# Legal AI Glossary: full definitions (plain text) > Plain-English definitions of the AI terms legal professionals encounter: core concepts, how models work, usage terms, risk and governance vocabulary, and legal AI applications. Each definition is self-contained and citable, with the legal-practice angle noted where one exists. Maintained by MHSB Solutions. Educational information, not legal advice. Attribution: Legal AI Glossary (https://legalai.help), maintained by MHSB Solutions. Free to read and quote with attribution. Site last updated: 2026-08-04 Total terms: 50 Generated from the same data as the glossary pages on every build. For what bars and courts require of lawyers using AI, see the companion Legal AI Compliance Tracker: https://legalaicompliance.help/ ## Core concepts ### Agentic AI AI designed to pursue goals through multi-step actions—planning, using software tools, and adjusting course—with limited human direction, rather than answering a single prompt. The umbrella term for systems built around AI agents. Where a chatbot produces one response per request, agentic AI breaks a goal into steps, executes them—searching, drafting, calling other software—and evaluates its own progress. This autonomy raises the stakes of errors, since a mistake early in a chain can compound across later steps. Organizations adopting agentic tools generally pair them with human review checkpoints and explicit limits on what actions the system may take on its own. - Glossary page: https://legalai.help/term/agentic-ai/ ### AI agent A software system that uses an AI model to make decisions and take actions toward a goal—running searches, filling forms, or managing a workflow—without step-by-step human instructions. An AI agent wraps a language model in software that lets it use tools: browsing, querying databases, drafting documents, or triggering other applications. Agents typically work in a loop—act, observe the result, decide the next step—until the task is done or a limit is reached. In legal operations, agents are being applied to intake, document assembly, and research workflows, with the supervising professional remaining responsible for the results. - Glossary page: https://legalai.help/term/ai-agent/ - Reference: https://en.wikipedia.org/wiki/Intelligent_agent ### Deep learning A type of machine learning that uses artificial neural networks with many layers to learn complex patterns from large amounts of data. It underlies modern image recognition, speech recognition, and language models. Deep learning networks pass data through stacked layers of simple mathematical units, with each layer learning progressively more abstract features—edges become shapes, words become meanings. The approach became dominant in the 2010s as computing power and available data grew, and it is the technology behind transformers and large language models. Its results are powerful but difficult to interpret, which is one reason explaining exactly why an AI system reached a particular conclusion can be hard. - Glossary page: https://legalai.help/term/deep-learning/ - Reference: https://en.wikipedia.org/wiki/Deep_learning ### Foundation model A large AI model trained on broad data at scale so it can be adapted to many downstream tasks—drafting, translation, coding—rather than built for a single purpose. Large language models are the most prominent examples. Foundation models serve as a base layer: vendors adapt them through fine-tuning or retrieval-augmented generation to build specialized products, including legal drafting and research tools. Training such models requires data and computing resources only a few organizations can afford, so most AI products are built on a small number of underlying models. Strengths and flaws of the base model—such as a tendency to hallucinate—carry through to the tools built on top of it. - Glossary page: https://legalai.help/term/foundation-model/ - Reference: https://en.wikipedia.org/wiki/Foundation_model ### Generative AI Artificial intelligence that creates new content—text, images, audio, video, or code—by learning patterns from large datasets, rather than only analyzing or classifying existing information. Generative AI systems are trained on large collections of existing material and then produce new output in response to instructions called prompts. Large language models, which generate text, are the type most used in legal work, but image, audio, and video generators follow the same principle. Output is a statistical construction rather than a retrieval of stored documents, so it can be fluent yet wrong. Law firms typically govern generative AI use through policies covering confidentiality, verification of output, and client communication. - Glossary page: https://legalai.help/term/generative-ai/ - Reference: https://en.wikipedia.org/wiki/Generative_artificial_intelligence ### Large language model (LLM) A neural network trained on massive text corpora to predict the next token, producing fluent text. The engine behind ChatGPT, Claude, and Gemini. Large language models learn statistical relationships between words by processing enormous volumes of text during training. At run time they generate answers one token at a time, choosing each next piece by probability rather than looking facts up in a database, which is why they can draft and summarize well yet still state falsehoods confidently. Lawyers using LLMs remain responsible for verifying output and for keeping client information out of tools whose data terms have not been reviewed. - Glossary page: https://legalai.help/term/large-language-model/ - Reference: https://en.wikipedia.org/wiki/Large_language_model ### Machine learning (ML) A field of computer science in which systems learn patterns from data and improve at tasks through experience, instead of following rules written by hand. The