The AI Law Podcast-The law of AI and Using AI in Law
All Episodes
Legal AI Procurement: Research Platforms vs. APIs

Legal AI Procurement: Research Platforms vs. APIs

0:00|0:00

We analyze the architectural differences between commercial legal research platforms and primary-law APIs, examining how each layer impacts legal procurement and introduces distinct error modes. Drawing on independent benchmark findings and ABA Formal Opinion 512, we highlight verified hallucination rates and reinforce the litigator's non-delegable duty to verify citations.


Chapter 1

Two Categories, Two Buying Decisions

Erick S. Robinson

Two products can both be called legal AI and still answer completely different questions. Lexis+ AI with Protégé and Westlaw with CoCounsel Legal are research platforms. A lawyer logs in, asks a question, reads cases, checks citations, and drafts. A primary-law data API, such as Vaquill AI, is built for software. A developer sends a request and gets back statute, regulation, or session law text as structured data. No case law, no chat window. This episode is education, not legal or patent advice.

Dr. Angela Liu

And the central question for today is procurement. If a firm confuses those two categories, it buys the wrong thing. I'd add a caution on our sourcing, too. Much of the architectural description comes from Vaquill's own blog, and Vaquill sells an API. That's a vendor describing its own category, so treat it as a framing, not an independent finding.

Erick S. Robinson

Agreed. The vendor description does line up with how the products are positioned, so let's use it as a map. The research platforms bundle an editorial corpus, a citator, a generative layer, and workflow tools. Westlaw pairs CoCounsel Legal with KeyCite and agentic Deep Research. Lexis pairs Protégé with Shepard's and, per that source, more than 300 agentic workflows. Both are sold per seat through a sales conversation. Neither publishes seat prices for its AI tiers, so I can't give you a number, and anyone quoting one should show you the quote.

Dr. Angela Liu

Now the API side. It supplies the law itself, over six million sections according to the vendor, with section level citations. But it doesn't tell you how a court treated a statute. It doesn't give you a citator signal. It doesn't synthesize anything. If your team builds an internal tool on it, you've solved one layer, the source text, and you've inherited every other layer.

Erick S. Robinson

Which brings us to the evidence on errors, and it's the part that vendor marketing tends to blur. The Stanford and Yale study by Magesh and colleagues, published in the Journal of Empirical Legal Studies in 2025, tested commercial legal research tools built on retrieval augmented generation. It found they reduce errors compared with a general chatbot, but they don't eliminate them. The reported rates: Lexis+ AI hallucinated about 17 percent of the time, Westlaw's AI Assisted Research about 33 percent, and Practical Law's AI tool gave incomplete answers or refusals in over 60 percent of queries.

Dr. Angela Liu

Those numbers need care. The study measured specific versions of tools at a specific time, and vendors have updated products since. It's a snapshot, not a permanent score. Also notice the error type. Not just invented cases. The study's concern includes real authority that's misgrounded, meaning a real case cited for a proposition it doesn't support. A citation that exists will pass a cursory check.

Erick S. Robinson

That's why tool layer matters. An API grounds you in primary text, which narrows one error type, fabricated statutory language. A research platform adds case law and citators, which helps with treatment, but adds a generative synthesis step where misgrounding can enter. Different layers, different failure modes.

Chapter 2

What It Means for Litigators and Patent Work

Dr. Angela Liu

For litigators and patent practitioners, this gets concrete. In a patent matter, the proposition often lives in a file wrapper or prosecution history, or in the precise holding of a Federal Circuit decision. A tool can cite a real Federal Circuit case and still overstate what it held. No platform output substitutes for opening the document and reading the passage.

Erick S. Robinson

And the ethical rule doesn't move with the architecture. ABA Formal Opinion 512, issued in July 2024, places the duty of competent verification on the lawyer. The source we reviewed puts it plainly: the verification duty belongs to the lawyer on every citation, whatever the tool. That holds whether you subscribe to a platform or your engineers build on an API. The 2023 Mata v. Avianca sanctions turned on citations nobody opened, not on which tool produced them.

Dr. Angela Liu

So here are the recommendations. Firms: ask what problem you're solving before you compare products. A lawyer who needs a researched memo wants a platform. A team building an internal tool needs source data, and probably case law from somewhere else. Trial platforms on matters where you already know the right answer, and record the errors.

Erick S. Robinson

Vendors should publish what was tested, which version, and what the error categories were. Clients can ask outside counsel how AI output is verified, not whether it's used. Courts will keep judging the filing, not the software.

Dr. Angela Liu

Bottom line. Platforms and APIs aren't rivals, they're different layers, and neither one transfers your duty to check.

Erick S. Robinson

What remains unknown is plenty. Current error rates on the latest versions, independent testing of agentic workflows, and real enterprise pricing are all unpublished or unsettled. Until that changes, design verification into the workflow, and don't count on the license to do it.