In June 2023 a New York lawyer filed a federal brief citing six cases. Five of them did not exist — they had been hallucinated by ChatGPT, which the lawyer had used to research without disclosing the tool and without verifying the output. Mata v. Avianca, Inc. (S.D.N.Y. Case No. 22-cv-1461) became the textbook cautionary tale for the legal profession's AI era and produced one of the most-cited sanctions orders of 2023. The case is also exhibit A for a different point: the legal AI risk in 2026 is not whether to use AI tools, but whether the documents your AI sees are properly redacted before they arrive. Contract redaction, litigation redaction, and the broader discipline of preparing documents for AI-assisted review are now the bottleneck between law firms and the productivity gains every managing partner has been promised.
This is the fourth article in our redaction guide series. The pillar guide covers the underlying technology and the general GDPR framework. The HR cluster covers shadow AI inside HR teams. The accountants cluster covers tax filings and payroll. This article zooms specifically into legal practice — contracts, pleadings, witness statements, e-discovery, M&A data rooms, regulatory filings — where the redaction problem intersects with attorney-client privilege, court rules on public access, cross-border discovery, and the bar's professional conduct rules in ways that no other vertical experiences.
It is also the most directly useful article for any litigation or transactional practice doing cross-border work. A London partner running an arbitration with European witnesses, a Milan litigator handling discovery requests from a US federal court, a Frankfurt M&A team preparing a data room for a buyer in Singapore — every cross-border legal workflow generates a document set whose handling is governed by at least two procedural codes, one or two privilege regimes, the GDPR, and frequently a national blocking statute. The redaction layer is the technical control that operates across all of them.
Terminology note: in this article we use "contract redaction" and "legal redaction software" workflows interchangeably for the same operational layer; "ediscovery redaction", "litigation hold redaction", and "legal hold redaction software" describe overlapping sub-workflows that share the same architecture; "sensitive clauses redaction", "contract review AI" and "AI for lawyers GDPR" compliance refer to the same combined GDPR + bar conduct + privilege framework discussed below. Tooling names like "Harvey AI contract redaction" describe the broader market category of which Redactic is a component. The Mata v Avianca case (often searched without the period) is the legal-AI cautionary precedent we return to repeatedly.
1. Why legal practice is uniquely exposed to redaction risk
Five structural facts explain why legal document redaction sits at a different risk level from any other professional category.
First, the destination of the document is often public by default. Court filings in most common-law jurisdictions, and many civil-law ones, become part of a public record unless specifically sealed. US Federal Rule of Civil Procedure 5.2 (Privacy Protection for Filings Made with the Court) requires social-security numbers, taxpayer-identification numbers, birth dates, names of minors, and financial-account numbers to be redacted in any electronic filing. England's CPR Part 31 governs disclosure and inspection. Italy's Codice di Procedura Civile (CPC) art. 71 and the disposizioni di attuazione regulate access to civil case files. Germany's ZPO §299 governs Akteneinsicht. Spain's LEC art. 235 governs the same area. In each jurisdiction the default — what happens if the lawyer does nothing — is some form of public exposure. Redact court filings is therefore not an optional improvement; it is the operation that prevents the firm from being responsible for the disclosure.
Second, the privilege regime is layered and inconsistent across jurisdictions. US attorney-client privilege and work-product doctrine differ from the English iteration of legal professional privilege, which differs from the EU Court of Justice's interpretation in Akzo Nobel Chemicals Ltd v Commission (C-550/07, 2010) which held that in-house counsel communications do not enjoy LPP under EU competition law. France's secret professionnel de l'avocat is reinforced by the Loi du 31 décembre 1971; Italy's segreto professionale forense is art. 622 Codice Penale plus the codice deontologico forense; Germany's anwaltliches Berufsgeheimnis is StPO §53 + BRAO §43a; Spain's secreto profesional is LO 1/1979 + EGAE. Attorney-client privilege AI exposure occurs whenever a lawyer pastes privileged communications into a consumer LLM, with the privilege analysis differing depending on which jurisdiction's regime applies to the document.