foundation of modern AI, including large language models. Machine learning algorithms are trained on examples—emails labeled spam or not, contracts labeled by clause type—until they can make accurate predictions about new data. The approach powers everyday tools such as spam filters and document classification as well as generative AI. Performance depends heavily on the quality and representativeness of the training data, which is why biased or thin data produces unreliable results. - Glossary page: https://legalai.help/term/machine-learning/ - Reference: https://en.wikipedia.org/wiki/Machine_learning ### Multimodal AI AI systems that can process and generate more than one type of data—such as text, images, audio, and video—within a single model. A multimodal model can, for example, describe a photograph or read a scanned document. Earlier AI models handled one data type at a time; multimodal models accept mixed input and can reason across formats, such as answering questions about a chart or extracting terms from a scanned exhibit. Major commercial AI assistants are now multimodal by default. For legal work this enables direct analysis of discovery images, recorded audio, and scanned filings, though accuracy on poor-quality scans and handwriting still calls for human checking. - Glossary page: https://legalai.help/term/multimodal-ai/ ### Natural language processing (NLP) The branch of AI concerned with enabling computers to read, interpret, and generate human language. It covers tasks such as translation, summarization, classification, and the text generation performed by large language models. NLP combines linguistics and machine learning to turn unstructured text into something software can work with. Earlier NLP relied on hand-built rules and statistical models; today most of the field is built on transformer-based language models. In legal settings NLP underpins contract review, e-discovery search, and research tools that read case law. - Glossary page: https://legalai.help/term/natural-language-processing/ - Reference: https://en.wikipedia.org/wiki/Natural_language_processing ### Transformer The neural network architecture behind modern language models, introduced by Google researchers in 2017. Its attention mechanism lets a model weigh how every word in a passage relates to every other, enabling fluent long-form text. Transformers replaced earlier architectures that processed text strictly one word at a time, making it practical to train on vastly larger datasets in parallel. The self-attention mechanism is what lets models track references and context across long documents. Nearly every prominent generative AI system is built on this architecture; the letter T in GPT stands for transformer. - Glossary page: https://legalai.help/term/transformer/ - Reference: https://en.wikipedia.org/wiki/Transformer_(deep_learning_architecture) ## How models work ### Context window The maximum amount of text, measured in tokens, an AI model can consider at one time—covering the instructions, any documents provided, and the response. Material beyond the window is invisible to the model. The context window functions as the model's working memory for a single exchange. Modern models accept very large windows—enough for entire contracts or briefs—but accuracy in recalling details can degrade as input grows, and anything outside the window is simply not seen. When a long conversation exceeds the window, earlier content may be dropped or summarized, which can cause the model to lose track of instructions or facts stated earlier. - Glossary page: https://legalai.help/term/context-window/ ### Embedding A list of numbers that represents the meaning of a piece of text, so that similar content sits close together mathematically. Embeddings let software search by concept rather than by exact keywords. An embedding model converts words, sentences, or documents into vectors—coordinates in a high-dimensional space—where distance reflects similarity of meaning. This is the basis of semantic search: a query about termination for cause can surface clauses that never use those exact words. Embeddings are stored and searched in vector databases and are a core component of retrieval-augmented generation. - Glossary page: https://legalai.help/term/embedding/ - Reference: https://en.wikipedia.org/wiki/Word_embedding ### Fine-tuning Additional training applied to an existing AI model using a smaller, specialized dataset so it performs better on a particular domain or task—such as legal drafting conventions—without building a model from scratch. Fine-tuning adjusts a pretrained model's weights with targeted examples, in contrast to retrieval-augmented generation, which leaves the model unchanged and supplies documents at question time. Legal technology vendors use fine-tuning to teach models domain vocabulary, formats, and tone. Material used for fine-tuning becomes part of the model's learned behavior, so client data should not be used for it without clear authority and safeguards. - Glossary page: https://legalai.help/term/fine-tuning/ - Reference: https://en.wikipedia.org/wiki/Fine-tuning_(deep_learning) ### Inference The stage at which a trained AI model is put to work—taking an input and producing output. Every chatbot answer is an act of inference; training, by contrast, is when the model learned. During inference the model's weights stay fixed; the system runs the input through the network to compute a response, one token at a time. For most commercial tools inference happens on the provider's servers, which is why confidentiality review focuses on where prompts and documents are sent, processed, and stored. Inference costs scale with usage, which is why providers meter and bill by token. - Glossary page: https://legalai.help/term/inference/ ### Model weights The numerical parameters—often