Redaction software where agents propose, cite and leak-test every box, and you approve the law, not each box
Every box cites its legal basis, and nothing leaves until a leak check proves it clean.
See one case, screen by screen ↓Document Redaction is redaction software that finds and redacts personal and privileged data in documents, scans, photographs, audio and video, and cites the legal basis for every box. Seven agents detect, cite, cross-check, argue against over-redaction and leak-test the export; the analyst approves each basis layer and the release, and counsel approves privilege.
How to choose AI solutions for back-office and enterprise work →
Updated
Redaction is still done one file at a time. An analyst scrubs body-camera video frame by frame, mutes names in the audio, boxes names, dates of birth and phone numbers page by page, then types an exemption code onto each box and retypes the same reasons into a redaction log and the response letter. Video is the expensive part: one published estimate is 10 staff minutes for every minute of footage.
The risk sits where nobody looks. A black box that hides the text but leaves it underneath, an officer's home address in a PDF's document properties, a bystander's name in the original file name, the same person boxed in five files and left visible in the sixth or in last month's release. And over-redaction costs too: a public employee's name withheld without a basis can be reversed on appeal, and in Washington a court may award up to $100 a day for each record wrongfully withheld.
Estimated split for an installment like the one on this page (9 files, 27.6 minutes of media, 28 pages), by hand and with the solution.
Seven agents detect, cite, cross-check, challenge and leak-test every box; the analyst approves each layer and the export.
Maria Lopez is a senior redaction analyst. She does not draw boxes all day: she works with the agents. They propose every box and explain why, she asks back, accepts, rejects or changes a basis, and she signs each layer and the export. Request PR-2026-0563 asks for every record of a police response in the 1400 block of Ferris Street on 2 October, body-camera video included. Installment 1 is due Friday. Here is what happens, screen by screen, in the working solution.
Nine open jobs, three due in two business days, 410 boxes waiting for a decision and 11 leaks caught before release this month. PR-2026-0563, from Noel Hartigan of the Larkhaven Ledger, holds 9 files: 27.6 minutes of media and 28 pages. The agents have proposed 153 boxes, each with a basis; 137 wait for her decision. The clock reads Fri 16 Oct, 2 business days, the date promised in the five-day response.
Agents at work: “Consistency agent found Elena Marsh visible on CAD-26-1002-118 p. 2 while 5 other files and its own page 1 redact her; proposed 1 box.”
On Officer Brandt's 14-minute body-camera clip the detection agent tracked 41 faces (38 adults, 3 children), 3 licence plates and the in-car terminal screen at 08:02, and heard 4 spoken identifiers: a birth date at 03:27, a name at 06:12, a phone number at 09:48 and a child's name at 11:30. At 06:12 the witness says her own name; the transcript shows it bleeped and coded PD. Two partly turned faces scored below 0.80 and wait under “Needs a look”.
The detection agent proposes only. Anything below the confidence threshold goes to a person.
On the incident report each box carries its code: LE 240(2) on the victim and on the second witness, who both asked for non-disclosure; PD 240(1) on the first witness and the subject's date of birth and address; JUV 240(14) on the two children. The layer panel counts each basis with its citation: personal data under RCW 42.56.240(1) and (14), law enforcement under 240(2), juvenile under 240(14). Privilege and trade secret are already approved by Samuel Achebe, the assistant city attorney, on Tue 13 Oct at 16:05.
Personal data: “1 consistency conflict open · Fix first.”
Maria clicks the first witness's name. Under the box: Elena Marsh, visible to reviewers only. Basis personal data, citation RCW 42.56.240(1), proposed by the detection agent at 0.96, with the reason: a bystander who is not a party; naming her is not of legitimate public concern. The sources it read are listed. She can accept, reject or move the box to another basis.
Why: “A bystander who is not a party; naming her is not of legitimate public concern and would be highly offensive (RCW 42.56.050).”
The consistency agent keeps one register of 13 people and identifiers across the 9 files. Elena Marsh is redacted everywhere except page 2 of the dispatch log, where the 22:09:12 entry reads “RP MARSH, E.”, an abbreviation the detection agent did not match. The earlier release PR-2026-0498 redacted her on the same log. One click on “Fix across the set” adds the box.