Third, legal documents combine identifiers, sensitive substantive content, and third-party rights. A typical witness statement contains the witness's PII, names of third parties who have not consented to disclosure, sensitive substantive narrative (allegations, contractual disputes, sometimes medical history), and frequently information protected by another regime (whistleblower identity under Directive 2019/1937, trade secret under the EU Trade Secrets Directive 2016/943). The redaction operation must understand all three dimensions simultaneously.
Fourth, the bar's professional conduct rules add their own layer. ABA Model Rule 1.6 (Confidentiality of Information) makes confidentiality a strict-liability obligation; Rule 5.3 governs supervision of non-lawyer assistants, which in 2024-2025 became the basis for several state-bar opinions on AI tools. The CCBE Code of Conduct for European Lawyers and the IBA International Principles on Conduct for the Legal Profession set the European baseline. The State Bar of California Opinion 2024-1 and the Florida Bar Opinion 24-1 on lawyer use of generative AI both require informed client consent for input of confidential information.
Fifth, cross-border discovery is the area where the regulatory contradictions are sharpest. The Sedona Conference Working Group 6 has documented the conflict between US-style broad discovery (FRCP 26-37) and the EU's data-minimisation principle. France's Loi 80-538 (the French blocking statute), updated by the 2022 décret, criminalises some forms of cross-border data production for foreign judicial proceedings. The In re Mercedes-Benz and In re NSV cases in US courts have set conflicting precedents on the weight to give the blocking statute. Cross-border legal data redaction is therefore not just a compliance question but a defence-strategy question: a properly redacted production may satisfy both the US court's order and the French statute.
2. The six legal document categories and what each redaction looks like
Different legal documents need different redaction patterns. The six below cover ~95% of what a litigation or transactional practice handles.
1. Contracts and engagement letters. The most common contract redaction workflow. Sensitive clauses (fee structures, termination triggers, change-of-control provisions, MFN clauses, exclusivity grants) and party identifiers are masked before the contract is shared with a third party (a buyer's counsel in M&A, a court in litigation, a regulator in a filing, an AI tool for review). The redaction must preserve the document's legal structure (sections, clauses, numbering, defined terms) while masking the sensitive content. To redact contracts for AI-assisted review, the structural preservation is non-negotiable because the AI's analysis depends on the clause structure.
2. Court filings and pleadings. Complaints, answers, motions, briefs, declarations, exhibits. The redaction is governed by the relevant procedural code — FRCP 5.2 in US federal court, the equivalent in state court, CPR Part 5 / Practice Direction 5B in England, the local court rules in EU civil-law jurisdictions. The risk: PII that survives in an exhibit attached to a publicly filed motion is on the public docket forever. The case Doe v. Cabrera 307 F.R.D. 1 (D.D.C. 2014) became one of the canonical examples of inadequate redaction in court filings.
3. Witness statements, depositions, and declarations. Witness anonymization is its own discipline. The witness's PII, identities of third parties named in the testimony, and any information that could allow re-identification through the witness's role or location must be considered. In whistleblower-protected matters (Directive 2019/1937 + national transpositions), the witness identity itself may be statutorily confidential. Witness statement redaction is also where the trade-off between completeness and protection is most acute — over-redaction destroys the testimony's evidentiary value, under-redaction exposes the witness.
4. E-discovery productions. Document review and production in litigation. The volume problem is severe (a mid-sized matter routinely involves tens of thousands of documents) and the redaction must be performed at scale with consistency across the production. The privilege log itself often becomes a redaction artefact — entries must describe withheld documents without revealing privileged content. E-discovery redaction (also searched as ediscovery redaction) is the use case where AI tooling provides the largest productivity gain and creates the largest shadow-AI risk simultaneously.