billions of them—inside a neural network that encode everything the model learned during training. The weights are the model: whoever holds them can run or modify it. Training gradually adjusts the weights until the model performs well; afterward, the weights are a large file of numbers that, with the right software, reproduces the model's behavior anywhere. Open-weight models publish this file for anyone to download, while closed models keep weights on the provider's servers. Weights do not store documents verbatim the way a database does, although models can sometimes memorize and reproduce fragments of training data. - Glossary page: https://legalai.help/term/model-weights/ ### Reinforcement learning from human feedback (RLHF) A training method in which people rate a model's answers and those ratings teach it to produce responses humans prefer—more helpful, better aligned with instructions, less harmful. A key step in turning raw language models into usable assistants. After initial training on text, models are refined by collecting human judgments on sample outputs and using them to steer the model toward preferred behavior. RLHF is a large part of why modern chatbots follow instructions and decline inappropriate requests. It also contributes to a known side effect: models tuned to please can sound confident and agreeable even when they are wrong. - Glossary page: https://legalai.help/term/reinforcement-learning-from-human-feedback/ - Reference: https://en.wikipedia.org/wiki/Reinforcement_learning_from_human_feedback ### Retrieval-augmented generation (RAG) A technique that has an AI model answer using documents fetched from a trusted source at question time, instead of relying only on its training. It grounds responses in citable material and reduces—but does not eliminate—hallucination. A RAG system first searches a designated corpus—case law, firm documents, a contract repository—usually via embeddings in a vector database, then passes the best matches to the model along with the question. Because answers are tied to retrieved sources, they are easier to verify, which is why most legal research and document-question tools are built this way. Quality depends on the retrieval step: if the right passage is not found, the model may still guess, so verification duties remain. - Glossary page: https://legalai.help/term/retrieval-augmented-generation/ - Reference: https://en.wikipedia.org/wiki/Retrieval-augmented_generation ### System prompt Standing instructions given to an AI model before any user input, defining its role, rules, tone, and limits. Users of a product usually never see the system prompt steering their conversation. Product builders use system prompts to make a general-purpose model behave like a specific assistant—telling it, for example, to answer only from provided documents or to decline to give legal advice. System prompts shape output significantly but are not unbreakable; crafted user input can sometimes override them, a technique known as prompt injection. When evaluating an AI tool, it is reasonable to ask the vendor how its instructions constrain the model and what safeguards back them up. - Glossary page: https://legalai.help/term/system-prompt/ ### Temperature A setting that controls how predictable or varied an AI model's output is. Low temperature makes responses more consistent and conservative; high temperature makes them more varied and creative. Technically, temperature reshapes the probability distribution the model samples from when choosing each next token. At low settings the model nearly always picks the most likely continuation, which suits drafting and analysis; higher settings encourage variety, which suits brainstorming. Temperature affects variability, not truthfulness—a model at the lowest setting can still hallucinate. - Glossary page: https://legalai.help/term/temperature/ ### Token The small unit of text a language model actually reads and writes—a word, part of a word, or punctuation mark. Models measure input size, output length, pricing, and usage limits in tokens. Before processing, text is split into tokens using a fixed vocabulary; common words may be a single token while rare or technical terms are broken into several. The model then predicts its output one token at a time. Token counts matter practically because they determine how much material fits in a model's context window and how API usage is billed—long documents such as contracts or deposition transcripts consume many thousands of tokens. - Glossary page: https://legalai.help/term/token/ ### Training data The collection of text, images, or other examples a machine-learning model learns from. A model's knowledge, abilities, and biases all derive from what its training data did and did not contain. For large language models, training data typically includes web pages, books, code, and licensed datasets gathered up to a cutoff date, after which the model knows nothing unless given new information. Errors, gaps, and biases in the data surface as errors and biases in the model. The provenance of training data is the subject of ongoing copyright litigation against AI developers, and whether user inputs become training data is a central confidentiality question when evaluating any AI tool. - Glossary page: https://legalai.help/term/training-data/ - Reference: https://en.wikipedia.org/wiki/Training,_validation,_and_test_data_sets ### Vector database A database built to store embeddings—numerical representations of meaning—and quickly find the entries most similar to a query. The storage layer behind semantic search and most retrieval-augmented generation systems. Instead of matching keywords, a vector database compares the geometry of embeddings to return conceptually similar documents, even across millions of records. In legal AI products, firm or matter