Seven Terms Every Legal-AI Data Buyer Has to Stop Nodding Through

Legal-tech teams building research assistants, contract analyzers, and case-law search buy data in a vocabulary half-inherited from software procurement — and the words that matter most in legal data are exactly the ones everyone nods through without defining. This is a working lexicon: seven terms, what they should mean when a data vendor says them, and the follow-up question that separates meaning from vocabulary.

The Lexicon

1. “Authority”
In legal, data isn’t just content — it’s a hierarchy of sources, and a research product that cites a blog post where a statute belongs fails its first real user. The follow-up: *what source tiers does your corpus distinguish, and is tier a field?* A legal dataset for AI training with source-type fields (statute, regulation, case, secondary commentary) lets your retrieval layer honor the hierarchy; a corpus of “legal text” with no tier field makes your model a very confident equal-opportunity misciter.

2. “Currency”
The law changes. A corpus frozen before the amendment is worse than absent — it’s confidently wrong. The follow-up: *what’s the update lag between a statute/case publication and its arrival in the corpus, and is each record version-stamped?* Version stamps turn “out of date” from a corpus-level indictment into a per-record join you can handle in retrieval. This is the same freshness-as-a-number discipline that governs every serious data category — cadence and timestamps in writing, “regularly updated” being the marketing phrase that means nothing.

3. “Jurisdiction”
The silent killer of legal-tech demos: the model that mixes jurisdictions without knowing it. The follow-up: *is jurisdiction a structured field — country, state/province, court level — or a guess from the text?* Structured jurisdiction fields are what let your product promise “answers valid where you practice” instead of “answers.”

4. “Provenance”
Your users are lawyers; the product’s citation is their professional liability. The follow-up: *can a citation trace to a specific captured source with a date?* Record-level provenance — source, capture timestamp, license scope per record, the fields Thordata’s dataset structure carries catalog-wide (standard datasets from about $0.25 per 1,000 records, sample-first) — is what turns a generated answer into a checkable one. An answer a lawyer can’t verify is a liability generator, not an assistant.

5. “Anonymization”
Case data and filings contain parties, minors, sealed-adjacent material. The follow-up: *what’s the redaction methodology, and can we audit samples?* Redaction that’s a pipeline stage with documentation is a spec; redaction that’s an adjective is a discovery request you’re pre-addressing.

6. “License scope”
“Licensed for legal data” is three words doing no work. The follow-up: *named uses — retrieval product, model training, commercial SaaS deployment — which are permitted, in writing?* The license schedule by use is now the category norm across serious dataset vendors; a legal-data quote without one is a quote from someone who hasn’t met your counsel yet.

7. “Coverage”
The least glamorous term and the one your roadmap actually depends on. The follow-up: *what courts, what years, what document types — with per-slice counts?* Coverage in legal is asymmetric by nature (federal deep, administrative thin, older decades spotty), and the vendor who documents the asymmetry is handing you the spec for what your product can honestly claim. “Comprehensive” is the word that should end the meeting.

Why a Lexicon and Not a Checklist

Because legal-tech buyers negotiate with counsel in the room, and the seven terms above are the vocabulary that conversation runs on. A team that walks in defining authority, currency, jurisdiction, provenance, anonymization, license, and coverage — with the follow-up questions ready — converts the vendor call from a demo into a specification review. Teams that don’t, buy on features and litigate on data.

The Pilot That Exercises All Seven

One jurisdiction, one practice area, ninety days: license the corpus slice, run your retrieval product against it, and score every failure against one of the seven terms. A miscitation is an authority failure; a stale answer is currency; a cross-jurisdiction blend is jurisdiction — the taxonomy turns your error log into a vendor scorecard and your next negotiation into documentation. And run the demand side in parallel: what lawyers and legal-ops teams actually search — visible through SERP monitoring at about $0.70 per 1,000 structured responses — is the product roadmap’s most honest input, because legal search vocabulary is precise in a way almost no other market’s is.

FAQ

Do legal teams buy web data at all, or just official repositories?
Both, deliberately: official sources for authority, web-sourced corpora for the surrounding reality — commentary, firm publications, regulatory news, secondary analysis — that makes a research assistant actually useful. The corpus problem is the same either way: the seven terms apply to whatever you buy, from whoever sells it.

What about our own document corpus — same lexicon?
Yes, pointed inward: your contract repository has a currency problem (templates drift from current law), a jurisdiction problem (governing-law clauses vary), and a provenance problem (clauses copied from where?). Running the lexicon against internal data before external data is the cheapest audit in legal tech.

How large should the pilot corpus be?
Practice-area-scoped, not volume-scoped: one jurisdiction × one area × 90 days of currency is a real product test; ten million records of mixed everything is a data swamp with a demo on top. Per-record catalog pricing means the scoped pilot is a few hundred dollars — the error-log taxonomy you build during it is worth more than the corpus.

Where does search data fit a legal product’s roadmap?
Three places: feature prioritization (which practice-area queries are rising — continuous SERP data crawling tracks it per market), competitive visibility (where your product surfaces against incumbent research tools), and content strategy (the queries lawyers type are the titles your content team should write). One feed, about $0.70 per 1,000 responses, all three uses — and the same query-level monitoring that shapes your roadmap doubles as the answer-layer visibility check every AI product now needs.

The lexicon’s one-line summary: legal-tech data quality isn’t a number, it’s seven defined words — and the buyer who defines them first writes the spec the rest of the market has to meet.