“Fixed across the set · Elena Marsh redacted in all 6 files.”
The segregability agent challenges every box. It lifts two: Ofc. Tomas Brandt's name on pages 1 and 4 of the incident report. An officer acting in his public role is not exempt, under the privacy officer's rule N-2, and withholding risks a penalty of up to $100 a day. It also checked every home address against published council minutes and permits, every box for being wider than its detail, and the 14 privilege passages for business-only text, and kept them all.
Either lift can be reversed with one click, and a lift on privilege always goes to counsel.
With the conflict fixed, Maria approves her three layers at once: personal data, 107 boxes; law enforcement, 24; juvenile, 7. She picks a reason code, ticks “I have reviewed every box in these layers and its basis” and types her name. The log records one approval per layer. She cannot approve privilege or trade secret: those show “Legal reviewer only”.
“One signature, recorded per layer.”
The leak agent opens the export as an outsider would: it copy-pastes under every box, reads every metadata field, compares file names with the register and listens under the bleeps. It finds three leaks, none of them visible on a page: the officer's home address in the incident report's document properties, “Marsh” in the original video file name, and the scanner's OCR text under 5 boxes on page 2 of the witness statement. “Fix all 3 and check again” scrubs, renames and flattens, then re-checks: clean.
Export stays locked until all five gates are green: layers, consistency, over-redaction, leak check, letter.
Every box is a row: file, page or time, basis, citation, what it covers, who proposed it, confidence, and who approved it and when. 156 rows: 154 boxes and the 2 lifts with their rule. The exemption index for the response letter is built from the same rows, and the log downloads as CSV for an appeal or a privilege log.
Maria approves the export with the reason code “Release: installment 1, as redacted”. The package is produced: 28 pages and 3 media files, 154 redactions burned in, each coded on the record, numbered LRK-0563-000001 to 000028 and M001 to M003, with its SHA-256. The disclosure record carries the $19.89 charge under RCW 42.56.240(14), delivery by secure link after Grace Whitcombe, the public records officer, countersigns, and installment 2 by Fri 30 Oct.
“Package produced · sent to Grace Whitcombe to countersign.”
The same release, seen by the five people who carry it: what their week looked like, and what it looks like now.
Seven agents detect, cite, cross-check, challenge and leak-test every box; the analyst approves each layer and the export.
Classifies the request, applies the template, computes the response clock from the rule and the holiday calendar, and inventories every file with its SHA-256.
Finds names, dates of birth, addresses, phone numbers and signatures in text and scans; faces, plates and screens in images and video, tracked across frames; and names and numbers spoken in audio.
Gives each box its legal basis and citation from the active rule sets, with its reasons and the sources it read, and routes each layer to the role that approves it.
Keeps one register of people and identifiers across the set and earlier releases, so each is treated the same way in every file.
Argues against over-redaction: boxes wider than the detail, information already public and the never-redact rules.
Opens the export as an outsider: text under boxes, document properties, file names, hidden layers, thumbnails and residual audio. Fixes what it finds and checks again.
Burns in the redactions, numbers the pages and media, renames files, and writes the redaction log, exemption index, letter, disclosure record and charge.
Approves each basis layer and the export; counsel approves privilege; the records officer countersigns. The agents propose; a named person decides.
Ask about any request, file, box or basis in plain words. Answers cite their sources and open the screen they come from.
Why is the name at 06:12 redacted?
At 06:12 on BWC-4471 the witness says her own name, Elena Marsh. She is a bystander, not a party, so the basis agent put it in the personal data layer under RCW 42.56.240(14), which protects people's privacy in body-camera recordings. The consistency agent linked it to the same person in the other files. The clip opens at 06:12 so you can hear the bleep.
What would leak if I exported now?
Three things, none of them visible on a page. The incident report's properties hold the officer's home address, exempt under RCW 42.56.250. The video's original file name contains “Marsh”. And the witness statement scan still has OCR text under 5 boxes. Run the leak check to confirm and fix all three.
Which boxes did the agents lift, and why?
