AI Redaction for Court Filings: Acrobat vs Everlaw vs Custom

By Jude Lee · · Comparison

Two lawyers reviewing redacted exhibits and a PDF on a laptop in a small firm conference room

The job nobody puts on the automation list

Ask what to automate first in a small firm and you’ll mostly hear intake, billing, and calendaring. Redaction rarely comes up. That’s strange, because it has every property that makes a task worth handing to software: it’s repetitive, it follows written rules, it scales with page count rather than skill, and the cost of a miss is public and permanent.

It’s also one of the least-marketed real uses of AI in firms, because it’s unglamorous. Large language models are genuinely good at spotting that a nine-digit number in a medical bill is a social-security number and that “D.O.B. 4/2/1981” is a birth date, including in messy OCR output where a regular expression fails. They are not good at deciding whether a paragraph is covered by a protective order. That split — recognition versus judgment — is the entire design brief.

Know the rule before you automate the rule

In federal civil cases, Fed. R. Civ. P. 5.2(a) limits what personal data can appear in a filing: social-security and taxpayer-identification numbers, birth dates, the names of individuals known to be minors, and financial-account numbers, with permitted partial forms (last four digits, birth year, minor’s initials). Criminal and bankruptcy rules have parallel provisions. State courts vary — some have their own privacy rules and their own e-filing system requirements — so check your court’s rules and local rules directly rather than trusting a vendor’s default template.

On top of that sit protective orders, sealing orders, privilege, and any client-specific confidentiality obligations. Those are drafted per matter and are exactly where automation stops being a rules engine and starts needing a lawyer.

5
Categories of personal identifiers enumerated in Fed. R. Civ. P. 5.2(a) — count them in the rule text yourself
Fed. R. Civ. P. 5.2(a)
pages × min/page
Model your own redaction time this way — don't trust a vendor's headline

What an agentic redaction workflow actually is

A chatbot answers a question. An agent carries out a multi-step task and takes actions in your systems. For redaction, that sequence looks like: retrieve the document set from your DMS, OCR anything image-based, classify each document type, apply a written redaction policy, produce a list of proposed redactions with page coordinates and a reason code for each, queue them for human review, then burn in the approved set, strip metadata, save a new version back to the DMS, and write an audit log.

The glue for steps one and nine is MCP, the open protocol for giving an AI assistant governed access to your tools and data. The policy itself is best expressed as a skill — a packaged, versioned instruction set that makes the agent redact the same way every time, the same discipline described in repeatable document review skills. A skill is where you write things like “in this matter, redact settlement amounts but not carrier names” so the rule lives in one reviewable file instead of in someone’s head.

A redaction agent’s job isn’t to be right. It’s to be reviewable — every proposed mark carries a page, a span, and a reason you can defend.

Comparing the three realistic paths

Adobe Acrobat Pro. Its search-and-redact feature finds patterns and phrases across a document and applies true redaction on burn-in. It is cheap, already on most firm desktops, and completely adequate for a handful of exhibits. It has no matter-level memory, no bulk queue across thousands of files, and no audit trail beyond what you keep yourself. Verify current capabilities in Adobe’s own documentation — features move.

eDiscovery platforms. Everlaw and Relativity both market redaction tooling built for volume, including pattern-based and bulk application, with review workflows, QC passes, and logging attached. Check each vendor’s current redaction documentation before you rely on any specific capability — these feature sets change release to release, the same caveat that applies to Acrobat. If your documents are already in the platform because you’re running a review, this is usually the correct answer and building anything custom is a waste of money. The catch is that everything has to be in the platform, and hosting fees mean small matters often aren’t.

