The Data Is In on Body-Camera Reporting, And It's Not Saving Anyone Time

Every few months, a new story surfaces confirming what many of us in this space have been saying since AI report writing tools first hit the market: generating police reports from body-worn camera audio is not the efficiency win it was sold as. The latest comes from Forbes, where Thomas Brewster obtained a cache of emails through public records requests from the Lafayette, Indiana Police Department, an agency of just over 70,000 residents that spent seven months testing Axon's Form One tool (Brewster, 2026). The findings should give any administrator pause before signing that AI Era Plan contract at $199 per officer, per month.

Form One is Axon's newer addition to its flagship Draft One product. Where Draft One writes the narrative of what happened, Form One is supposed to handle the administrative side, pulling officer and citizen names, ID documentation, license plates, and vehicle information out of the body camera audio and filling out the forms and PDFs officers have to complete. On paper, that is exactly the kind of duplicative data entry any cop would love to hand off. In practice, one Lafayette officer put it bluntly in an email: "I know it doesn't save me time and I know it has inaccuracies that I will have to edit" (Brewster, 2026). Another described how a simple form that used to take thirty seconds to complete manually now takes three minutes because of all the errors he has to fix. Three minutes. On a thirty-second task. When the tool struggles to capture the right names and license plates even when they are clearly stated on the footage, you have not automated the paperwork; you have added a proofreading assignment on top of it. As that officer wrote, the tool "dramatically increases the time it takes to finish reports" (Brewster, 2026).

None of this is new, and the research is thinner than the AI skeptics claim. The study everyone cites, No man's hand: artificial intelligence does not improve police report writing speed (Adams et al., 2024), found no time savings from Draft One, but look at what it actually measured: the clock ran from the moment an officer opened a report to the moment it was submitted. That is not writing time, it is elapsed time, and any cop who has started a report on scene and finished it three calls later knows those are nowhere close to the same thing. Strip out that broken stopwatch and the durable finding is simpler: officers spent their time editing, handed a narrative built from raw audio and left to repair it. It also left every property file, crime code, and record management field untouched, the tedious data entry that still had to be done by hand. Manchester's Lieutenant Matthew Barter put it plainly, "It was just easier to type the report themselves" (Brewster, 2026), and Anchorage walked away in 2024 because its audio-only design forced officers to hand-enter every visual detail they never said out loud on scene. That is an architecture problem, not a rollout hiccup, and no customer survey boasting 300,000 saved hours makes it disappear.

It fails the same way in every department. As any cop who has worn a body camera knows, the overwhelming majority of the audio captured on a scene has nothing to do with the incident being investigated. When you feed all of that ambient chatter and background noise into a large language model and ask it to build a report, you are asking the AI to decide what mattered. That is how you end up with the errors King County, Washington cited when its prosecutor's office announced it would not accept AI-assisted reports at all: names swapped, or officers placed at a scene when they were only on the radio. As prosecutor Daniel Clark noted, those are "the type of error that can really challenge an officer's credibility when they're on the stand" (Brewster, 2026).

Some competitors have tried to solve this by adding more raw data rather than less. Code Four, a Y Combinator startup cited in the Forbes piece, uses both body camera audio and video to generate its reports. As I've written previously, layering in video visualization doesn't fix the problem; it compounds it. Any supervisor who has reviewed use-of-force footage knows the camera angle does not capture everything an officer sees and processes in the live moment. Take an officer running Standardized Field Sobriety Tests: if the footage catches a suspect stepping off the line during the walk-and-turn but the officer didn't observe it on scene, a video-driven report would fold that observation in as if it factored into the arrest decision, creating real evidentiary problems. More input the officer doesn't control is not the answer.

This is precisely why we built Policereports.ai the way we did. Our system is not a body-camera transcription engine. It is an active-input platform that keeps the officer in control of the narrative from the very first word. The officer dictates the incident as he or she remembers it, or uploads comprehensive written details, controlling what is relevant, what the legal elements are, and what the reporting structure of their specific agency requires. The AI's job is to take the officer's own account and organize, format, and complete the documentation efficiently, not to guess at what happened from hours of captured audio. When the officer supplies the facts, the system is not "filling gaps," and the propensity for the kind of confabulation that produces phantom license plates and shape-shifting frogs drops dramatically.

That design is also where genuine efficiency comes from, and it is where we do what Form One only promises. The tools in these studies stalled for a specific reason: the officer received a draft built from a noisy transcript and then spent the "saved" time repairing it. Remove the corrupted draft and you remove the editing trap. Because the officer supplies the facts up front, in the structure the agency requires, the narrative comes together faster and there is far less to fix on the back end. On top of that, we auto-complete the agency's PDFs and forms, the ones that live inside and outside the records management system, directly from the officer's own report rather than scraping them out of ambient audio. That piece is more tedious than cognitive, but it is real time handed back to an officer's shift, and it is the same task Form One is chasing, done from the officer's account instead of a noisy transcript. We integrate into existing agency workflows, build unlimited custom modules for each specific task, and cut the duplicative data entry no one should have to do twice. We are not an AI wrap producing a generic draft that then has to be reformatted by hand.

And because these documents sit at the foundation of the criminal justice system, we do not treat accountability as an afterthought. Our platform runs a two-tiered verification architecture, a generator that drafts from the officer's input and a quality assurance checker that validates that draft back against the original input, flagging fabrications, contradictions, speculation, and inconsistencies for the officer to resolve. We store every iteration of the process, the original input, the prompt, the generated draft, and the final report, all encrypted, timestamped, and fully auditable. We never train on agency report data, keeping PII, CJIS, and law enforcement sensitive information out of the models entirely.

The lesson from Lafayette, Manchester, and Anchorage is not that AI has no place in the report writing process. It is that how you implement it determines everything. Feed a model raw body-camera audio and make the officer an editor, and you get credibility problems on the stand and a technology that quietly gets abandoned after a few months. Keep the officer in control of the input, build the system around the agency's real workflow, and put genuine verification and audit infrastructure behind it, and you get what agencies were promised in the first place: faster, more consistent, legally sufficient documentation that lets officers get back to the communities they serve. And the fair way to settle it is honest measurement of the work itself, not customer surveys, and not a stopwatch that runs straight through an officer's downtime. That is a bar every vendor should welcome, ours included. We built our company around the approach that can meet it.

-Chris Ryan is a former law enforcement officer command staff member and the current Chief Product Officer at Policereports.ai

References

Adams, I.T., Barter, M., McLean, K. et al. No man's hand: artificial intelligence does not improve police report writing speed. J Exp Criminol (2024). https://doi.org/10.1007/s11292-024-09644-7

Brewster, T. (2026, July 22). Axon Says AI Police Reports Save Time. Public Records Show They Get Facts Wrong. Forbes. https://www.forbes.com/sites/thomasbrewster/2026/07/22/axon-says-ai-police-reports-save-time-public-records-show-they-get-facts-wrong/