spine of the operating model is the design principle that every workflow produces output that holds up in court. Time savings and evidentiary integrity are compatible, but only if the operating model is designed to produce both from the start.
• Brady v. Maryland and Giglio v. United States make AI involvement in the evidence chain a constitutional disclosure matter, not a technology preference. The King County, Washington, prosecutor's bar on AI-written reports is the governance line the operating model must address proactively.
• Bundled sole-vendor contracts reaching approximately 45 million dollars with up to ten-year terms create lock-in risk that the operating model must mitigate through data portability terms, audit rights, and modular contract structures.
• CJIS Security Policy obligations remain with the agency regardless of vendor agreements. No contract transfers the agency's accountability for criminal justice information handling to the vendor.
• The three audiences of a transformation, command staff, city council, and community, need the same factual account in different levels of detail. Divergent stories across audiences destroy the credibility of the program.
• The agencies that succeed are not the ones that move fastest. They are the ones that build a model that can be explained to a prosecutor, a council member, a community advocate, and a judge with the same answer, and that answer is: "Here is what we do, here is who checks it, here is the record."