5. M&A data rooms and due diligence. The seller assembles a virtual data room with hundreds or thousands of documents (commercial contracts, employment agreements, IP assignments, litigation files, financials, IT inventories). The redaction question is what to expose at which round (initial NDA-protected access, second-round access for short-listed bidders, full access post-LOI). M&A data room redaction (also referenced as M&A due diligence redaction) is the area where bad redaction has produced multiple documented "I saw a name I shouldn't have" incidents that have led to deal collapse or post-deal disputes.
6. Regulatory filings and correspondence with authorities. Antitrust submissions, securities filings, GDPR breach notifications, data-protection authority correspondence, sanctions and AML disclosures. Each regulator has its own redaction expectations and its own confidentiality regime. The EU Commission's TR/03 redaction standards for merger filings differ from CMA's UK equivalent. National data-protection authorities accept different redaction formats for breach notification annexes. The pattern: a single document may need three different redaction passes for three different audiences.
The pattern across all six categories is identical to the patterns we covered in the pillar and earlier clusters: structured data needs structured detection, free-text data needs ML detection, and the audit certificate is the artefact that makes the operation defensible.
3. Privilege, GDPR, and bar conduct rules — how they interact
Four legal regimes apply simultaneously to AI use in legal practice. They are not aligned, and the interactions are not always intuitive.
Attorney-client privilege and legal professional privilege (LPP). US privilege analysis follows the Upjohn framework for corporate clients (Upjohn Co. v. United States, 449 U.S. 383, 1981). English LPP covers both legal advice privilege and litigation privilege (Three Rivers (No 6), 2004). The CJEU's Akzo Nobel (C-550/07, 2010), commonly cited as akzo nobel privilege case, excluded in-house counsel from LPP under EU competition law — a holding that has shaped European M&A and antitrust practice for fifteen years. The VEB v BP case (CJEU C-499/21, 2023) and the more recent Halet v Luxembourg (ECtHR Grand Chamber, 2023) have continued to refine the privilege boundary in EU enforcement contexts. The take-away for AI use: pasting privileged communications into a consumer LLM is, in most analyses, a third-party disclosure that destroys privilege — and the destruction is not curable by subsequent withdrawal of the AI prompt. Legal professional privilege EU is jurisdiction-specific; check the applicable regime before assuming a privilege is preserved.
GDPR Articles 5, 6, 9, 22, 28, and 30. Lawyers process personal data on behalf of clients (Art. 28 processor relationship), often including special-category data (Art. 9 — health, religion, criminal convictions Art. 10) and frequently make decisions that affect data subjects (Art. 22). The records of processing under Art. 30 must reflect the full picture of legal-practice data flows, including AI tool use. A 2025 EDPB statement on AI in legal practice (publicly anticipated based on the 2024 work programme) is expected to clarify the processor status of AI tools used by external counsel.
Bar professional conduct rules. ABA Model Rule 1.6 makes confidentiality strict-liability; the Comment to Rule 1.6 explains that a "lawyer who has reasonable cause to believe that confidential client information may be disclosed" by use of a tool must take "reasonable efforts" to prevent disclosure. State-bar opinions in 2024-2025 (California 2024-1, Florida 24-1, Michigan State Bar Ethics Opinion JI-159) have applied the rule to generative AI: informed client consent is required, supervision under Rule 5.3 applies to AI as a non-lawyer assistant, and competence under Rule 1.1 includes the duty to understand the AI tool's limitations. The CCBE 2024 Guidelines on AI Use by European Lawyers reach largely the same conclusions.
National blocking statutes. France's Loi 80-538 (the blocking statute) criminalises the production of certain commercial documents to foreign judicial authorities outside the Hague Evidence Convention channels. The 2022 décret modernised the regime and gave the SISSE (Service de l'information stratégique et de la sécurité économiques) a coordinating role. Italian law follows broadly similar logic through the antitrust and trade-secret regimes. Swiss law via art. 271 CP. The interaction with AI use: cross-border legal data routed through a US LLM is not just a GDPR transfer issue but potentially a blocking-statute violation. US blocking statute discovery is also relevant in reverse — US courts have authority to compel production from EU litigants notwithstanding the blocking statute, and the case-by-case balancing has produced inconsistent outcomes.
