The Legal AI Shared Vocabulary

LB
Lakshita Bhargava

27 min read

Purpose

Lawyers and technologists often use the same terms to mean different things, and this divergence is a principal source of misalignment in legal technology work. A specification drafted in one sense and built in another departs from its intent. An evaluation framed on the wrong meaning measures the wrong thing. Supervision directed at the wrong control fails without notice. This vocabulary attempts to gives a single source for the two professions use a term differently.

Two further developments shape its scope. The terms for agents, permissions, tool use and prompt injection are now settling into standard usage, and the vocabulary records them as that usage has formed. The EU AI Act has also introduced several defined terms, among them provider, deployer and regulatory sandbox. Practitioners in other jurisdictions will increasingly need to read these terms alongside their own law. The vocabulary records them as the Act states them, with a view to aligning them across jurisdictions.

A

Action log

Record of every action an agent took.

A chronological record of each tool call an agent made, the instruction that prompted it, the result returned and any human approval given. It differs from the audit trail, which also covers inputs, sources and model version; the action log is the part that records conduct.

Agent ★

Software pursuing a goal through autonomous steps.

Software that pursues a goal through a sequence of steps (calling tools, retrieving documents, drafting, revising) with limited human intervention between steps. Agents now also operate interfaces built for people, such as browsers, desktops and document systems.

Agent harness ★

Software layer that runs an agent.

The layer that decides which model is called, in what order, with which tools, and with access to what. It is where permissions, boundaries and approval steps are set.

Agent-to-agent protocol (A2A)

Standard for agents delegating to other agents.

An open protocol through which one agent discovers another agent's capabilities and hands it a task. The Model Context Protocol (MCP) connects an agent to tools; A2A connects agents to each other.

AI agent and agentic AI ★

The system and the design approach behind it.

An AI agent is a single system that is given a goal, a degree of autonomy and access to tools, and that works through a sequence of steps until the goal is met or a limit is reached. Agentic AI is the wider design approach, and the class of products built on it, in which models plan, act, check their own results and hand work to other agents. The distinction is one of scope: the agent is the unit that acts, and agentic AI describes how systems of such units are built and governed.

AI feature and AI agent ★

The line between assisted and autonomous work.

A feature is a model placed in software for turn-based use: a person requests, reviews and decides at each step. An agent has a goal, a degree of autonomy and access to real systems, and works toward an outcome without a human turn between steps. The distinction lies in goal, autonomy and system access; a more sophisticated feature is still a feature.

Alignment

Training a model to behave as intended.

The engineering goal of making a model behave as its operator intends, for example by refusing improper requests and staying within instructions.

API (application programming interface) ★

Machine-to-machine channel for requesting model services.

A defined channel through which one program requests services from another. It is the route by which a document system calls a language model without a person using a chat screen.

Approval gate

Mandatory human sign-off before an action proceeds.

A point in an agent’s workflow where execution pauses until a named person approves the next action. That action is typically irreversible or outward-facing, such as sending, filing, paying or deleting. It is the concrete form that human-in-the-loop takes for agents.

Artificial intelligence (AI)

Umbrella term for machine performance of cognitive tasks.

An umbrella term for systems that perform tasks associated with human cognition. In legal practice it spans rule-based systems, statistical machine learning such as technology-assisted review, and generative models. Precision requires naming which kind is meant.

Audit trail

Tamper-evident record of system activity.

A tamper-evident record of what a system did: inputs, retrieved sources, model version, tool calls, outputs and human edits.

Autonomy level

How much an agent may do before a human intervenes.

The degree of independent action granted to an agent, from suggesting a step for a person to take, through acting with approval at each gate, to acting end to end with review afterwards. It is set in the agent harness, and the same model can run at any level.

B

Benchmark

Standardised test set for scoring models.