documents are typically split into chunks, embedded, and indexed this way so a model can be handed the most relevant passages. Because the indexed content may include client material, vector stores fall within the scope of confidentiality and security review like any other document repository. - Glossary page: https://legalai.help/term/vector-database/ - Reference: https://en.wikipedia.org/wiki/Vector_database ## Using AI ### API (application programming interface) A defined way for software programs to communicate with each other. AI providers expose their models through APIs, letting firms and vendors build the models into their own tools and workflows. Instead of a person typing into a chatbot, a program sends a request—prompt, documents, settings—to the provider's API and receives the model's output for use inside an application. Most legal AI products are built this way on top of a few major model providers. API terms often differ from consumer-app terms, commonly including no training on customer inputs and shorter retention, which matters when tracing where client data actually flows. - Glossary page: https://legalai.help/term/api/ - Reference: https://en.wikipedia.org/wiki/API ### Chain-of-thought (CoT) A technique in which an AI model works through a problem step by step before giving its answer, either because the prompt requests it or because the model was trained to reason this way. It improves performance on multi-step problems. Asking a model to think step by step encourages it to decompose a problem—identify the rule, apply it to the facts, then conclude—which helps on logic, math, and analysis tasks. Newer reasoning-focused models generate extended chains of thought automatically before answering. The visible reasoning is itself model output rather than a guaranteed record of computation, so a plausible-looking chain of reasoning still requires checking. - Glossary page: https://legalai.help/term/chain-of-thought/ ### Closed model An AI model whose weights remain private, accessible only through the provider's app or API—as with the leading commercial models. Users send data to the provider's servers and accept its terms to use the model. Closed models are typically the most capable available and come with managed infrastructure, safety tuning, and support, but the model itself cannot be inspected or self-hosted. Data handling is governed by the provider's contract, making terms about retention, training on inputs, and security certifications the central diligence questions. Most legal AI products are built on closed models accessed through enterprise agreements. - Glossary page: https://legalai.help/term/closed-model/ ### Enterprise AI AI services sold to organizations under commercial terms—typically with administrative controls, security certifications, contractual data protections, and commitments not to train on customer data—as distinct from free or consumer versions of the same tools. The same underlying model often behaves identically in consumer and enterprise versions; what differs is the wrapper: authentication, usage controls, audit logs, retention settings, and a negotiated contract. For law firms the distinction matters because consumer terms frequently permit the provider to use inputs for training, while enterprise terms typically do not. Confidential client material generally belongs only in tools whose terms the organization has reviewed and approved. - Glossary page: https://legalai.help/term/enterprise-ai/ ### Few-shot prompting Including a handful of worked examples in a prompt—inputs paired with the desired outputs—so the AI model learns the pattern and applies it to new material. Often markedly improves accuracy and formatting consistency. The examples demonstrate, rather than describe, what good output looks like: show three clauses labeled the way you want, and the model labels the fourth the same way. No training occurs—the examples sit in the prompt and influence only that exchange. Few-shot prompting is among the most reliable techniques for getting consistent results on repetitive legal tasks such as document coding or clause extraction. - Glossary page: https://legalai.help/term/few-shot-prompting/ ### Hallucination An AI model's confident production of false information—invented facts, quotes, or citations—presented as though true. A structural byproduct of how language models generate text by probability rather than by looking facts up. Because a language model composes each answer token by token from learned patterns, it can produce output that is fluent, specific, and wrong—including case citations that do not exist. Hallucination can be reduced through retrieval-augmented generation and careful prompting, but it has not been eliminated in any current system. Fabricated citations in court filings have led to sanctions against lawyers, and bar authorities generally treat verification of AI output as part of the duty of competence. - Glossary page: https://legalai.help/term/hallucination/ - Reference: https://en.wikipedia.org/wiki/Hallucination_(artificial_intelligence) ### Open model An AI model whose weights are published for anyone to download, run, and modify—on their own hardware if they choose. Contrasts with closed models accessible only through a provider's service. Open models—often called open-weight models, since training data and code may stay private—allow organizations to run AI entirely inside their own infrastructure, keeping data on premises. That control comes with responsibility: the hosting organization must secure, update, and evaluate the model itself. Licenses vary, with some permitting unrestricted commercial use and others carrying conditions, so open does not mean license review can be skipped. - Glossary page: https://legalai.help/term/open-model/ ### Prompt The input a user gives an AI