The compound effect is that a single AI workflow in a cross-border legal matter can trigger privilege analysis under three regimes, GDPR analysis under one or two, bar conduct rules under one or two, and a blocking-statute analysis under one. The redaction layer is the single technical control that addresses the cumulative risk by eliminating the underlying PII and confidential content before any AI processing.
4. Litigation: court filings, exhibits, and discovery productions
Litigation workflows have the most concentrated redaction risk and the strongest external rules. The four sub-workflows below cover the litigation lifecycle.
Pleading-stage redaction. The complaint or motion is being prepared. PII must be redacted under the relevant procedural rule before filing. FRCP 5.2(a) lists the categories for US federal filings; state-court rules vary; English courts require redaction of personal data and confidential information in publicly accessible parts of the bundle (Civil Procedure (Amendment No. 2) Rules 2024 codified existing practice). Italian filings under art. 71 CPC and the new processo civile telematico require redaction of any data not strictly necessary. The classic failure mode: an exhibit is attached to a public motion with PII that the pleading itself carefully omitted.
Discovery production redaction. The producing party reviews documents for privilege and relevance, redacts privileged material, and produces the redacted documents to the requesting party. The Sedona Conference Cooperation Proclamation Series provides best-practice guidance. Privilege log entries themselves must describe the withheld documents without revealing the privileged content — a redaction problem within the privilege log itself. Privilege log redaction is among the most failure-prone sub-workflows because the description balance is hard.
Witness deposition redaction. Deposition transcripts run hundreds or thousands of pages and contain PII, third-party names, sensitive substantive content, and frequently inadvertent privileged disclosures. The redaction must be applied consistently across the transcript while preserving line-and-page citation. Public-access redacted transcripts (used in summary judgment briefing) are a separate workstream from sealed full transcripts.
Trial preparation. Trial exhibits, witness lists, and demonstratives may be filed publicly or used in open court. Redaction here is the last line of defence before the matter becomes public record. The case United States v. Manafort (D.D.C. 2018) became a textbook example of inadequate sealing when the table of contents of a sealed filing was published with section titles that revealed the underlying confidential information. Sealing court records is therefore a partial protection only; the redaction must survive future de-sealing.
Litigation hold and document preservation. The duty to preserve relevant documents arises at the moment litigation is reasonably anticipated. Litigation hold redaction — masking PII from documents retained under a hold but accessed for non-litigation purposes — is a specific sub-workflow. The legal hold redaction software that ships with most e-discovery platforms covers the core, but the redaction-layer interaction with the hold is firm-specific. Deposition prep also brings the redact deposition transcript step into the workflow. Public access court records (PACER in the US, equivalent in EU jurisdictions) make this work visible to opponents and the press.
The architecture for all four sub-workflows is the same: detect PII and privileged content, surface to a human reviewer for confirmation, apply masking or blackout, generate audit certificate, route to filing or production system. The human review step is materially more important for litigation than for any other use case because the consequences of an undetected miss are typically irreversible — once a document is on a court docket or in a production set, recall is rarely possible.
5. Contracts and M&A data rooms: where AI helps and where it leaks
Contract review and M&A due diligence are the workflows where AI tools have produced the largest productivity gains and the largest risk surfaces simultaneously. Three sub-workflows.
Contract review at scale (Contract review AI). An incoming commercial contract is fed to an AI tool for clause-level analysis (extraction of key terms, comparison to standard playbook, risk flagging). Without a redaction layer, the AI sees party names, executives' names, deal economics, sometimes trade secrets. With a redaction layer that masks the parties and the consideration figures, the AI gets a clean structural view, the lawyer reads an output that references placeholders, and the restoration is local. AI contract review GDPR compliance hinges on this redaction layer being upstream of the AI invocation.