A standardised test set used to score model performance. A score can overstate real ability for three reasons: test questions may have leaked into training data; labs may have tuned the model to the test; and even a structured test measures general capability rather than performance on a given set of matters. A benchmark works like the bar exam: it shows a general standard has been met and says little about competence on a particular matter.

Bias ★

Systematic error disadvantaged identifiable groups.

Systematic error in outputs that disadvantages identifiable groups or skews results. It may be systemic, statistical or computational, or human-cognitive in origin.

C

Chain-of-thought

Generated reasoning steps preceding an answer.

A technique in which the model produces intermediate reasoning steps before its answer; reasoning models are trained to do this at length. The visible reasoning is itself generated text. It is not a faithful record of how the output was produced and should not be relied on as an explanation of it.

Chunking

Segmentation of documents for retrieval.

Splitting documents into segments before they are embedded for retrieval.

Citator ★

Service reporting subsequent treatment of authority.

A research service, such as Shepard’s or KeyCite, that reports how an authority has been treated since (followed, distinguished, overruled), so that a lawyer knows whether it remains good law.

Computer use

Agent operation of a screen, keyboard and mouse.

The capability by which a model views a screen and operates software built for people by moving a cursor, clicking and typing, rather than through an API. It extends agents to systems that offer no programmatic interface.

Connector

Link between an AI tool and an outside system.

A configured link that lets an AI tool read from or act in another system, such as a document management system, email, a calendar or a research database. A connector runs under an account and permissions set when it is installed, so it decides what the tool can see and what it can do in that system.

Context engineering

Deliberate selection of what the model sees.

The practice of deciding what a model sees on each call (instructions, tool definitions, retrieved passages, memory and conversation history), selected and ordered deliberately within the context window.

Context window ★

Text span the model can attend to at once.

The limited span of text a model can attend to in one exchange: instructions, documents and conversation together. Anything outside it is invisible to the model, and the window does not persist once the session ends.

Contract lifecycle management (CLM)

Contract management software from request to expiry.

Software that manages contracts from request through drafting, negotiation and execution to post-signature obligations, built around a clause library and approval workflows. Generative AI now commonly sits on top for first drafts and clause extraction.

D

Data residency

Where data is stored and processed.

The physical and legal location in which data, including prompts, documents and outputs, is stored and processed. It determines which jurisdiction’s law applies to the data, and it is fixed by contract, by the deployment model and by where inference runs.

Deployer and provider ★

The EU AI Act’s two principal roles.

The two principal roles under the EU AI Act. A provider develops an AI system or model, or has one developed, and places it on the market or puts it into service under its own name or trademark. A deployer uses an AI system under its own authority, other than in a personal, non-professional activity. An organisation using a vendor’s tool is a deployer.

Deterministic and probabilistic ★

Repeatable output against sampled output.

A deterministic system returns the same output for the same input every time, as a document-assembly template or a calculation does. A probabilistic system samples from a distribution, so identical prompts can produce different outputs. Generative models are probabilistic, and even fixed settings do not guarantee identical output.

Document assembly

Template-driven drafting from questionnaire answers.

Deterministic drafting automation: templates with variables and conditional logic produce documents from questionnaire answers. It predates generative AI and cannot produce language that no one has approved.

Dry run

Execution that reports actions without performing them.

A mode in which an agent carries out its reasoning and reports the actions it would take, without taking them. It is the cheapest way to see what an agent will do with a mandate before granting it.

Duty of technology competence ★

Professional duty to understand tools used in practice.

The professional obligation to understand the benefits and risks of technology used in practice. United States professional-conduct guidance frames the duty and applies it to generative AI, requiring a reasonable understanding rather than expertise.

E

E-discovery ★

Discovery of electronically stored information.

The identification, preservation, collection, review and production of electronically stored information in litigation and investigations. It was the first legal domain to industrialise machine learning, through technology-assisted review.

Embedding ★

Numeric vector encoding the meaning of text.

