Superficially, law should be the perfect domain for large language models. The inputs are language, the outputs are language, precedent is textual, statutory interpretation is textual, and contracts are textual. Yet despite impressive demo-level capabilities from legal-AI platforms, enterprise transformation inside law firms and corporate legal departments remains surprisingly incremental.
The missing piece is that law is expressed in words, but legal work is not merely manipulation of words.
Early vs. Late Wittgenstein
Philosophy of language provides the cleanest explanation for this gap.
In the Tractatus Logico-Philosophicus, early Ludwig Wittgenstein argued that language gives propositions a logical structure through which the world can be represented: "The limits of my language mean the limits of my world." What can meaningfully be said is bounded by that representational structure. The naïve expectation for legal AI was fundamentally early-Wittgensteinian: if we can model language at scale, we can model the professional world represented through language.
In his later Philosophical Investigations, Wittgenstein complicates this view entirely. Meaning does not sit inside words like data inside containers. Meaning arises from use—from "language games," institutional practices, implicit conventions, and forms of life.
An LLM is extraordinarily good at the linguistic surface of law, but legal practice is a language game embedded in an institution.
Consider the chain of execution in legal practice:
Words → Doctrine → Procedure → Evidence → Authority → Incentives → Judgment → Action
Language models are strongest on the left side of this chain. Enterprise value lives almost entirely on the right.
The Contextual Gap
Analyzing a contract provision is a textual task; advising a client on it is an institutional move.
Consider the contrast between what an AI tool is asked to do and what a practicing lawyer actually does when reviewing an M&A document:
- The Model Task: Read 800 pages and identify change-of-control provisions inconsistent with the firm's standard template.
- The Lawyer Task: Determine whether this clause is commercially problematic for this specific client, given what the counterparty is likely to accept, what the partner promised the client yesterday, what occurred in the previous transaction, the current regulatory jurisdiction, how aggressive opposing counsel is acting, which issues are worth spending negotiating capital on, and whether the risk is grave enough to awaken the General Counsel at 11 p.m.
Most of that knowledge is absent from the documents being analyzed. It is distributed across people, systems, history, and tacit institutional memory.
Legal knowledge is highly textual, but legal work is highly contextual. LLMs solved much more of the first problem than the second.
The Capability vs. Institutionalization Gap
Current industry data highlights this exact boundary. A 2026 survey reported 69% individual generative AI usage among legal professionals, yet firm-level implementation stands at 46% for general AI tools and just 34% for legal-specific platforms.1 The primary obstacles cited were not raw intelligence or model accuracy, but security, ethics, privilege boundaries, and lack of trust.
To understand the language, you have to participate in the practice in which the language has meaning.
This gap persists because legal AI has primarily been designed around an individual user loop:
Lawyer → AI tool → Better answer → Lawyer
Real enterprise transformation requires a system architecture that connects the full institutional context:
Matter → Documents → Client Context → Institutional Precedent → Workflow → AI Reasoning → Human Decision → Approval → Execution → Audit Trail → Institutional Memory
Why "Better Models" Will Not Bridge the Gap
Doubling model intelligence tomorrow does not resolve the structural friction points that govern legal institutions:
- Permissions & Privilege: Strict ethical walls and confidentiality boundaries across clients and matters.
- Fragmented Stores: Unstructured document repositories without standardized matter metadata or source provenance.
- Verification & Liability: Unclear lines of professional accountability when AI outputs inform formal opinions.
- Economic Disincentive: Much of Big Law still monetizes billable hours. Recent American Bar Association commentary explicitly notes that AI-driven efficiency creates a direct tension with traditional law firm billing models.
| Layer | Domain | Model Capability | Current Enterprise Adoption |
|---|---|---|---|
| 1. Language | Read, draft, summarize, compare text | Very strong | High individual uptake |
| 2. Knowledge | Retrieve precedent, connect documents, apply doctrine | Increasingly strong | Selective integration |
| 3. Institution | Authority, workflow, incentives, accountability, action | Weak / Uncoded | Structural bottleneck |
Encoding the Language Game
A contract is not merely 80 pages of text. It is a strategic move in a game involving buyers, sellers, regulators, leverage, deadlines, and future liability.
The model understands the sentence. The system has to understand the game.
The next wave of transformative enterprise AI companies will not win simply by deploying larger foundation models. They will be the platforms that successfully encode the language games of institutions: authority, state, workflow, permissions, provenance, and action. The breakthrough from LLM to enterprise AI is fundamentally a system-design challenge, not a model-training one.
Footnotes
-
69% of legal professionals now use general-purpose generative AI tools for work, more than double the 31% reported in last year's survey. At the firm level, 46% have implemented general AI tools, and 34% legal-specific tools. 8am (MyCase / LawPay), Generative AI data from the 8am 2026 Legal Industry Report (2026-03-20). ↩