M&A data room preparation. The seller's counsel assembles thousands of documents and applies tiered redaction by access level. The redaction tool must understand contract structure (preserve clause numbering when masking party names), preserve calculation logic (do not corrupt financial models), and produce an audit certificate per document. The buyer's counsel uses an AI tool for analysis of the data room; the redacted content the buyer sees is what the seller has decided to expose at that round. M&A data room redaction is also the area where "what was disclosed" becomes a litigated question post-closing, and the audit certificate becomes the primary documentary defence.
Counterparty diligence. The firm reviews the counterparty's KYC pack, beneficial owners, related parties, sanctions screening. Personal data of individuals who never consented to disclosure to your firm flows through this workflow. Redaction is needed both when the data is fed to AI tools and when summaries are shared internally for risk assessment.
The architectural pattern is the same as litigation: redact before the AI sees the document, keep the placeholder map locally, restore on output. The key tuning for contracts is structural preservation — a redactor that breaks the numbered clause structure destroys the AI tool's ability to reason about the contract's logic. Redactic preserves DOCX structure run-by-run and XLSX cell-by-cell, which is what makes the contract review workflow operationally viable. To redact contracts reliably for AI review, the structural preservation must be the default, not an opt-in.
6. The shadow AI problem in law firms — and the Mata v. Avianca precedent
Shadow AI in law firms takes specific forms that differ from the patterns we saw in HR or accounting.
The research shadow workflow. An associate uses ChatGPT to research a legal point at 11pm. The case-citation hallucination problem documented in Mata v. Avianca, Inc. (S.D.N.Y. Case No. 22-cv-1461, June 2023) — where ChatGPT invented six non-existent cases that the lawyer cited to the court — became the textbook example. Sanctions of USD 5,000, mandatory disclosure to the court of the AI use, and the matter referred to the state bar. Subsequent cases (Park v. Kim 91 F.4th 610, 2d Cir. 2024; the 2024-2025 series of state court sanctions orders against lawyers who cited hallucinated cases) have continued the trend.
The contract-summarisation shadow workflow. A junior associate uses a consumer LLM to summarise an inbound contract before sending it to the partner. The model sees party names, business terms, sometimes trade secrets. The partner does not know AI was used. The client did not consent. The privilege analysis depends on whether the consumer LLM's terms of service create a confidentiality protection sufficient to preserve privilege — typically they do not.
The brief-drafting shadow workflow. A senior associate uses ChatGPT to draft a section of a brief. The draft contains arguments based on facts from the client's matter that the associate provided as context. The matter facts are now in the model provider's logs. Some bar opinions treat this as a confidentiality breach by itself (irrespective of whether the LLM trained on the input).
The translation shadow workflow. A litigator working on a cross-border arbitration uses a free translation tool to render a German witness statement into English for internal use. The witness's identity, the substantive testimony, and any third-party names go through the free tool's compute. The legal-practice translation use case is among the most insidious because the lawyer perceives translation as a routine, non-substantive task.
The fix is the same architecture we have described in the pillar and earlier clusters, with two legal-specific tunings. First, the privilege-aware policy: the redaction layer must flag any content that appears to be privileged (client-lawyer communications, work product, litigation strategy) and either block the AI invocation or require an additional confirmation. Second, the citation-integrity guard for AI research: any case citation an AI produces must be verified against an authoritative database (Westlaw, Lexis, EUR-Lex, BAILII) before use. ChatGPT lawyers use that ignores either guard reproduces the Mata failure pattern.
Two widely cited telemetry studies of sensitive data reaching AI tools report shares that differ by a factor of about fifteen (2.6% and 39.7%), for reasons explained in our review of shadow AI statistics, so any figure quoted to a partner or a client should come with its source and its definition.
7. The redaction workflow tuned for legal documents
The 5-step workflow from the pillar guide applies to legal documents with three legal-specific tunings.