A numeric vector representing the meaning of a piece of text, placed so that passages with similar meanings sit near each other. Embeddings power semantic search, which retrieves by meaning rather than by keyword.

Evaluation (evals) ★

Structured testing against a defined task set.

Structured testing of a system against a defined set of tasks with graded outputs. It is distinct from a vendor demonstration and from public benchmarks. A senior lawyer should produce or approve the correct answers. Its highest-value use is regression testing after a vendor or prompt change.

Existence-checker

Confirmation that a cited authority is real.

A tool or workflow step that confirms a cited authority exists in an authoritative database. It answers only whether the citation is real.

Explainability

Traceability of output to assessable reasons.

The degree to which an output can be traced to reasons a human can assess. For generative models, explanation of the internal mechanism is largely unavailable; in practice, explainability means showing sources and steps.

Extension

Add-on that gives an AI tool access to another application.

A small program installed in a host application, most often a browser, that lets an AI tool read and operate that application on the user's behalf. A browser extension acts with the user's own sign-ins, so the tool can reach whatever the user can reach.

Extractive AI ★

Systems that pull existing text from sources.

Systems that locate and pull existing text or data from source documents, such as clause extraction, entity recognition and passage retrieval. Because output is verbatim or close to it, extraction is far less prone to fabrication. It fails mainly by missing or mis-selecting material, although features built on generative models can still paraphrase or invent.

F

Few-shot prompting

Worked examples supplied within the prompt.

Including worked examples in a prompt so that the model imitates the pattern; zero-shot prompting gives instructions only. It is often the cheapest way to improve the format and consistency of output.

Fine-tuning ★

Further training that alters model weights.

Further training of an existing model on additional examples, which changes its weights to shift style, format or task behaviour. It is not a reliable way to teach a model facts it can cite. Unlike retrieval, it cannot be updated instantly or attributed to sources.

Fleet and swarm

Several agents working one problem.

A fleet is several agents with defined roles and access working on parts of a problem. A swarm adds heavy cross-communication and task-sharing between them. Mixed designs pair a large model for reasoning with small models for narrow steps.

Foundation model

Broad base model adapted to many uses.

A large model trained on broad data and adapted to many downstream uses. It is the base layer beneath most legal AI products. The EU AI Act regulates these models as general-purpose AI models.

Frontier model

Largest and most capable class of model.

The tier of largest models, requiring hardware most organisations cannot own. Smaller models handle well-defined narrow tasks at lower cost.

G

Generative AI ★

Systems composing new content by prediction.

Systems that compose new content (text, images, code) by predicting plausible continuations, rather than by retrieving or extracting existing content. Fluency is built in; factual reliability is not.

Grounding ★

Constraint of output to identified sources.

Constraining output to identified source material, such as retrieved documents, a contract under review or a research database, rather than the model’s internal parameters. Grounding reduces fabrication but does not eliminate error, because the model can still misread or overstate its sources.

Guardrails

Technical controls around model inputs and outputs.

Technical controls around a model, such as input filters, output filters, refusal rules and restricted tool access. Guardrails are probabilistic defences layered on a probabilistic system. They reduce the rate at which policy is breached and cannot replace review.

H

Hallucination ★

Fluent output that is false or unsupported.

Output that is fluent but false or unsupported, such as invented cases, misattributed quotations or confident misstatements of a document's terms. The rate of hallucination is material for general-purpose chatbots and remains material for specialised legal research tools. Courts are increasingly addressing AI-fabricated material in filings.

Handoff

Transfer of a task between agents or to a human.

The point at which one agent passes a task, with its context and results so far, to another agent or to a person. What is passed determines what the receiver knows; anything not passed is lost.

Human-in-the-loop ★

Human review before output takes effect.

A workflow design in which a human reviews or approves output before it takes effect.

I

Inference

Running a trained model to produce output.

Running a trained model to produce output, as opposed to training it. The cost an organisation pays for per query is inference. Use does not change the model unless inputs are separately kept for training. It is the unit vendors bill against.

