The Vendor Conversation Every Firm Is Having Right Now
In 2026, Anthropic and OpenAI launched legal-specific products: Claude for Legal and Codex for Legal. Both platforms were previously consumer-focused. That shift is worth sitting with for a moment, because it changes the shape of every procurement conversation a firm will have this year.
I work with firms that have been evaluating legal AI tools for the past two or three years. The question I hear most often is some version of: "Should we go with a purpose-built legal tool or just use one of the big foundation models?" That question used to have a reasonably clean answer. Now it doesn't.
How the Market Was Structured Before
The legal AI market had a rough but legible shape: general-purpose foundation models like Claude, ChatGPT, and Gemini on one side, and purpose-built legal tools like Harvey AI and CoCounsel on the other. The foundation models competed on price and raw capability. The purpose-built tools competed on citation verification, ethical walls, and data confidentiality. Firms had to choose a lane.
That structure made sense. It also made evaluation manageable. You knew what you were comparing.
When Anthropic and OpenAI move into legal-specific products, they're attempting to occupy both sides at once: the brand recognition of a foundation model with the workflow specificity of a legal tool. That's a rational commercial move. It also deserves scrutiny.
The Confidentiality Question Is Architectural, Not Cosmetic
Purpose-built legal AI tools were designed with ethical walls and data confidentiality features from the ground up. General-purpose models, even when repackaged for legal use, were not built around law firm requirements.
That's not a minor distinction.
A tool's confidentiality controls are either baked into its architecture or they're not. Renaming a product and adding a legal-facing interface doesn't retroactively change how data is handled at the infrastructure level. For firms working on sensitive client matters, M&A transactions, or litigation with significant exposure, that gap carries real compliance weight. The question to put to any vendor, including Anthropic and OpenAI, is not "do you have confidentiality features?" but "how were they built, and what third-party validation exists?"
What Ethical Walls Mean in Practice
Ethical walls in legal AI aren't just about keeping client data off public training sets. They're about preventing matter-level data from crossing between clients within the same firm's environment. A conflicts-of-interest scenario where two clients are adverse to each other is exactly the situation where architectural controls matter. Consider a firm handling both sides of a transaction in different practice groups: the question is whether the AI environment enforces information barriers the same way the firm does, or whether it assumes a single unified knowledge base.
Purpose-built tools were designed to answer that question. Foundation models adapted for legal use may or may not be. Ask before you sign.
Citation Accuracy Is a Liability Question, Not a Quality Preference
The second pressure point is citation verification. Purpose-built platforms offer citation verification that foundation models don't match.
In legal work, a fabricated or unverified citation isn't a formatting error. It's a professional liability. Attorneys have faced sanctions for submitting AI-generated briefs containing citations to cases that don't exist. The verification burden, when it falls on the associate or the partner reviewing the work, adds time and risk back into the workflow that the tool was supposed to reduce.
Purpose-built legal AI platforms handle citation verification internally. That's a core feature, not an add-on. When Anthropic and OpenAI enter this space with Claude for Legal and Codex for Legal, the right question is specific: how does citation verification work in your legal product, and who carries the risk when a citation is wrong? Get that answer in writing.
The Verification Burden Shifts Depending on the Tool
Firms using foundation models for legal research take on a verification burden that purpose-built tools handle internally. That's worth quantifying for your own context. If associates are spending time verifying AI-generated citations on every research task, the time savings from the AI layer shrink considerably. The math on ROI changes depending on where the verification work lands.
More Players, More Pressure, More Clarity Required
The entry of Anthropic and OpenAI into legal-specific products makes the vendor market more crowded and more competitive. That creates useful pressure. It doesn't simplify the decision.
None of this is an argument against Claude for Legal or Codex for Legal. It's an argument for articulating your requirements before any sales conversation starts. The firms that get the best outcomes from these tools aren't the ones that move fastest. They're the ones that showed up with specific questions.
Here's a short version of the framework I use with firms:
First, what confidentiality architecture does your client base demand? Not what you can get away with, but what your clients would expect if they asked directly. Second, what is your acceptable risk threshold on citation accuracy, given the matter types you handle? Third, which specific workflows are you trying to change, and which tools were designed for exactly those workflows, not adapted for them after the fact?
The vendor market will keep shifting. Anthropic and OpenAI will not be the last general-purpose players to move into legal. The firms that build a clear internal standard for what they need will be able to evaluate each new entrant against that standard, rather than starting from scratch with every product launch.
When a sales team walks in this year, you want to be asking the questions, not answering them.
FAQ
What is the difference between Claude for Legal and purpose-built legal AI tools like Harvey AI?
Claude for Legal is Anthropic's legal-specific product launched in 2026, adapted from a general-purpose foundation model. Purpose-built legal AI tools like Harvey AI and CoCounsel were designed from the ground up with features like ethical walls, data confidentiality controls, and citation verification built into their architecture. The core distinction is that general-purpose models repackaged for legal use were not originally built around law firm requirements, whereas purpose-built tools were. For firms handling sensitive client matters, that architectural difference carries compliance weight that a product rename does not resolve. The right approach is to ask vendors directly how their confidentiality controls were built and what third-party validation exists.
Why does citation verification matter when choosing a legal AI tool?
In legal work, a fabricated or unverified citation is a professional liability, not a formatting issue. Purpose-built legal AI platforms offer citation verification that foundation models do not match, handling the verification process internally rather than passing that burden to the attorney reviewing the output. When firms use foundation models for legal research, associates or partners must verify AI-generated citations manually, which reduces the time savings the tool was supposed to provide. The question to put to any vendor, including Anthropic and OpenAI with their legal products, is how citation verification works and who carries the risk when a citation is wrong. Getting that answer in writing before signing is a reasonable precaution.
How should law firms evaluate AI tools now that Anthropic and OpenAI have entered the legal market?
The entry of Anthropic and OpenAI into legal-specific products with Claude for Legal and Codex for Legal makes the vendor market more crowded and more competitive. Firms that get the best outcomes are not the ones that move fastest, but the ones that arrive at vendor conversations with specific requirements already defined. The key questions are: what confidentiality architecture does your client base demand, what is your acceptable risk threshold on citation accuracy given the matter types you handle, and which specific workflows are you trying to change. Firms should evaluate each new entrant against that internal standard rather than restarting the evaluation from scratch with each product launch.
What are ethical walls in legal AI and why do they matter?
Ethical walls in legal AI refer to information barriers that prevent matter-level data from crossing between clients within the same firm's environment. They are particularly important in conflicts-of-interest scenarios, such as when a firm represents clients who are adverse to each other in different practice groups. Purpose-built legal AI tools were designed with these controls built into their architecture from the start. General-purpose foundation models adapted for legal use may or may not enforce information barriers at the same level, and that gap carries real compliance weight for firms handling sensitive client matters. The evaluation question is not whether a vendor has confidentiality features, but how those features were built and whether they enforce barriers the same way the firm's own policies do.
Are Claude for Legal and Codex for Legal good choices for law firms?
Claude for Legal and Codex for Legal are not automatically wrong choices for law firms, but the entry of Anthropic and OpenAI into the legal market makes it more important, not less, for firms to define their requirements before signing. The line between general-purpose foundation models and purpose-built legal tools is now blurring, which increases the complexity of the vendor decision. Firms should ask specifically about confidentiality architecture, citation verification processes, and which workflows each tool was designed to support. The competitive pressure created by more vendors entering the market is useful, but only if firms use it to get clearer on what they need rather than moving quickly on brand recognition alone.
