A renewal pitch landed in my inbox this week from one of the two big legal research incumbents. The offer: their research platform now has AI built in — named frontier models, specialized legal workflows, drafting and analysis and document review, all “within the secure environment” of the platform. Nine months free if I sign before the end of the month.
I wrote back a reply that I suspect the salesperson has started seeing more often, and that I want to unpack here because the reasoning behind it is the whole ballgame for how firms should be buying legal technology right now. The reply, lightly cleaned up:
I have zero interest in accessing an AI assistant inside your environment. You have no access to outside APIs or tools, so that is a nonstarter. I do, however, have a huge interest in allowing my AI environment to access you.
That is the entire dispute, in two sentences. Not which model. Not whose “skills.” The direction of the integration.
The assistant is the environment now
Here is the assumption buried in every walled-garden AI pitch: that your work happens inside their product, so the assistant should live there too.
That assumption was defensible in 2023. It is wrong now, and it is worth being precise about why.
An AI assistant that runs in your environment — on your machine, under your control — is not a chat window. It is a working layer that sits above every tool you use. Mine reads the file system where my matters live. It queries my practice management system. It reads and labels the firm’s email. It pulls Minnesota statutes as raw HTML from the Revisor’s site because summarizers paraphrase and paraphrase is not good enough for quoted text. It drafts in Word from the firm’s templates, checks its own citations against archival sources, and runs the same verification gate before anything leaves the building. Each of those capabilities compounds the others, because they all live in one place and share one context.
Now consider the walled-garden version. The assistant inside a research platform can see exactly one thing: the research platform. It cannot see the matter file the research is for. It cannot see the draft the authority is going into. It cannot check the calendar, cannot read the demand letter you are responding to, cannot know that the client agreement caps the fee. It is an assistant with one book open and amnesia about everything else in your practice — and every hour you spend inside it is an hour spent teaching your workflow to live inside a subscription.
The integration only works in one direction. A tool can join an environment. An environment cannot fit inside a tool.
The plumbing for this already exists
None of this requires the vendor to invent anything. There is an open standard — the Model Context Protocol — for exactly this: exposing a service’s search, retrieval, and data to an AI assistant in a controlled, auditable way. It is the same standard my practice management connector speaks, the same one this section has written about before. Implementing it is weeks of engineering, not years.
And the proof that it works for legal research specifically is not hypothetical, because the nonprofit side of this industry already did it. CourtListener, run by the Free Law Project, exposes its full corpus — case law, federal dockets via RECAP, oral arguments, judges — through a documented REST API that anyone can build against. My firm connects to it the same afternoon-project way we connect to anything else, and the assistant uses it the way an associate uses a library: goes in with a question, comes back with an answer and a citation you can check.
That is what I mean when I tell the salesperson firms are switching. Not that a nonprofit’s database matches the incumbents shelf-for-shelf — it does not, and I will be specific below about where it falls short. The point is that the architecture everyone is switching to is the open one, because it is the only architecture in which the assistant can do its job. The research service that participates in that architecture gets used. The one that requires my assistant to stop, hand off to a human, log into a browser, and re-type the question does not — and “has AI inside” does not fix that, any more than a branch library installing a librarian fixes the fact that the books cannot leave the building.
What we actually did without them
Concreteness is worth more than argument here, so: over the past two days, this firm researched, drafted, and published thirty long-form articles on Minnesota law — family law, tort, business torts, consumer protection, expungement, construction payment. Every statutory quotation was pulled from the Revisor’s raw HTML. Every case was read in full from the Caselaw Access Project’s archival scans or the Minnesota Judicial Branch’s own slip opinions, with pin cites resolved against star pagination, not guessed. Recent session laws were read in struck-and-underscored form to catch three separate 2026 amendments that the codified statute pages did not yet display. Federal statutes came from the U.S. House’s official code site; federal regulations from the eCFR; Supreme Court quotes from the Court’s own prints. And then an independent verification pass re-derived all of it before anything published.
Marginal research cost from commercial databases: zero. Not because we are cheap, but because for primary authority — what the statute says, what the court held, what the session law changed — the free and archival sources are not a compromise anymore. They are the ground truth the commercial databases are themselves built on, and an assistant that can read them directly, at machine speed, in bulk, erases most of what the subscription was paying for.
Most. Not all.
The sliver they still own
Honesty about the remainder, because it is real and it is exactly what I would pay for if it were sold correctly.
The one thing our entirely-primary-source pipeline cannot do is subsequent treatment. When we verify an authority, we can confirm it exists, confirm what it held, confirm the quote is verbatim — and we cannot confirm that it was not reversed, superseded, or abrogated last Tuesday. There is no free citator. A citator is genuinely hard: it is editorial judgment applied at corpus scale, continuously. KeyCite and Shepard’s are the two incumbents’ crown jewels, and they have earned that status.