Integration

Built-in connection between two products.

A connection between two products that the vendors have built and maintained, through which data passes automatically, for example between a contract lifecycle tool and an e-signature service. Integrations are fixed by the vendor, while connectors and plug-ins can usually be added and configured by the user.

Interrupt

Stopping an agent mid-task.

The ability to halt an agent while it is running, before it completes or takes its next action, and to inspect its state at that point.

J

Jagged frontier ★

Uneven capability across similar tasks.

The finding that AI capability is uneven in ways that defy intuition: strong on some tasks and weak on neighbouring tasks that look equally difficult. In a field experiment with consultants, using AI improved performance on tasks inside the frontier and reduced accuracy on a task just outside it.

Jailbreak

User input crafted to bypass guardrails.

An input crafted by the user to make a model bypass its guardrails. It differs from prompt injection, where the hostile instruction arrives through content the model processes rather than from the user.

K

Knowledge cutoff ★

Date at which training data ends.

The date after which a model’s training data ends. Without retrieval, the model knows nothing after that date, including new decisions, amended statutes and overruled authority, and it will not reliably signal that it is out of date.

Knowledge management (KM)

Capture and reuse of collective work products.

The discipline of capturing and reusing an organisation’s collective work product, such as precedent banks, playbooks and matter know-how. KM systems are the natural retrieval corpus for grounded AI, so the quality of KM directly determines the quality of AI output.

L

Large language model (LLM) ★

Next-token predictor trained on vast text.

A neural network, typically a transformer, trained on vast text corpora to predict the next token. Drafting, summarisation and question-answering emerge from that training. It is the engine inside most legal generative AI tools.

Legal designer

Design methods applied to legal services.

A practitioner who applies design methods (user research, prototyping, visual and plain-language design) to legal services and documents, so that contracts, notices and tools work for the people who must use them.

Legal engineer ★

Translator of legal judgment into executable systems.

A hybrid role that turns legal expertise into working systems. The work includes encoding playbooks into CLM logic, building document automation, designing and testing AI workflows, and defining evaluation criteria.

Legal operations (legal ops)

Business management of legal functions.

The business-management discipline of legal departments and firms, covering vendor selection, process design, metrics, spend and technology adoption. Legal ops usually owns the decision to buy AI tools; supervising their use remains a lawyer’s duty.

Legal technologist

Selection and administration of legal technology.

A practitioner who selects, configures, integrates and administers legal technology. The legal technologist makes tools run in the organisation’s environment; the legal engineer encodes legal substance into them. In practice the roles overlap.

Loop

The repeated cycle an agent runs.

The cycle of observe, reason, act and check that an agent repeats until its goal is met or a limit is reached. Each pass through the loop is a step; the full sequence is a trajectory.

M

Machine learning (ML)

Pattern learning from data in place of hand-written rules.

Systems that learn patterns from data rather than following hand-written rules. The term covers both the classifiers behind technology-assisted review, which predict rather than generate, and the large models behind generative AI.

MCP server

Program that exposes one system to AI tools through MCP.

A program that describes a system's data and actions in the format set by the Model Context Protocol, so that any compatible AI tool can discover and call them. The server holds the system's credentials and decides which actions it offers, which makes it the point where access to that system is controlled.

Memory (persistent) ★

Cross-session storage supplied to future sessions.

A product feature that stores information from past sessions, such as facts about the user, prior matters and preferences, and supplies it to future sessions. It differs from the context window, which does not persist after a session, and from training, which memory does not involve.

Model

Trained artefact of architecture and weights.

The trained artefact, made up of an architecture and learned weights, that produces outputs. Models, tools and products are different layers: a legal AI product typically wraps a third-party model with retrieval, guardrails and an interface. Models are named by input and output type ("text in, text out"), and not every model is a neural network.

Model Context Protocol (MCP) ★

Open standard for connecting models to tools.