Two, both on the incident report: Ofc. Tomas Brandt on page 1 and page 4. Officers acting in their public role are not exempt, and rule N-2 says never redact them. Withholding them could count as a wrongful withholding, with a penalty of up to $100 a day per record. You can reverse either lift on the checks screen.
Draft the exemption index for the response letter
Drafted from the redaction log, not retyped: personal data, law enforcement, juvenile, privilege and trade secret, each with its count and the law it cites, at the place of each redaction. The two officers' names are released. Grace Whitcombe countersigns the letter after export.
13 screens from the working solution, on its sample data. Pick one to see it large.
The rules the agents cite and the thresholds they work to are settings, versioned and approved before they apply.
| Setting | Default | Choose from |
|---|---|---|
| Confidence below which a box needs a look | 0.80 | Slider; the screen shows how many boxes it would send for a look |
| Face detection sensitivity | High (min face 24 px) | High (min face 24 px) · Standard (min face 40 px) |
| Spoken identifier confidence | 0.85 | 0.85 · 0.90 |
| Bleep style | Tone, 1 kHz | Tone, 1 kHz · Silence |
| Video redaction style | Pixelate, code label on box | Pixelate, code label on box · Solid box, code label |
| Rule sets in force | Washington Public Records Act, juvenile records, vehicle records, criminal history, litigation privilege | Also available: federal FOIA, EU / UK GDPR data-subject access, HIPAA health information |
| Who approves which layer | Senior analyst: personal data, law enforcement, juvenile · Legal reviewer: privilege, trade secret | Set by role; roles come from single sign-on groups |
| Body-camera redaction charge | $0.85 per minute · RCW 42.56.240(14) | Your rate and the law that allows it |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample data · “typical” = published figures (City of Monroe, WA, 2026 fee schedule: 10 staff minutes per minute of footage; DOJ Office of Information Policy, Summary of FY 2025 Annual FOIA Reports: 1,707,197 requests; RCW 42.56.550(4): up to $100 a day per record wrongfully withheld) · “estimated” = our estimate. The City of Larkhaven, the Larkhaven Ledger and the people, cases, addresses and records named on this page are fictional: characters and sample data in the working solution.
Document Redaction is redaction software that finds and redacts personal and privileged data in documents, scans, photographs, audio and video, and cites the legal basis for every box. Seven agents detect, cite, cross-check, argue against over-redaction and leak-test the export; the analyst approves each basis layer and the release, and counsel approves privilege.
Both. In video the detection agent tracks faces, licence plates and screens across frames, and in audio it finds names, numbers and dates spoken aloud and bleeps them, in the audio and in the transcript. The analyst reviews by exception on a timeline under the player, with low-confidence tracks marked “needs a look”.
The basis agent gives every box a basis and citation from the rule sets you switch on, such as a public records exemption, a GDPR article or privilege. The code is printed on the box, and the redaction log carries the citation, the sources read, the confidence and the approver. The letter's exemption index is built from the same rows.
Before export the leak agent checks for text left under boxes, document properties, names in file names, hidden layers and annotations, thumbnails and residual audio. It fixes what it finds and checks again, and the export stays locked until the check is clean.
Yes. The demo is a public records request, but the same queue holds litigation productions, employee-file requests and claim files, and rule sets for federal FOIA, EU and UK GDPR data-subject access, HIPAA and litigation privilege can be switched on.
Yes. Agents propose and never approve or release. The analyst approves each basis layer and the export with a reason code and a typed name, only the legal reviewer can approve privilege and trade secret, and the records officer countersigns the letter. Every action is on a hash-chained audit trail.
The Agentic Solution Engine builds and deploys it from your requirements and documents, such as your disclosure rules, request templates and a sample of past releases, and it goes live once every quality gate has passed. We will walk you through it on your own records first.
Every deadline met, and every page traced to the request it answers.
1-min demoPublic Records Request ManagementEvery withheld line cites its law, and every release goes out before the clock runs.
1-min demoLegal Intake & TriageRoutine questions answered in minutes; anything risky reaches a lawyer first.
1-min demoWe'll run Document Redaction on a sample of your own past releases.