Public Records Request and Redaction Software Problems: The 7 That Cost Real Money, and How to Avoid Them

Why does the exemption log get bolted on at the end?

Because it is invisible in every demonstration. Buyers evaluate intake, the public portal, the request number, the correspondence history and the deadline clock, because that is the part of the process everyone can see failing today. The exemption log is a back office artefact nobody looks at until a denial is challenged, so it lands in a later phase and arrives as a report built over data that was never designed to support it.

The result is structurally weak. A summary written after the review cannot be tied to the released product, because the reviewer redacted in one tool and typed the citation in another. Two reviewers applying the same exemption to the same category of content will disagree and nobody will notice. And the response letter carries a paragraph of general language rather than an index, which is an invitation to a challenge.

Build it as the spine instead. Every withheld element, whether a page, a paragraph, a field in a spreadsheet or a range of frames in a video, carries a citation to the specific statutory exemption, the reviewer who applied it, the timestamp and a short justification note, generated as a consequence of doing the review rather than described afterwards. It cannot drift from the released product because it is the same object.

Three things follow immediately. The response letter generates itself with an accurate index. An appeal is answered in an afternoon instead of a fortnight. And a supervisor can finally see inconsistent exemption use across reviewers, which no tracking tool will ever surface for you.

What goes wrong when you migrate past requests and evidence?

The instinct is to migrate everything, and it is usually wrong. Historic requests in a tracking tool carry a request number, a requester, a date and some correspondence. What they almost never carry is the linkage you actually want: which items were responsive, which were withheld, and under which citation. Moving that data into a system built around per item exemption logging produces records that look complete and are hollow.

The specific traps are worth naming. Released packages stored as flattened files with no relationship to the source records they came from, so you cannot tell what was withheld, only what was given. Redacted video exported as a new file with the original retained separately and no link between them. Fee calculations recorded as a total with no breakdown, which matters because most state rules distinguish search time, review time, duplication and the cost of a redaction technician.

Migrate the request metadata and the correspondence, keep the released packages in a readable archive, and mark everything pre conversion so nobody mistakes an imported record for one produced under the new discipline. Then be honest with staff about the boundary. An office that believes its history is fully searchable, when it is not, will make a commitment in a response letter it cannot keep.

One thing genuinely worth extracting from the old data: last year's requests sorted by hours spent rather than by count. That is the analysis that tells you what to build, and in most agencies the answer is that a small number of video requests consumed the office.

Why do records, evidence and email connectors break after launch?

Because they belong to vendors and departments with no obligation to your roadmap. A body camera platform changes its export interface or its rate limits. The records management system is upgraded and the case number format shifts. The email archive is migrated to a new tenant and the credentials with it. A department reorganises and the account your connector authenticated with is disabled during an offboarding.

Vendor cooperation varies enormously and it is a procurement fact rather than a technical one. Pulling media out of a camera vendor's platform at volume, without paying that vendor for the privilege, is a specific capability and it can be withdrawn or repriced. Ask any developer to name the platforms they have done it against in production, and separately ask your own contract manager what your agreement says about bulk export.

Operationally, the pattern that survives is the same as everywhere else and is skipped just as often. Version the connector per source. Alert on silence as well as on error, because a search connector returning nothing looks identical to a search with no hits, and that misreading produces a response letter that says no responsive records exist. That is the single worst outcome available in this domain and it is entirely preventable with a volume expectation and an alarm.

Scope connectors one at a time by request volume, starting with whichever system generates the most responsive material. For police agencies that is almost always the evidence store rather than the records system.

What happens when fee rules and appeal workflow are not covered?

Fees generate their own appeals, which is a category of work most offices do not anticipate. Fee rules are fiddlier than they look because most state acts distinguish between search time, review time, duplication and the cost of a redaction technician, and some cap or waive categories depending on the requester and the purpose. A system that calculates a single hourly figure and multiplies produces estimates that are indefensible in exactly the same way an undocumented withholding is.

The second gap is the appeal itself. Offices track requests and then handle appeals in email, so the record of what was argued, what was reconsidered and what was released on review lives nowhere. When the same requester files again on related material, the office starts from nothing.

Cover both explicitly. Fee estimates itemised by the categories your state act recognises, with the underlying time and page counts traceable, and the estimate archived as sent. The appeal tracked as a case linked to the original request, with the withheld items it concerns, the outcome per item, and any change written back to the exemption log with a reason. Litigation holds belong in the same structure, because a hold that exists only as an email to the records officer is a hold that fails when that person is on leave.

Should you build custom or configure what you already own?

If you are a small or mid sized city clerk with document requests, no body camera programme and a handful of appeals a year, buy. GovQA, NextRequest and JustFOIA are competent products and they solve the front of the process well: portal intake, request numbers, correspondence history, fee estimates, deadline reminders and requester visibility without a phone call. They will improve your life next month for a fraction of what discovery alone costs on a custom project, and the honest diagnosis is usually that your bottleneck is intake discipline rather than software architecture.

Take Veritone Redact and its peers seriously for the video work itself. Automated face and plate detection with tracking across frames is genuinely hard and those vendors have invested in it. The gap is not detection quality, it is that redaction is sold as a media tool rather than as part of a records decision, so the audit trail you need has to live in whatever you build or maintain around it.

Build when at least two of these describe you. Requests routinely span video, records and staff email in one filing. You have lost a fee award, or been warned about one, over withholding you could not document. You are a police records unit whose video backlog is measured in months. Or the same request arrives at five departments and nobody can see it is one request being worked five times.

How do hidden costs get into the quote?

• Connector count. Each source system is real integration work and vendor cooperation varies. A quote listing integrations as a single line has not been costed.
• Video infrastructure. At volume you are making storage and transcode capacity decisions a document only office never faces, and those are recurring costs rather than build costs.
• Fee calculation. Assumed simple, actually a rules problem shaped by your state act and by requester category.
• Accessibility. Public facing pages for a government body carry standards that are not optional and are cheaper designed in than retrofitted.
• Retention and disposition. Released packages, originals and redaction artefacts all sit under records schedules, which is design work rather than a policy attachment.
• Reviewer training and parallel running. Your most experienced reviewers doing the job twice for several weeks, and their hours are rarely in the plan.

What separates a build that works from one that fails here?

The builds that work measure redaction throughput as review minutes per footage minute after automation, not as detector accuracy. Accuracy is easy to make look good on clean daytime video and tells you nothing about a foot pursuit at night. Throughput is the number that determines whether you meet a statutory deadline, and it is the number to hold a vendor to in writing.

They design the human as a checker rather than an artist. Automated detection of faces, plates, screens and documents in frame, object tracking so a subject stays redacted through a pan, audio redaction driven by a transcript so a reviewer marks a spoken address in text rather than scrubbing a waveform, and a confirmation pass where the reviewer corrects rather than draws. Be honest about the residual: detection will miss a face at the edge of frame in low light and will over redact a reflection. The goal is not zero human review.

They start with the shared intake and the statutory clock when the same request hits several departments, then add departmental fulfilment workspaces as each one is ready. Forcing every department onto a single workflow on day one is how these projects stall, because departments have different source systems and different appetite for shared visibility.

They make the responsiveness decision explicit. Each item marked responsive, non responsive or withheld by a named reviewer, with withheld items carrying citations automatically. That is what turns the exemption log from a document into a byproduct.

And they settle ownership before kickoff: the repositories, the cloud accounts, any model artefacts trained on your footage, and the unrestricted right to bring in another firm. At Digital Heroes that is the default from the first commit. A developer who wants to retain redaction models trained on your video is retaining control over your records pipeline, and that should end the conversation.