Resources

The Legal AI Glossary

The vocabulary of legal AI, in plain English — the terms you'll meet in vendor pitches, and what they actually commit a vendor to.

How AI answers

Grounding: Building an answer from identified sources (documents, emails, statutes, case law) rather than from the model's general training. A grounded answer can show you where each part came from.
Hallucination: Confident, fluent output that is factually wrong — invented cases, misquoted rules, fabricated details. A property of generative models, managed by architecture, not by asking the model to be careful.
Evidence gate: A control that checks each claim in an answer against actual evidence before the answer is shown — and blocks or flags the claims that have none, instead of letting them blend in.
Typed evidence: Provenance that records what kind of source supports a claim — a document span, an email, a verified citation, a rule authority, a billing ledger entry — so a reviewer knows how to check it.
Claim-level citation: Citing each individual assertion to its source, rather than appending a bibliography to a whole answer. The difference between checkable and decorative.
Retrieval-augmented generation (RAG): The common technique behind grounding: fetch relevant source material first, then have the model write from it. Necessary but not sufficient — retrieval without verification still hallucinates.

How AI acts

AI agent: An AI system that takes actions — calling tools, creating records, sending communications — rather than only producing text. The useful question is never whether something is 'an agent' but what authority it has and what stops it.
Standing delegation: A persistent instruction given once in plain language ('whenever a document is linked to this matter, summarize it and message me') that the AI keeps executing when triggering events occur — with defined scope, approvals, and limits.
Tool allowlist: The explicit list of capabilities a given AI task may use. A summarization task with no email tool on its allowlist is structurally unable to send email — a stronger guarantee than a policy asking it not to.
Approval gate: A hard pause before a consequential action (send, file, delete, pay) that requires a named human to approve — producing an audit record of who approved exactly what.
Human-in-the-loop: The general pattern of routing AI decisions through people. Approval gates are its enforceable form; 'a human reviews the output eventually' is its unenforceable one.
Bounded run: Autonomous work with hard limits — iterations, budget, progress checks — so a task that stops making progress stops running instead of looping and spending.
Kill switch: A single control that halts all autonomous AI activity immediately. If a vendor cannot point to one, the firm is the kill switch.
Immutable versioning: Instructions that cannot be silently edited: changes create a new version, old versions survive with their history, and every run records which version it executed under.

Citations & authority

Authority ledger: A local, queryable corpus of legal authority used to verify citations deterministically — checking that a cited case actually exists and resolves to the right decision, before any AI judgment is applied.
Deterministic verification: Checking by lookup, not by model opinion. The same citation checked twice gives the same answer, and the answer doesn't depend on a model's mood.
CourtListener / Free Law Project: The nonprofit that maintains the largest freely available corpus of US case law and citations — the public backbone for citation verification at scale.
Pin cite: The page-level reference inside a citation (31 Cal.4th 1187, 1196). Verifying the case exists is step one; verifying the pin cite supports the proposition is the harder, second step.
Citation graph: A record of which documents rely on which authorities, and how — holding, dicta, criticism — so a firm can see its exposure when an authority is questioned.

Deployment & control

Self-hosted: The software runs on infrastructure the firm controls. The complete form of data sovereignty — and meaningful only if the AI, not just the database, runs there too.
Private cloud: A dedicated deployment operated for one firm in its own tenancy — between multi-tenant SaaS and on-premise in both control and effort.
Local model: A language model running on the firm's own hardware. Zero marginal token cost, nothing leaves the building — typically smaller than frontier models and best matched to the right tasks.
Frontier model: The largest commercial models (Claude, GPT, Gemini families). Strongest reasoning; cloud-hosted by their makers, so using them is a data-flow decision as much as a quality one.
Fine-tuning (LoRA-class): Parameter-efficient training that adapts an open model to a firm's own curated examples. Sharpens voice and task performance; does not make a small model a frontier model.
Default-deny governance: No model, tool, or action is available unless affirmatively enabled by firm policy — the opposite of 'everything works until someone objects.'
BYOK: Bring Your Own Key: the firm controls the encryption keys, so the vendor cannot read firm data even in its own infrastructure.
Cost attribution: Metering AI usage to the client matter that consumed it, so AI cost can be billed, absorbed, or examined deliberately — the billing-grade version of a usage dashboard.

Legal operations

Matter-centric: Organizing every workstream — documents, email, deadlines, tasks, billing — around the legal matter as the unit of record, so context stays whole for people and AI alike.
UTBMS codes: The Uniform Task-Based Management System: standard task and activity codes (L110, A103…) that classify legal work on invoices, required by many corporate clients and insurers.
LEDES: The standard electronic-billing file format corporate clients and carriers ingest. If your invoices can't export LEDES, your biggest clients re-key them.
Block billing: Combining multiple activities in one time entry. Carriers and clients discount for it; catching it before the bill goes out is cheaper than the write-down.
Trust accounting (IOLTA): Holding client funds in trust with per-client ledgers, reconciliation, and strict no-commingling rules — the accounting a bar audit examines first.
Court-day math: Deadline calculation that respects court days, holidays, service-method extensions, and dependent chains — where 'add 30 days' goes to die.