- Upload (with privilege-aware classification). Accept native formats: DOCX (contracts, briefs), PDF (filings, exhibits), XLSX (financial schedules, data-room indices), email exports (PST/MBOX with attachments for e-discovery). On upload, classify by matter and by privilege tier; route privileged documents to a separate workflow with stricter controls (no AI, no external sharing, fuller audit trail).
- Detect (with legal cascade). Dictionary lookup on the matter-specific terms (party names, witnesses, custodians, key executives the firm has previously masked), national regex for jurisdiction-specific identifiers (US SSN, EU national IDs, IBAN), multilingual NER for residual names and organisations, and OCR for scanned documents. The legal-specific addition: a privilege-flag pass that looks for indicators of privileged content (attorney sign-off blocks, "privileged and confidential" headers, signature blocks of inside or outside counsel).
- Review (with category grouping). Show detected spans grouped by category. For litigation use cases, also show by relevance category (PII, third-party identifier, privileged content, trade secret). The reviewer typically wants to verify the privilege log entries first, the PII pass second, the substantive content third. The grouped view also feeds the privilege log generation directly.
- Apply (mode by destination). Court filings and public productions: irreversible blackout. AI-assisted review: reversible masking with local restoration. M&A data room: tiered (the same document may have different redactions at different access levels). The tool must support all three from the same upload.
- Certify and route. Generate the SHA-256 redaction certificate. For court filings, include the certificate with the filing as supporting documentation of the redaction process. For productions, attach the certificate to the privilege log. For M&A, the certificate is the post-closing evidence of what was disclosed at which round.
A concrete example. A litigation team uploads a 240-page deposition transcript ahead of summary judgment briefing. The dictionary contains the matter-specific witness and counsel rosters. The regex catches PII (4 SSNs in financial exhibits attached to the transcript, 6 dates of birth, 12 phone numbers). The NER catches 38 third-party names in the substantive testimony. The privilege-flag pass catches 3 sections that appear to contain privileged work product. The review groups the 63 detected spans by category. The reviewer accepts the PII masks, accepts the third-party redactions, escalates the 3 privileged sections to the supervising partner. The export is two versions: a public-redacted version for the summary judgment exhibit, and a sealed version with only PII redacted. The certificate logs both versions, with separate hashes and category counts. Total time including the partner escalation: under 30 minutes.
8. Frequently asked questions
Does pasting a contract into ChatGPT destroy attorney-client privilege?
In most analyses, yes. The consumer ChatGPT terms of service do not create a confidentiality protection that prevents the disclosure from being a third-party disclosure. The 2024-2025 state-bar opinions (California 2024-1, Florida 24-1, NY State Bar Opinion 1240) have aligned on this. Some enterprise LLM tiers with no-training, no-retention contracts may preserve privilege, but the analysis is contract-specific and the safe default is to assume privilege is destroyed.
What does the Mata v. Avianca case actually require lawyers to do?
The case did not establish new rules; it applied existing FRCP 11 and FRCP 26(g) standards to AI-assisted legal research. The operational takeaway is that any case citation produced by an AI tool must be verified against an authoritative database before being relied on in a filing. Multiple subsequent sanctions orders have confirmed this is the standard. Court rules now in many federal districts require disclosure of AI use in filings (NDOK, EDTX, and others).
Can I use AI to redact a deposition transcript automatically?
Yes, but with the human-review step in the loop. The recall-biased detection will over-redact intentionally; the lawyer must review the suggested redactions and accept, reject, or extend each one. End-to-end automation without human review is not compliant with most bar conduct rules under the supervision and competence duties. Redactic's manual-review console is designed for this workflow.
How do I handle a US discovery request that conflicts with the French blocking statute?
The honest answer is "it depends and you need French counsel". The operational answer is that a properly redacted production that removes the trade-secret content the blocking statute protects, while preserving the relevance the US court needs, is often the negotiated outcome. The redaction certificate is the evidence both sides need: the US court sees that the production happened, the French regulator sees that the protected content did not cross the border.
Does the GDPR apply to a US litigation involving European witnesses?