Platform redaction (Everlaw, Relativity)
Best when documents are already loaded for review. Bulk tools, QC workflow, and audit trails are built and tested. No engineering. Costs scale with hosted data; rules are the platform’s rules; won’t touch the PDFs sitting in your DMS or a client’s shared drive.
Custom agent over your own systems
Best when redaction spans systems and follows firm-specific policy — medical records into a demand package, exhibits in NetDocuments, PDFs a client emailed. You own the policy file and the log format. Costs an engineering build plus ongoing maintenance, and you inherit responsibility for the burn-in step being genuinely destructive.

A fourth option deserves naming: rules-based automation with no AI at all. If you only ever redact SSNs and account numbers from clean, text-layer PDFs, a scripted pattern match is faster, cheaper, and more predictable than a model. Use the model when documents are scanned, inconsistent, or when the target is contextual (“the treating physician’s home address”) rather than a fixed format.

Which slice of redaction work is worth automating

Concretely: in a personal-injury matter, the automatable band is usually medical bills and records, insurance correspondence, and pharmacy printouts — high page counts, formatted identifiers, the same five things redacted every time. In an employment matter, it’s payroll registers, benefits enrollment forms, and HR files with third-party employee identifiers. What stays manual in both is anything where the redaction call depends on the protective order’s definition of “Confidential” or on privilege.

  1. Sort by rule clarity, not by volume

    Identifiers under FRCP 5.2 and your local rules are objective. Protective-order categories are interpretive. Automate the first group; have the agent flag the second for a human rather than act on it.

  2. Write the policy down before you build

    One page per matter type: what gets redacted, what gets partially masked, what never gets touched. This becomes the skill. If you can’t write it, an agent can’t execute it.

  3. Make the agent propose, never file

    Output is a review queue with reason codes, not a finished PDF, and every item in that queue gets approved or rejected by a person. Pair it with an explicit oversight model so it’s clear who signs off and what they’re attesting to.

  4. Verify the burn-in independently

    After approval, extract text from the output file and search for the redacted strings. If anything comes back, the pipeline is broken. Automate this check; don’t rely on eyeballs.

  5. Log everything

    Document, page, span, rule cited, model proposal, human decision, timestamp. This is what makes the process defensible if it’s ever questioned.

What it’s worth, using your numbers not ours

Skip the vendor ROI headline and build your own: (pages per month × minutes per page ÷ 60) × the blended hourly cost of whoever does it today = current spend. Then estimate the residual — review time on proposed redactions is real and won’t be zero. The difference is your ceiling, not your result.

Don’t anchor on a generic hourly-rate benchmark; use your own blended cost. What matters more than the rate is what the recovered hours get reallocated to. Hours moved from paralegal redaction to billable case work convert to revenue; hours moved to nothing convert to nothing. The same discipline applies across other automation candidates ranked by hours recovered.

Where these agents break

Handwriting and poor scans defeat OCR, and a model will happily redact confidently around text it misread. Spreadsheets and native files often need different handling than PDFs. Redactions in images, charts, and signature blocks get missed more than text does. Partial masking is its own failure mode: on a hyphenated SSN like 123-45-6789, a tool that counts characters rather than digits can mask the wrong span and leave part of the number exposed while appearing to comply. And models sometimes over-redact, which is safer but can gut the evidentiary value of an exhibit.

Choosing between the four options

There is no single best tool, and any list that names one winner is selling something. The best tool is the one matched to the job: platform redaction if your documents are already in review, Acrobat if you handle a dozen exhibits a month, a scripted rule if the pattern never varies, and a custom agent when the work spans systems and follows policy nobody else’s product knows. The same logic that governs discovery triage decisions applies here — pick the narrowest tool that covers the actual workflow, and be willing to conclude that the answer is “nothing new.”

My own assessment, as of 2026 and offered as opinion rather than a measured finding: redaction is among the steadier AI use cases in legal work precisely because the recognition problem is well-defined. That apparent reliability is exactly why it’s tempting to skip the review step. Don’t.

Where is your firm losing billable hours?

Get a free automation audit: we map your intake-to-invoice workflow and show you exactly what's worth automating — before you spend a dollar.

Get a free automation audit

Related guides