AI Suggests What to Redact. Your Reviewers Decide.

ThinkRedact autonomously processes each page, surfaces redaction recommendations with FOIA exemptions and statutory justifications, and orchestrates the review workflow from intake to final release. Every release requires human sign-off.

Book a demo See how it works

A sample lease amendment, shown in the ThinkRedact review screen. All names and numbers are made up.

97%of sensitive data found before review
180 pages/minaverage throughput of AI detection
9supported file families, with flexibility to add more
100%of released redactions approved by a person

From Upload to Release in Five Steps

Every document follows the same path. Each step below opens the screen your team works in, with a sample document moving through it.

  1. Step 1UploadDrop in one file or a whole batch. Each one becomes a readable PDF.
  2. Step 2DetectRules, a language model and vision read every page for what to withhold.
  3. Step 3ReviewReviewers accept, reject or adjust each suggestion, and add their own.
  4. Step 4ApproveThe document moves through your review steps to a named approver.
  5. Step 5ReleaseApproved redactions are burned in, and the file and its log are ready to send.
Step 1

One File or Hundreds, in a Single Upload

Drop in one document, a mailbox export or a ZIP of a large batch of records. ThinkRedact unpacks the batch, identifies each file by its contents, converts it to PDF and reads it, so every document opens in the same viewer.

Nine File Families

PDF, Word, Excel, PowerPoint, Google exports, Outlook and other email, mailbox archives, images and ZIP files.

OCR Only Where It Is Needed

Scans and photos are read with OCR. Pages that already carry text are left as they are.

Large Batches, Tracked Together

Files are processed in parallel. The batch shows how many are ready and how long it took, and each file shows the stage it has reached.

Step 2

Three Ways of Reading Every Page

Each kind of sensitive content is found by the method that suits it. Every suggestion names its exemption, how confident it is, and the evidence behind it.

Pattern Rules

Tax ID and Social Security numbers, phone numbers, personal email addresses, bank and card numbers, dates of birth and home addresses, read together with the labels around them.

Language Understanding

Rents, rates, pricing and other commercial terms under b(4), and privileged legal advice under b(5), offered only where the criteria your administrator sets are met.

Signature Detection

Handwritten signatures and initials on scans and photos under b(6), boxed on the ink itself.

Signature Boxes That Fit the Ink

The model says where a signature is. ThinkRedact then measures the ink itself and draws the redaction around every stroke, with a small margin, so no part of a signature is left showing.

Scans, Photos and Initials

Signatures and initials are found on scanned pages, phone photos and pictures pasted into documents.

Privilege Within Your Criteria

A b(5) suggestion is offered only when a message meets the criteria in your configuration, such as counsel on the thread, and the model decides what part of it is privileged.

Step 3

Every Suggestion Explains Itself

Reviewers work page by page or straight down the list. Each suggestion says why it was flagged and how sure the system is, so a decision takes seconds.

Accept, Reject or Adjust

One click or one key per suggestion, or decide in bulk by page, exemption or confidence.

Add What Was Missed

Draw a box, select text, redact whole pages, or use Find & redact to cover every place a name or number appears.

AI suggestion. Dashed, awaiting a decision.
Reviewer drawn. Solid outline, added by a person.
Approved. Filled, applied in the released file.
Step 4

Reviewed by the Right People, in the Right Order

Each document follows a route: the review steps it passes through and who may act at each one. A sensitive release can add a senior review before approval. A simple one can go straight to an approver.

Routes Set in One File

An administrator lists the routes and the roles allowed at each step. Changing them needs no code.

A Named Person at Each Step

Assign every step to someone, see who has each document, and reassign when someone is out.

Sent Back When Needed

Any step can return a document to the one before it, and every decision keeps the name of the person who made it.

Step 5

Nothing is Left Under the Black Box

Release builds a new version from the original. The text and images under each approved redaction are removed rather than covered, and the file is rebuilt with its metadata cleared.

An Exemption on Every Redaction

Release waits until each approved redaction cites at least one exemption.

Original and Released, Both Kept

The original stays untouched beside every released version, and only a released version can be exported.

Accessible Output, With Its Log

Released PDFs carry their language, title and page tags. The exemption log that goes with them lists what was withheld and why, never the withheld text.

Every Step, Properly Audited

From the first upload to the last export, every action is recorded with who took it and when. The audit trail cannot be edited or deleted by anyone, including administrators.

Reports and Exports

Processing times, decisions by person and step, exemption counts and the audit trail, as files your team can share. None of them contain withheld text.

Quality Checks

A sample of each document's decisions goes to a second reviewer, and the agreement rate appears in your reports.

Feedback on Detection

Reviewers can mark a suggestion as wrong or a redaction as missed, so you can see where detection needs tuning.

Safeguards Built Into the Platform

ThinkRedact runs in your environment, keeps people in charge of every decision, and keeps your documents out of model training.

A Person Decides Every Redaction

Nothing reaches a released file without an approval recorded under a person's name.

Your Documents Never Train a Model

Page content is used to make suggestions and nothing else. Sensitive data is never used to train, tune or evaluate a model.

Runs in Your Environment

ThinkRedact runs as a containerized application within your own cloud infrastructure with no external dependencies.

Works with any AI Provider

The model connection sits behind one interface, so ThinkRedact can work with the provider your agency approves.

Accessible to Every Reviewer

Every tool works from the keyboard, screens are checked for Section 508, and released PDFs are tagged for assistive technology.

Configured with Ease

Detection labels, privilege criteria, review routes and the quality sample are plain configuration files an administrator can read and change.

Related Modules

Documents
Document AI
Extract, classify and validate content across large document collections.
Governance
Policy & Guardrails
PII and policy guardrails that run inside the runtime, with a record of what fired and why.
Federal
AI in Federal
How ThinkTrends supports federal agencies with secure, governed AI.

Frequently Asked Questions

Common questions about ThinkRedact.

Does ThinkRedact redact anything on its own?+
No. It suggests redactions. A reviewer accepts, rejects or changes each one, and only redactions approved at the final review step are applied to the released file.
Which FOIA exemptions does it suggest?+
Currently, we support b(3) for tax identification numbers, b(4) for commercial and financial information such as lease terms and pricing, b(5) for privileged legal communications that meet your configured criteria, and b(6) for personal information, including handwritten signatures with flexibility to add more exemptions as needed. A reviewer can cite any of the four on any redaction, and more than one where it applies.
Can someone recover the text under a redaction?+
Redactions are burned in. The text under each one is deleted, the image content beneath it is destroyed, and the document is rebuilt with its metadata cleared. The original stays inside the system, and only the released version can be exported.
Are our documents used to train AI models?+
No. Page content is sent to the AI model only to make suggestions, and it is never used to train, tune or evaluate any model. Suggestions are stored when a document is processed, so opening it again shows the same suggestions without calling the model.
How long does a large document take?+
A 100-page document is processed in less than a minute, which can scale to hundreds of documents. Pages are read in parallel, and each document shows its progress while it runs and its total time when it is done.
Can we use it for records other than FOIA requests?+
Yes. Personal information, financial details and signatures are found the same way in any document, and the review steps and detection labels are set in configuration. Talk to us about the categories your releases need.