Who Owns the Graph?
Capture the reasoning behind your decisions and your AI stops plateauing. Connect it, and things that used to take a lot of digging show up in a click. What gets asked far less often is where that context ends up, and who holds the keys. It is the least examined assumption in the stack.
Three things decide how good an AI agent is at its job: the model underneath it, the harness around it (the app or assistant you actually open), and the context it can draw on. Models are rented, and they improve on a lab's schedule, not yours. Harnesses get swapped constantly, and most organizations already run more than one. Context is the only layer you can own, and the only one that compounds instead of resetting every time a tool changes.
That stopped being theoretical this year. Anthropic launched Claude for Legal in May, with practice-area plugins that open with a setup interview about your team's playbooks, escalation chains and risk calibration. OpenAI followed in September with Astra for Law and a set of partner plugins. Both are serious products, and both ask your team to teach them how you work. So the ownership question gets practical fast. When your team answers that interview, where does the answer go?
Three places your context can live
In the vendor. The tool keeps your corrections, preferences and decision history in its own database, in its own schema. You can see it through the interface and usually nowhere else. If you leave, the documents come with you and the reasoning stays behind. In the model. Some context gets absorbed into an assistant's memory or a provider-managed profile. It feels like ownership, but you cannot open it, audit it or hand it to another tool, and when the provider swaps the model underneath you, you may not know what survived. Nowhere. The default. The reasoning lives in a senior partner's head or a chat message that explained an exception, and you rent your own judgment back from whoever remembers it.
There is a fourth option, and it is the one worth building toward: a governed record you own, which every tool reads from instead of keeping a private copy. The tools become interchangeable. The record does not.
The legal AI apps built on the big models don't fix this on their own. Many work across several models and pick the best one per task, which is real value, but model-agnostic is not context-agnostic. If an app keeps your playbooks in its own store, the lock-in has moved up a layer. Gartner put the stakes plainly in an April release: "control of enterprise context is economic power." It added that durable advantage comes from "controlled openness," and that vendors who rely on restriction risk being bypassed by orchestration layers enterprises trust more. Read that as a buyer. Vendors will be tempted to hold your context close, and the ones worth keeping will let you determine how your context should travel.
Picture a mid-sized firm that adopted an AI drafting assistant two years ago. Its associates corrected the output constantly, and each correction taught the tool something: this client never accepts a landlord termination right, that lender always wants the guarantor language in section 9. When the renewal arrives 40 percent higher, IT confirms the documents can move. The two years of corrections cannot. They live in the vendor's system as model behavior, with no export and no way to read them. The firm renews. Nobody chose to be locked in. They just never asked while it was cheap to ask.
You can buy AI tools, but you can't buy a well-structured corpus of institutional knowledge. Firms that have invested in it have a durable competitive advantage because their data is ready to be leveraged in ways competitors simply can't match.
- Eric Duncan, Legal Solutions Director, NetDocuments
Ownership is also where the value is
Lock-in is the defensive case. The offensive case is that context is where the return on AI sits, and NetDocuments' 2026 Legal Context Engineering Benchmark Report puts numbers on it. Written by Scott Kelly, VP of Product, it holds the model and the harness constant, changes only the context the agent can reach, and prices a correct answer with and without the Legal Context Graph. The test covers 300 questions across 10 real matters, built from 874 public filings and docket documents, none of them synthetic. On the frontier model, a correct answer fell from $0.68 to $0.36 with quality held almost level, about 48 percent less.
As pricing moves from all-you-can-eat to consumption, the question becomes what a correct answer costs. Context is the layer that carries forward across every model upgrade and every agent you adopt, so whoever holds it holds that margin.
The practical home for that context is the system that already holds the record, the permissions and the audit trail, which in most firms is the DMS. That is the design behind the NetDocuments Legal Context Graph. It works at three levels: what a document is and contains, how documents relate inside a matter, and where expertise and precedent sit across the firm. Existing permissions, ethical walls and audit controls travel with the knowledge, and MCP lets Claude, ChatGPT, Copilot, Gemini, Perplexity and legal specific apps like Harvey and Legora draw on it without anyone uploading files into a session. The AI comes to the document instead of the document going to the AI. When the context lives in the DMS and the tools connect to it, swapping a harness or a vendor becomes a configuration choice instead of a migration, and lock-in matters far less. The AI-powered graph is rolling out in private preview now.
The law firms getting this right aren't asking AI to change how they govern documents. They are asking AI to work within the governance they have already built and invested in.
- Brandon Nelson, Legal Solutions Director, NetDocuments
Five questions to ask before you sign
You don't need a longer RFP. You need these in your next vendor conversation, with written answers.
- Where is our context stored? In a system we control, in the vendor's database, or in the model itself?
- What happens when we switch tools? Do decisions, reasoning and relationships stay in a governed record the next tool can read, or do they leave with the vendor?
- Whose permissions apply? Do matter-level access rules and ethical walls govern what the AI can retrieve, every time?
- Can a second tool read it? If we ran two assistants side by side, would both draw on the same record?
- What is the evidence? Show the effect on answer quality and cost, with a method we can rerun ourselves.
Vendors with good answers give them quickly, and the hesitation tells you most of what you need to know. How it lands depends on where you sit. In private practice, matter history is a firm asset, and letting a vendor hold it privately means someone else can reprice your accumulated edge every renewal cycle. In corporate legal, it is continuity: context locked to one product is a turnover problem wearing a different hat. In legal operations, it is a governance line item you already know how to manage.
Start this week with the AI tool your team uses most, and answer the five questions about it. If you can't, that is your first finding. Then take an afternoon to map where your team's decision context lives today and how much of it would leave with the next departure, so you know whether you are holding it, renting it or losing it.
Once you ask who holds context and who can take it with them, the legal tech market looks different. Systems of record, point solutions, assistants and orchestration layers sort themselves by an answer that has nothing to do with feature lists. The next piece redraws the market map on exactly that basis.