model—an instruction, question, or document—to which the model responds. The quality and specificity of the prompt strongly influence the quality of the output. A prompt can be a single sentence or many pages of instructions, examples, and source material. Everything in the prompt occupies the model's context window and is processed on the provider's systems, so prompts containing client information should only go to tools with appropriate confidentiality terms. Writing effective prompts is a learnable skill known as prompt engineering. - Glossary page: https://legalai.help/term/prompt/ ### Prompt engineering The practice of designing and refining the instructions given to AI models to get accurate, useful, and consistent output—specifying role, context, format, examples, and constraints rather than asking open-ended questions. Common techniques include assigning the model a role, supplying relevant documents, showing worked examples, and asking the model to reason step by step. Small wording changes can produce materially different answers, so teams often standardize tested prompts for recurring tasks such as summarizing depositions or extracting contract terms. Good prompting improves reliability but does not remove the need to verify output. - Glossary page: https://legalai.help/term/prompt-engineering/ - Reference: https://en.wikipedia.org/wiki/Prompt_engineering ### Zero data retention (ZDR) A contractual arrangement in which an AI provider does not store a customer's prompts or outputs after processing the request—data is used to generate the response and then deleted rather than logged. Standard API terms often retain inputs briefly for abuse monitoring; zero-data-retention agreements remove or sharply limit that storage so content exists on the provider's systems only transiently. ZDR reduces exposure if the provider is breached or subpoenaed, and law firms handling privileged material increasingly request it. Verifying what retention actually applies—across the vendor and any subprocessors—is a standard part of AI tool diligence. - Glossary page: https://legalai.help/term/zero-data-retention/ ### Zero-shot prompting Asking an AI model to perform a task with instructions alone—no worked examples included in the prompt. The model relies entirely on what it learned in training to understand the request. Zero-shot prompting works well for tasks common in the model's training, such as summarizing or translating, and it is how most people naturally use chatbots. For specialized formats or judgment calls—classifying clauses against firm-specific criteria, for instance—results often improve when examples are added, turning the request into few-shot prompting. Comparing zero-shot and few-shot results is a quick way to test whether a task needs more guidance. - Glossary page: https://legalai.help/term/zero-shot-prompting/ ## Risk and governance ### AI audit A structured review of an AI system or an organization's AI use, testing whether it works as claimed and complies with applicable policies, contracts, and laws—examining accuracy, bias, security, and data handling. An audit may target a specific tool (does the contract-review system find what it claims?), a vendor (do its practices match its security questionnaire answers?), or the organization's own program (is the AI policy actually followed?). Emerging AI regulations and client outside-counsel guidelines increasingly call for periodic review of this kind. Audits also produce documentation that an organization exercised reasonable diligence, which matters if an AI-related failure is later challenged. - Glossary page: https://legalai.help/term/ai-audit/ ### AI disclosure Informing affected parties—clients, courts, counterparties, or the public—that artificial intelligence was used in producing work or making a decision. Disclosure obligations may arise from court orders, client agreements, regulation, or professional-conduct duties. Disclosure questions in legal practice arise at several levels: some courts have standing orders addressing AI use in filings, some clients require notice or consent through outside-counsel guidelines, and consumer-protection and AI-specific laws increasingly require telling people when they are interacting with an AI system such as a chatbot. Bar authorities generally frame the analysis around existing duties of candor, communication, and competence rather than a single blanket rule. Firms commonly address disclosure expectations in engagement letters and matter-intake checklists. - Glossary page: https://legalai.help/term/ai-disclosure/ ### AI governance The framework of policies, roles, processes, and controls an organization uses to direct and oversee its use of artificial intelligence—covering tool approval, acceptable use, risk review, training, and accountability. Governance translates abstract principles—accuracy, confidentiality, accountability—into concrete mechanisms: who may use which tools, with what data, under what review, and who answers when something fails. In legal organizations it typically spans professional-responsibility duties, client confidentiality commitments, and emerging AI regulation. Mature programs cover the full lifecycle, from vetting and approval through monitoring, incident response, and periodic audit. - Glossary page: https://legalai.help/term/ai-governance/ ### AI policy An organization's written rules for AI use: which tools are approved, what data may be entered, what review output requires, and who is accountable. The baseline document of any AI governance program. Effective policies are specific enough to act on—naming approved tools, prohibited data categories such as client-identifying information in unapproved systems, verification requirements, and disclosure expectations. Law firm policies commonly also address client consent, engagement-letter