Add the secondary literature — treatises, practice guides, the accumulated editorial layer — and you have the honest inventory of what a Westlaw or Lexis subscription still uniquely provides in 2026.
So the renewal decision is not “do they have AI.” It is: is the citator-and-treatise sliver worth the full subscription price when it is locked behind a user interface my assistant cannot visit? Because that is what is actually on offer. The workflow the pitch imagines — a lawyer, in a browser, on their platform, chatting with their assistant — is not my workflow, and increasingly it is not anyone’s. My workflow is an assistant that touches the research service a hundred times in an afternoon, in two-second increments, mid-draft, mid-verification, mid-gate. Priced per seat and gated behind a UI, the sliver is not worth it. Priced like infrastructure — metered, per call, accessible to the tools that actually do the work — it would be irresistible.
Which brings me to the conclusion I want to state as plainly as the prediction pieces in this section state theirs, so it can be checked later:
Dropping Westlaw and Lexis is the rational move for a small firm — unless and until they sell low-cost API access. Not AI inside their walls. Not a copilot in their browser. A library card my assistant can hold: metered access to the citator and the editorial layer, at a price that reflects that the marginal cost of answering an API call is fractions of a cent. The first incumbent to sell that wins the agentic era, because every firm building what this firm has built will route through them without a second thought. Every year neither of them sells it, the free corpus gets deeper, the verification tooling gets better, and the sliver gets thinner.
The security objection, inverted
The pitch’s strongest-sounding word is “secure” — the AI runs within the secure environment of the platform. It is worth noticing that this inverts the actual confidentiality analysis.
A research query is very close to the least sensitive thing my firm emits. “Minnesota conversion elements punitive damages” discloses nothing a docket search would not. What is sensitive is everything the walled-garden workflow asks you to bring inside: the client’s documents, for review; the draft, for analysis; the facts, for research memos. The pitch says your data is safe because the AI lives here — but the only reason my data would be there at all is that their architecture requires me to carry it in. The assistant that runs in my environment sends the vendor nothing but queries. The assistant that runs in theirs needs my client file to be useful.
“Secure environment” is not a security argument. It is a data-gravity argument wearing one as a costume.
What to do at renewal
For the firm looking at its own renewal notice, the practical sequence:
- Inventory what you actually use. Pull the usage logs. In most small firms the overwhelming share of sessions is primary-authority lookup — a statute, a case, a quick elements check. All of that is available free, at API speed, today.
- Stand up the free layer first. CourtListener’s API for case law and dockets. Your state legislature’s site for statutes and session laws — raw HTML, not summarized. The Caselaw Access Project’s archives for historical reporters. Wire them to your assistant through MCP and use it for a month while the subscription still runs. The overlap will tell you what the subscription is really for.
- Isolate the citator. Whatever remains is mostly treatment-checking. Price that honestly: a per-lawyer Fastcase-tier plan (many state bars include one as a member benefit), or the smallest incumbent tier that includes the citator, treated as a verification step rather than a research home.
- Then ask the salesperson one question: can my environment access your service — not your AI, your data — through an API, and what does it cost? Their answer is the renewal decision. As of this week, from one of the two incumbents, the answer I have is silence on the API and nine free months on the walled garden. That is an answer too.
The access-to-justice thread that runs through everything in this section runs through this too, and it is not decorative. The reason a solo practitioner can now verify law like a firm with a research department is that the law itself — the statutes, the opinions, the session laws — was pried into the open by people who believed public law should be publicly computable, and the reason that matters is that every dollar of subscription overhead a small firm sheds is a dollar that no longer has to come from a client who could barely afford the representation in the first place. The incumbents can join that architecture and get paid fairly inside it, or they can keep selling seats to a workflow that is disappearing. But the direction of the integration is settled. The assistant is not moving into the library. The library is going to have to learn to make house calls.
Sources
- Model Context Protocol — the open standard for connecting AI assistants to external tools and data sources
- CourtListener REST API documentation — public, documented API over case law, RECAP federal dockets, oral arguments, and judge data
- Free Law Project — the 501(c)(3) nonprofit operating CourtListener and RECAP
- Caselaw Access Project — archival scans of the printed reporters, served as static files
- The thirty-article verification run described above is this firm’s own work over 2026-08-18–19; its per-article source manifests and independent gate reports are maintained in the site’s repository. The vendor email quoted in reply is paraphrased from real correspondence; no vendor employee is named.
Commentary on legal technology and the business of law — the opinions are the author’s. Not legal advice, and not a recommendation for any particular firm’s research stack: what a firm can safely drop depends on its practice mix, and treatment-checking (a citator) remains a required verification step this firm currently sources separately. No client information appears in this article. Questions about anything here: Send us a message or 612-470-6529.