Yes, to the extent that personal data of EU data subjects is processed in connection with the litigation. The processing is typically lawful under Art. 6(1)(c) (legal obligation) or Art. 6(1)(f) (legitimate interests), but data minimisation under Art. 5 applies. The Sedona Conference International Principles on Discovery, Disclosure and Data Protection provide a framework; the redaction layer is the technical control that implements the minimisation.
What is the right format for an M&A data-room redaction certificate?
SHA-256 input hash, output hash, category counts, mode (reversible/irreversible), timestamp, and a reference to the access tier (round 1 NDA, round 2 LOI, etc.). The certificate should not contain any content from the redacted document. Some buyers' counsel ask to receive the certificate as part of the seller's representations; the format that Redactic produces is suitable for this.
Does the EU AI Act apply to AI tools used by lawyers?
The AI Act does not classify most legal-practice AI uses as high-risk under Annex III. The high-risk categories that touch legal practice are Annex III(8)(a) (AI in administration of justice and democratic processes) and Annex III(5)(d) (financial-services risk assessment). For most law-firm AI use, the AI Act's transparency obligations (Art. 13), deployer obligations (Art. 26 when applicable), and AI literacy obligation (Art. 4, effective Feb 2025) apply. These are operationally less demanding than the high-risk regime but still require documented AI literacy training and basic transparency to affected parties.
Should I redact the names of opposing parties' employees in a discovery production?
Generally not — they are typically discoverable as part of the substantive matter. But you may need to redact PII of those employees beyond their names (SSN, DOB, home address, salary information unrelated to the case) under FRCP 5.2 or its state-court equivalent, and under the GDPR if EU data subjects. The privilege log entry for any document withheld in full needs to describe the document without revealing privileged content.
How long should the redaction certificate be retained?
For the same period as the matter itself, plus the statute-of-repose buffer. For litigation matters, typically the file retention period set by the firm's record-keeping policy (commonly 7-10 years). For M&A matters, typically the duration of the representations and warranties insurance plus the survival periods. Certificates contain no PII, so retention raises no privacy concerns of its own.
Where does Redactic store legal documents?
Nowhere persistent. Input and output documents live in Redis with Fernet encryption and short TTL during the job, destroyed by the worker's finally block. The application has no database model for document content. The redaction certificate stores hashes and counts only. Compute and metadata storage are EU (Hetzner Helsinki + Cloudflare R2 EU jurisdiction). The sub-processor list and DPA are public on the trust pages. For legal practice the same architecture applies; nothing about the legal use case bypasses the no-trace guarantee.
A final note on the AI productivity trade-off. The legal profession is being told that AI tools will produce 30-50% efficiency gains across contract review, e-discovery, and brief drafting. The studies that support those numbers are real; the productivity gains are real. The catch is that the gains are conditional on the documents the AI sees being properly redacted first. A firm that has not built the redaction layer captures the AI productivity gains at the cost of compliance and privilege exposure; a firm that has built the redaction layer captures the same gains without the exposure. The architecture exists; the regulatory and bar landscape is settled enough to plan against; the build-vs-buy choice favours buy because the operational discipline around the model is most of the cost.
If your practice handles contracts, court filings, witness statements, e-discovery productions, M&A data rooms, or regulatory submissions — and it does — every one of those documents currently moves through email, drives, and consumer AI tools without a redaction layer in front of them. The cost of doing nothing is not abstract: it is the next Mata v. Avianca sanctions order, the next state-bar complaint, the next privilege-destruction inadvertent disclosure, the next blocking-statute investigation, and — increasingly — the next professional indemnity carrier rejecting a claim because the AI use was undocumented. The architecture exists, the regulatory clock is running, and the redaction layer pays for itself in the first matter that goes to trial. Start with the contracts and court filings on your top three matters this month.
For the foundational architecture see the pillar guide; for HR-adjacent workflows see the HR cluster; for accounting workflows see the accountants cluster. To see Redactic on a sample contract or pleading, the redactic.eu landing includes a live demo.