An open standard through which an AI application discovers and calls external tools and data sources over one common interface. Each system is exposed once, as an MCP server. Credentials can stay inside the server, so the model need not hold them.

Model-first and retrieval-first ★

Answering from parameters against answering from a corpus.

Two architectures for answering questions. A model-first tool answers from the model’s parameters: fluent and fast, but uncited and frozen at the knowledge cutoff. A retrieval-first tool searches a maintained corpus first and drafts from what it finds: it can cite and is as current as its corpus, but it is only as good as its retrieval.

Model version

Specific release of a model.

A specific release of a model. Vendors replace or update underlying models, and behaviour, failure modes and evaluation results do not carry over from one version to the next.

Multimodal

Models handling more than one medium.

Models that accept or produce more than one medium, such as text, images or audio. In practice this means reading scanned exhibits, extracting tables from PDFs and transcribing audio. Each added medium adds failure modes, such as misreads of the kind OCR produces, presented fluently.

N

Natural language processing (NLP)

Computational handling of human language.

The field of handling human language computationally. Older legal NLP, such as keyword expansion, entity extraction and classification, remains in use alongside generative models. Not everything called AI in a legal product is generative.

NIST AI Risk Management Framework (AI RMF)

Voluntary US framework for AI risk governance.

A voluntary US framework for governing AI risk through four functions: Govern, Map, Measure and Manage. Legal, risk and engineering teams increasingly use it as a shared checklist.

O

Observability

Visibility into what an agent is doing and why.

The instrumentation that lets a person see an agent’s steps, tool calls, inputs and outputs as they happen and afterwards. The action log and audit trail are its records; observability is the live view.

Open weights and closed

Whether the model can be run in-house.

Open weights means the weights can be obtained and the model run on the organisation's own systems. Closed means the vendor keeps the weights, the software or both, and access is through an API or interface. Open weights and open source are different claims.

Opinionated and unopinionated

Whose judgment is built into the tool.

An opinionated tool carries built-in decisions: the model, hidden system instructions, the workflow, the retrieval sources, the definition of good output and the refusals. An unopinionated setup leaves those judgments to the user.

Orchestration

Coordination layer sequencing models, tools and checkpoints.

The coordination layer, sometimes called the harness, that sequences models, retrieval, tools and human checkpoints into a workflow. It is where an AI feature becomes a process. Most decisions about reliability and oversight are made at this layer.

P

Parameters

Numeric weights set during training.

The numeric weights inside a model, set during training, that hold its learned knowledge and behaviour. Parameter count measures size and says nothing about quality. Knowledge held in parameters cannot be attributed to a source and is frozen at the point of training. Software and weights are separate and can be owned and distributed separately.

Permission scope

Bounded set of resources and actions an agent may use.

The defined set of data sources, tools and actions an agent may reach, fixed before it runs. Security frameworks identify the absence of a permission scope, known as "excessive agency", as a principal risk of LLM applications.

Playbook

Codified negotiating positions and fallbacks.

An organisation’s codified negotiating positions: preferred clauses, fallbacks and escalation triggers. In AI-assisted contract review, the playbook is the specification the tool applies. The tool’s output is therefore only as good and as current as the playbook.

Plug-in

Installable bundle that extends an AI tool.

A package installed into an AI tool that adds capabilities in one step. It can contain skills, connectors, commands and automated checks. A plug-in brings its own instructions and permissions with it, so installing one changes what the tool knows how to do and what it can reach.

Precedent ★

Judicial authority or reusable document.

The clearest homonym in the field: judicial authority in one usage, a reusable document in the other. The split also runs within the legal profession, between contentious and transactional practice.

Privilege ★

Protection from disclosure, or system access rights.

Protection from disclosure in one profession; access rights in the other. Sending privileged material to a vendor raises questions of confidentiality, and potentially of waiver, which US professional-conduct guidance addresses.

Prompt ★

Full input the model receives.

