Most everyone agrees legal AI splits into commodity capability and defensible differentiation. Almost all are looking for that defensibility in the wrong place.
TL;DR
A widely shared thesis says legal AI is settling into 70 percent commodity capability and 30 percent defensible differentiation. It is right about the shape and wrong about the source. The defensible 30 percent is not a better feature set - features get copied. It is the verification-and-validation layer: the discipline that proves the AI’s output can be trusted - that every citation, reference, reading of prior case law, and drafted argument has been cross-checked, is accurate, and can be stood behind. The bar is not “as good as a capable first-year associate.” The expectation is that it is closer to a perfectionist machine - reliably right, and able to show its work. That defensibility is about the output, and it is built into the platform, not bought with a model license.
What I learned building one
Building a legal practice platform - the kind that runs a firm end to end, every matter and every document, not a single-purpose contract tool - taught me the model was the least interesting decision. We spent months on model selection, skills, and agents, and in the end the technology was close enough that the choice barely moved the platform. What moved it was verification and validation. Could the platform prove its output was right? Could we reproduce a result from a week ago? When a decision on a document had to be defended, could the people who owned it stand behind the reasoning? That layer was not a feature we added near the end. It was the thing that made the platform trustworthy enough to use in production.
Who has to answer, and for what
Here is the part that gets miscast. When a lawyer flags, redlines, or changes a clause, someone may later ask why. The answer that matters is the legal reasoning, and an attorney stands behind it - not the machine’s internal logic. None of this means concealing the use of AI; disclosure obligations are real and a firm meets them. It means you do not hand a client the model’s reasoning as if that were the justification for what was decided. What they care about is whether the call was sound and who is accountable for it - a person. A throat to choke… or disbar. The platform earns its keep when it lets the people own and defend the work, with the reasoning intact and the AI’s role properly disclosed.
The 70/30 is about features. The defensibility is not.
A sharp commoditization thesis has been circulating in legal tech circles: vendors compress toward 70 percent commodity capability and 30 percent defensible differentiation. (fn. 1) The shape is right. But the 30 percent is not defensible because the features are better - vendors have similar features, and features get copied. It is defensible because of how the work - the output from the system - is verified and validated: whether the platform can prove, on any given output, that it is accurate and can be trusted. That discipline is not a line on a comparison chart. It is what carries the commodity 70 percent of the development across to the defensible 30 percent. Strip it out and the 30 percent is not differentiation; it is a claim that collapses the first time someone checks the citations.
Why the defensibility is in the output
Be precise about what makes this defensible. The verification-and-validation layer distinguishes, and records, what the model produced from what a person reviewed, corrected, and signed off on. This is not about patenting AI-assisted work - legal AI is not an invention-generation workflow, and the asset here is not a patent. It is output defensibility: the ability to stand behind a citation, a brief, or an analysis because it was checked and a qualified person put their name to it. File with a court and you still need a record that an attorney actually reviewed each citation and confirmed it is real and on point; that final review and sign-off is where the professional license earns its keep. The model is rented. The defensible work product - and the record of how it was verified and who stood behind it - is yours.
The audit trail is one piece, not the whole
None of this reduces to keeping an audit trail. The trail is one element. Reproducibility, quality, transparency, and a contemporaneous record of who reviewed and signed off all matter, and in some matters they matter more than the trail itself. Treating any one of them as the whole answer is how you end up with a platform that looks rigorous and fails under the first hard question.
What to ask before you buy or adopt
If a legal-AI evaluation skips the verification-and-validation conversation, you are paying a premium for differentiation and taking delivery of commodity. These three questions are the whole game: Can the platform prove an output is accurate and reproducible? Is there a contemporaneous record of who reviewed and validated the work? Can your people defend the reasoning for a decision on the merits, without leaning on the machine’s logic as the justification? Vendors who answer cleanly are selling the defensible 30 percent. Vendors who cannot are selling the 70 and pricing it like the 30.
In legal AI, the defensible 30 percent is the verification-and-validation layer - proof the output is accurate and someone qualified stands behind it, built into the platform, not bought with the model. The vendors who can show it are priced like the defensible 30 percent for a reason. Footnotes
(fn. 1) Pim Betist, Anthropic Just Climbed Up the Legal Stack, LinkedIn (2026), https://www.linkedin.com/posts/pimbetist_anthropic-just-climbed-up-the-legal-stack-ugcPost-7460561397173764096-yE0O/ (the commoditization thesis this piece builds on: that legal-AI vendors compress toward broadly available commodity capability plus a thin band of defensible differentiation).