62 Days to 7

The agency serves a county population through a disclosure unit of two people. Its body-worn camera programme started in 2022, and by the time the deployment described here began, 175 body-worn and in-car units were in service.

Those two staff handled every public information request the agency received: video, audio and documents alike, about 150 a month. Video and audio were the minority of the volume and the overwhelming majority of the work: 28 percent of requests, and 76 percent of redaction hours.

That imbalance is the whole story of this deployment. Everything that follows is a consequence of it.

The reason video consumed three quarters of the unit's hours is arithmetic, not inefficiency.

A typical video request runs five to ten hours of work for each hour of footage. A crowded scene with several objects that have to stay tracked runs fifteen to thirty hours per hour of footage. One request the unit handled took between eight and ten hours. It was a 30-minute foot pursuit recorded on a single body-worn camera, and the pursuit crossed crowded areas where the footage caught licence plates, house numbers and bystanders throughout.

Those numbers are not unusual. They are what it takes when a person has to watch a recording frame by frame and decide, repeatedly, what has to be obscured and for how long. A face is not redacted once. At thirty frames per second, a single minute of footage is 1,800 separate decisions about the same person, and the person moves, turns away, is occluded by a pillar, and comes back.

With two people and that much work behind each video request, the queue did what queues do. Video took an average of 62 days to release, against nine days for documents. About 130 requests sat open at any given time, and 75 of those had already passed the date the agency had committed to in writing.

State law required the agency to produce records promptly, and to send the requester a written notice naming a new date whenever it could not meet the deadline. Those notices were going out about fifty times a month, or one request in three.

Volume was only half the problem. Each of those redaction calls carried a legal basis, and the bases are not interchangeable.

A bystander and a juvenile in frame are treated differently. A licence plate on an uninvolved vehicle identifies someone who was never part of the case. An address read out over dispatch never appears on screen at all and still has to come out of the audio. The reviewer decided each one of these, and had to be able to account for the decision if a release was later questioned.

Nothing recorded the reasoning. The basis for every withholding lived in the memory of whoever made it, and the released file carried no trace of why a region had been obscured. For an agency whose releases are subject to challenge, that is the exposure that matters more than the backlog. A late release is a problem. An unexplainable one is a different category of problem.

The agency had looked at automating this before. Its evidence management system included a bundled redaction capability.

The unit did not consider it reliable enough to run without a reviewer checking the output, which meant it removed none of the review burden that was actually consuming the hours. At renewal the agency chose not to continue with that capability, and went looking instead for a redaction tool that would work alongside the evidence system it already ran, rather than replacing it.

That constraint shaped the whole evaluation. The camera programme, the storage, the chain of custody and the records workflows were all built around the existing system. Replacing it to fix one step would have been a disproportionate answer to the problem.

While the disclosure unit was drowning in redaction hours, investigators had a different problem with the same evidence.

A single vehicular homicide file held more than 300 photographs and between 70 and 90 hours of video. The evidence was all there. Getting an answer out of it was the difficulty.

Analysis was limited to search, and search only reached words. If something was said on a recording, an investigator could find it in the transcript and search the keywords and the section markers a long recording had been divided into. If something was only visible, such as a vehicle in the background or a person walking past, there was nothing to search against.

So the work was done by watching. An incident with thirty cameras in the surrounding area meant reviewing all thirty to work out which one held the angle that mattered, before the investigation itself could properly begin. Checking whether a suspect's account after arrest matched an earlier one meant playing both interviews from the start. Establishing whether a vehicle had come up in another case meant somebody remembering that it had.

Where a recording had no audio, and fixed cameras rarely have any, there was no transcript to review and no keywords to match. Those files could only be watched.

Building a case took about eight hours of investigator time, most of it spent watching material to find out what was in it, and querying the evidence system, records management and dispatch one at a time. Assembling a package for the prosecutor took three business days.

The agency deployed VIDIZMO Redactor and VIDIZMO AI Intelligence Hub on its own infrastructure, integrated with the cloud evidence management system it already ran.

Nothing was replaced. The evidence management system remained the system of record. The camera programme, the storage and the chain of custody stayed where they were. What was added was the step that had been missing, and a way to ask questions of evidence that was already there.

Redactor integrates with the agency's evidence management system over its API. Evidence moves into the redaction queue and the released copy comes back without anyone exporting a file by hand. Records staff work in a browser, with nothing to install on a workstation.

Detection covers faces, licence plates and people in frame, and finds spoken personal information on the audio track. Crucially, an object is tracked across frames rather than redrawn on each one, which is where the five-to-ten-hours-per-hour figure came from.

The unit built templates for the footage types that recur: traffic stops, foot pursuits, booking and custody, interview room recordings, crowded scenes at public events, and several others. A reviewer now starts from settings that suit the material rather than configuring each request from nothing. Footage submitted at the end of a day is processed overnight and is waiting for review the next morning.

Each redaction sits on its own layer, annotated with the exemption it was made under and a note giving the reason.

That is the part the previous workflow could not produce at all. The basis for a release now travels in the file rather than in a reviewer's memory. When a withholding is questioned, the answer is already recorded. Because redactions are organised as layers, one asset can also serve several disclosure tiers without keeping separate copies that drift apart.

The evidence management system already indexed the words. It transcribed speech and let staff search transcripts, keywords and section markers. What it never did was look at the picture.

AI Intelligence Hub pulls media from that system through its API and analyses it for what the system never captured , detecting the people, vehicles, weapons and objects that appear in footage. The evidence itself stays where it was.

An investigator can now ask a single question covering the evidence system, records management and dispatch at once, and get an answer showing which file it came from and where inside that file, even when the thing they are looking for was never spoken aloud or written into a report.

That changes where an investigator starts. Comparing a suspect's account after arrest against an earlier one used to mean playing both interviews from the beginning. The Hub now points to the passages in each where the subject comes up, so the investigator begins there and reads them side by side. Faced with thirty clips from surrounding cameras, an investigator can ask which of them show the person or vehicle in question and start with those. A silent camera feed can be queried the same way a body-worn recording with speech on it can.

The full recordings stay available, and nothing stops an investigator watching either one end to end. What changed is that finding the relevant minutes no longer takes hours.

Adding a second route to evidence raises an obvious question: could asking the Intelligence Hub surface something the evidence management system would have kept from that person?

It cannot, and the agency examined this before deploying.

Each item carries the access and sealing status of the system that owns it, kept current, so sealing a record in the evidence system also removes it from the Hub's results. A search against a connected system runs under the credentials of the person asking, so that system applies its own rules rather than being bypassed. The index sits on the agency's own servers under the agency's retention policy.

Where the Hub generates a summary it cites the timestamps it drew from, and the investigator reviews those passages directly. Workflows carry human review and approval steps. The finding of record is the investigator's, not the system's.

The figures below are reported by the agency and cover the first six months after deployment, measured against the six months before it.

In that period the unit processed about 340 video and audio requests, 252 of them new and roughly 90 drawn from the backlog. The six months before it, working by hand, produced about 250.

The gap between the median and mean review times is the handful of multi-camera incidents each month that still take twelve to fifteen hours. The 30-minute foot pursuit that had taken eight to ten hours now takes about an hour and a half.

The twelve remaining deadline notices are worth reading carefully, because they are not backlog. They are requests where the agency is waiting on a legal ruling about whether information may be withheld, and requests carrying an unusually large volume of records. The agency has met every date it has set since deployment.

The same two people handle the work. The late sittings have all but stopped.

The agency is pursuing CALEA accreditation, and the exemption record and named-user access log support the documentation its assessors ask for.