The full input a model receives: the user’s request together with system instructions, injected documents, tool definitions and conversation history. The model responds to all of it, which is why the same question behaves differently inside different tools.

Prompt engineering

Craft of structuring instructions for reliability.

The craft of structuring instructions, examples, output formats and source constraints to make a model behave more reliably. It reduces risk without guaranteeing anything, and it does not replace verification of output. In agent systems it has been absorbed into context engineering.

Prompt injection ★

Hostile instructions hidden in processed content.

An attack in which instructions hidden in content the model processes cause it to act against its operator’s intent. The content may be a document under review, a web page, an email or a tool result. Security frameworks rank prompt injection as the foremost risk to LLM applications, and no complete defence is known.

Provenance

Traceable origin of an output.

The traceable origin of an output: which sources, which model version, which prompt and whose edits.

Q

Query

Request sent to a search or AI system.

The request sent to a search or AI system. In retrieval systems, the query determines what the model sees as much as the corpus does: a poorly framed query grounds the answer in the wrong passages.

R

Reasoning model

Model trained to work through a problem before answering.

A model trained to generate extended intermediate reasoning before giving its answer, which improves performance on multi-step tasks such as analysis, calculation and code. It costs more and takes longer per query, and its visible reasoning carries the same limits as chain-of-thought.

Red-teaming

Adversarial testing before deployment.

Deliberate adversarial testing before deployment: trying to make a system fail, leak data or misbehave. It complements evaluations, which test expected use.

Reinforcement learning from human feedback (RLHF)

Training on human preference ratings.

A training stage in which human preference ratings steer a model towards helpful, acceptable answers. Because it optimises for outputs people rate highly, it is one reason confident, fluent errors survive training.

Retrieval-augmented generation (RAG) ★

Retrieved passages supplied to the model as context.

An architecture that retrieves relevant passages from a defined corpus and supplies them to the model as context for drafting an answer. Retrieval can fetch the wrong passages, and the model can still misstate what was fetched. Both happen at material rates, even in leading legal tools.

Rollback

Reversal of an agent’s action.

The ability to undo an action an agent has taken and restore the prior state. Some actions can be rolled back; others, such as sending, filing or paying, cannot. Actions that cannot be rolled back are the ones that call for an approval gate.

S

Sandbox

Isolated environment, or supervised regulatory space.

An isolated environment in one profession; a supervised regulatory space in the other. The EU AI Act requires each Member State to establish an AI regulatory sandbox. Regulatory sandboxes for legal services also exist in the United States.

Skill

Packaged instructions for one kind of task.

A stored set of instructions, templates and sometimes scripts that an AI tool loads when a task calls for it, such as drafting in a house style or reviewing a contract against a playbook. A skill changes how the tool performs a task. It does not give the tool access to any new system.

Small language model

Compact model for narrow, well-defined tasks.

A language model with far fewer parameters than a frontier model, cheap enough to run at high volume and small enough to run on an organisation’s own systems. It suits narrow tasks such as classification, extraction and routing, and agent systems often pair it with a larger model.

Structured data

Data with a defined schema.

Data with a defined schema (fields, dates, parties, amounts), as opposed to free text. Much legal AI work is really structuring: turning contracts and dockets into fields that deterministic systems can then process reliably.

Sub-agent

A subordinate agent assigned a bounded task by another.

An agent created or called by another agent to perform part of a larger task. It has its own context window and permission scope, and returns its result to the agent that called it. Sub-agents are the standard way to keep any one agent’s context small.

Supervision (of AI tools) ★

Lawyer’s non-delegable duty of oversight over tools.

The lawyer’s duty to oversee non-lawyer assistance extends to software. Under US professional-conduct rules and guidance, managerial lawyers must ensure, through clear policies and training, that AI use complies with professional obligations.

System prompt

Builder’s standing instructions preceding every input.

