On February 10, 2026, Judge Jed Rakoff held that client chats with consumer-tier Claude are not protected by the attorney-client privilege. The ruling sits underneath every legal-AI procurement decision F500 GCs will make this quarter, including the one Anthropic announced on May 12.
On February 10, 2026, Judge Jed S. Rakoff of the United States District Court for the Southern District of New York issued an opinion that the legal-AI industry has spent three months trying not to talk about. The holding was narrow on its face and load-bearing in its consequence. Client interactions with consumer-tier Claude, the version sold under the Pro and Max subscription plans, are not protected by the attorney-client privilege. The summary published by Herbert Smith Freehills Kramer captures the reasoning in a single sentence: Anthropic, as a consumer-tier vendor, is not contractually bound to the standard of confidentiality that a third-party agent in the attorney-client relationship would be.
The doctrinal hook is the agent-of-the-attorney rule. Privilege survives when an attorney shares client confidences with a third party who is bound to the same duty of confidentiality the attorney owes. Paralegals, contract reviewers, retained experts, and translation services have all been folded into the rule under decades of case law. The binding is what does the work. Consumer Claude fails the binding. The standard subscription terms permit model training on user inputs unless the user opts out, retain conversation logs for vendor purposes, and grant Anthropic employee access for safety review. The court read those terms and concluded that no agency relationship of the kind privilege requires can attach to them.
Rakoff did not write a referendum on AI in legal practice. He wrote a referendum on contracts. The opinion explicitly distinguishes the enterprise tier, noting that enterprise deployments present a "materially different analysis" because the vendor in that case is contractually bound to the confidentiality standard. The HSF Kramer note carries that line forward. Enterprise master service agreements typically include zero data retention, no training on customer inputs, customer-controlled retention windows, and indemnification language that imports vendor confidentiality obligations into the attorney-client chain. The word that matters in that list is "typically." Not every enterprise contract carries all four. The ones that do restore the binding. The ones that do not, do not.
What the ruling produced, in operational terms, is a single new question that now sits underneath every conversation a lawyer has with an AI tool on a client matter. Which tier is the lawyer on. Pro. Max. Team. Enterprise. The answer is now the privilege question. Not a privilege question. The privilege question. A lawyer on a Pro account drafting a settlement memo with client facts in the prompt is, on Rakoff's reasoning, generating a discovery target. A lawyer on a properly negotiated Enterprise account doing the same thing has the binding in place.
On May 12, 2026, Anthropic announced Claude for Legal, the company's first product offering branded specifically for law firms and in-house legal teams. The launch post is a marketing document. It walks through workflow integrations, the partnerships with Harvey and a set of AmLaw 100 firms, and the case-law search and contract analysis features. The post does not name the Rakoff ruling. The post does not explain that Claude for Legal is sold across Pro, Max, Team, and Enterprise tiers and that the tier the customer buys determines whether the product carries the binding privilege requires. The post does not contain the words "privilege," "zero data retention," "ZDR," or "MSA."
A general counsel reading the launch post would be forgiven for assuming that a product named "Claude for Legal" comes with the privilege-protective contract terms a legal product needs. It does not, automatically. The product name is a marketing layer on top of the same underlying tier structure that produced the Rakoff problem. Buying Claude for Legal on the Pro tier puts the same data in the same training pipeline and the same retention window as buying consumer Claude on the Pro tier. The contract is the product. The brand is not.
Join 1,847 litigation leaders who get weekly intelligence on strategy, technology, and the data that matters.
The Rakoff ruling does not stop at the GC's own legal department. F500 general counsel sit at the top of an outside-counsel panel that can run to fifty firms across litigation, regulatory, IP, employment, and corporate work. Every one of those firms now has a tier question of its own. The April 21 Sullivan & Cromwell apology letter to the Southern District of New York covered roughly forty AI-generated errors across five filings in a single Chapter 15 proceeding. The takeaway from that letter was that even AmLaw top-tier firms do not have AI use under management-level control. The Rakoff takeaway is the same problem rotated ninety degrees. If a panel firm has lawyers running consumer Claude on the F500's matters, the GC has inherited the privilege risk through the panel firm. The discoverability sits with the F500. The contract the GC needs is one the GC does not control.
The 2026 outside counsel guidelines most F500 legal departments are running on were written before any of this was visible. They cover hourly rates, staffing requirements, billing protocols, conflicts disclosures, and document retention. They do not cover what tier of AI subscription the firm's associates use to draft the memos that get filed under the F500's name.
The Rakoff ruling is not a theoretical risk. It is a discovery target that is already cited in privilege motions in the SDNY and is being briefed by the plaintiff bar in venues outside New York. A general counsel reading this article this quarter can produce a defensible posture with three concrete moves.
First, pull the Pro and Max tier user list from your IT or procurement department for every individual seat across the legal department and any business unit whose staff touch privileged material. Not a sample. The full list. The output of this step is a defensible record of which seats need to be migrated to Enterprise tier under a privilege-protective MSA and which need to be terminated. Map the seats against ongoing matters.
Second, inventory which panel firms use consumer-tier Claude or any consumer-tier AI tool on your matters. Send the question in writing on firm letterhead. The firms that cannot answer the question are the highest exposure on your panel. The firms that answer the question with "we don't track that" are the second highest. The firms that produce a clean per-matter audit trail are the model the panel needs to converge on.
Third, issue an outside counsel guideline amendment that requires Enterprise-tier or panel-firm-issued ZDR attestation for any AI used on your matters. The language can be one paragraph. The attestation requirement, not the punitive enforcement, is the operative move. Asking the question on the record changes what the firm-side workflow has to look like.
The Rakoff opinion will be the cited authority in the first wave of privilege motions on AI-generated work product, and the cite will be in a discovery dispute on a matter the GC is already managing. The three moves above take a quarter to run. The discovery motion that forces them does not give the GC a quarter to respond.
If you want a specific readout on where your portfolio sits against the AI-privilege exposure this ruling opens up, the Litigation Sentinel Executive Briefing is a six-question diagnostic built for that purpose. About four minutes. Produces a specific readout, not a generic one.
litigationsentinel.com/briefing
The Executive Briefing takes 2 minutes and shows you exactly where the gaps are.
Take the Executive Briefing →Join 1,847 litigation leaders who get weekly intelligence on strategy, technology, and the data that matters.