Standing instructions that the tool’s builder places before every user input, defining role, constraints and format. Most users never see it, yet it shapes every output. Two products built on the same model behave differently largely because of it.

T

Technology-assisted review (TAR) ★

Machine classification of documents in discovery.

Machine-learning classification of documents in discovery, also called predictive coding. Reviewers code seed documents, and the system ranks the rest by predicted relevance, refining as it goes. Courts have accepted it since Da Silva Moore v. Publicis Groupe. TAR ranks existing documents and cannot fabricate content.

Temperature

Sampling setting controlling output randomness.

A sampling setting that controls how random output is. Low temperature makes output more repeatable and conservative; higher temperature makes it more varied. Even at zero, a generative model is not fully deterministic in practice. Temperature tunes variance; it does not create reliability.

Token ★

Unit of text a model reads and writes.

The unit of text a model reads and writes: a word or word fragment, on average somewhat shorter than an English word. Context windows, pricing and length limits are all measured in tokens.

Tool use (function calling) ★

Model’s structured request for an external action.

The mechanism by which a model, instead of answering in prose, issues a structured request for an external action, such as searching a database, reading a file, sending an email or running code. The surrounding software carries out the action and returns the result to the model. Tool use is what turns a chat model into an agent; MCP standardises how tools are described and called.

Training data ★

Corpus a model learned from.

The corpus a model learned from. It determines the model’s knowledge, style, blind spots and biases, and users generally cannot inspect it. Vendor terms separately determine whether a user’s inputs become someone else’s training data.

Trajectory

The full sequence of an agent’s steps on one task.

The complete record of one run: every observation, reasoning step, tool call and result from the instruction to the outcome. It is what an evaluation examines and what a post-incident review reconstructs.

Transformer

Attention-based architecture underlying LLMs.

The neural-network architecture underlying modern LLMs. Its attention mechanism lets the model weigh relationships across an entire span of input, which is why context windows work as they do.

U

Unauthorized practice of law (UPL)

Prohibition on non-lawyer legal services.

The prohibition, as named in US law, on non-lawyers providing legal services. It varies by jurisdiction and shapes how consumer-facing legal AI must be positioned: it may provide legal information and must stop short of individualised advice. For tools that generate individualised legal answers, the boundary remains live and unsettled.

V

Vector database ★

Datastore retrieving by semantic similarity.

A datastore optimised for embeddings, which retrieves passages by semantic similarity. It is the retrieval backbone of most RAG systems.

Verification ★

Human confirmation of output before reliance.

The human act of confirming output before relying on it. For cited authorities, this means confirming that they exist, remain good law and support the stated proposition. For summaries, it means confirming that the documents say what the summary claims. Professional guidance makes the degree of independent verification depend on the tool and the task, and the court decisions on fabricated citations show the cost of skipping it.

Vertically integrated and interchangeable

Whether the model can be swapped beneath the tool.

Vertically integrated means one vendor supplies compute, model, harness and interface as a bundle. Interchangeable means the model can be replaced beneath a separate harness.

W

Workflow automation

Routing of tasks through defined steps.

Software that routes tasks, documents and approvals through defined steps, such as intake, review, signature and filing. It is often deterministic and often the higher-value form of automation; not every workflow needs a generative model.

Z

Zero data retention (ZDR) ★

Contractual non-storage of prompts and outputs.

A contract term under which the model provider does not store prompts or outputs once a request is complete. It differs from a promise not to train on data, since data may be kept without being trained on, and from encryption, since data may be stored securely.

Zero-shot prompting

Instruction with no examples.

Instructing a model to perform a task without examples, relying on what it learned in training. It is the cheapest form of prompt to write and the least controllable; adding examples usually improves consistency.

LB

Author

Lakshita Bhargava

A legal technology specialist, working on legal design, computational legal analysis and the deployment of AI systems in legal practice. Trained as a mathematician before practising law, also analyses cybersecurity, competition law, and other focused AI regulation .