language, and applicable court rules on AI use. A policy only functions alongside training and accessible approved tools; rules without alternatives tend to produce shadow AI. - Glossary page: https://legalai.help/term/ai-policy/ ### Data residency Where data is physically stored and processed—which country's or region's servers hold it. Residency determines which jurisdictions' laws can reach the data and is a common requirement in client and regulatory commitments. Cloud AI services may process requests in data centers around the world unless the contract pins processing to a particular region, an option some providers offer. Residency matters for legal work because privacy statutes, cross-border transfer rules, and some client engagement terms restrict where matter data may go. Residency is distinct from retention: data can sit in the right country and still be kept longer, or used more broadly, than intended. - Glossary page: https://legalai.help/term/data-residency/ ### Human in the loop (HITL) A design and oversight approach in which a person reviews, corrects, or approves an AI system's output before it takes effect—keeping human judgment between the machine and the consequence. Human-in-the-loop checkpoints are placed where errors would be costly: a lawyer reviews the AI-drafted brief before filing, a supervisor approves an AI-suggested privilege call. The approach works only when review is real—time-pressured rubber-stamping recreates the risk it was meant to remove, a failure mode often called automation bias. Professional-responsibility frameworks generally assume a responsible human remains accountable for AI-assisted work product. - Glossary page: https://legalai.help/term/human-in-the-loop/ - Reference: https://en.wikipedia.org/wiki/Human-in-the-loop ### Model training on inputs A provider practice in which the prompts and content users submit are used to train future versions of the AI model. Whether a tool trains on inputs is a central confidentiality question for legal users. When inputs feed training, information submitted by one user can influence—and in rare cases resurface in—the model's later behavior, placing the data beyond the submitting organization's control. Consumer AI services often reserve this right by default, sometimes with an opt-out, while enterprise and API terms typically promise no training on customer data. Reading the actual data-use terms, rather than the marketing page, is the reliable way to know which applies. - Glossary page: https://legalai.help/term/model-training-on-inputs/ ### Prompt injection An attack on AI systems in which crafted text—typed by a user or hidden in a document, email, or web page the system reads—overrides the model's instructions and steers it into unintended behavior. Language models read their operating instructions and the content they are given as one stream of text, so a document, email, or web page an AI tool ingests can smuggle in wording the model treats as a command. Injected instructions can override the vendor's system prompt, pull information the session has access to, or misdirect an AI agent into unintended actions—and the indirect, hidden-in-content form is the harder problem, because the user never sees it. No complete technical fix exists today; practical mitigations include restricting what data and tools an assistant can reach, filtering untrusted content, and keeping human approval in front of consequential actions. For legal teams it is a vendor-diligence question: before pointing a tool at client matter data, ask how it defends against instructions embedded in the material it reads. - Glossary page: https://legalai.help/term/prompt-injection/ - Reference: https://en.wikipedia.org/wiki/Prompt_injection ### Shadow AI Employees' use of AI tools without the organization's knowledge or approval—pasting work content into personal chatbot accounts, for example. The AI-era version of shadow IT, and a leading source of unmanaged data risk. Shadow AI typically emerges when approved tools lag behind what staff find useful, and it bypasses the contractual and security protections enterprise deployments provide. In a law firm, client information entered into an unapproved consumer tool may be retained by the provider or used for training, creating confidentiality exposure no one has assessed. Common responses pair clear policy and training with sanctioned tools good enough that workarounds feel unnecessary. - Glossary page: https://legalai.help/term/shadow-ai/ ### SOC 2 A widely used independent audit report on a service provider's controls for security, availability, processing integrity, confidentiality, and privacy, developed by the American Institute of Certified Public Accountants (AICPA). A standard item in vendor security diligence. A SOC 2 report is produced by an independent auditor who evaluates a provider's controls against the AICPA's Trust Services Criteria; a Type I report assesses control design at a point in time, while a Type II report tests how controls operated over a period. Law firms routinely request SOC 2 Type II reports when vetting AI vendors that will touch client data. A SOC 2 covers organizational controls—it does not measure whether an AI model's outputs are accurate. - Glossary page: https://legalai.help/term/soc-2/ ## Legal AI applications ### AI intake chatbot A conversational AI assistant on a law firm's website or phone line that greets prospective clients, gathers facts about their matter, screens for fit and conflicts information, and schedules consultations—at any hour, in natural language. Intake chatbots built on language models can ask follow-up questions and summarize a prospect's situation for staff, going beyond rigid web forms. They raise specific compliance considerations: avoiding unauthorized legal advice, making clear that no attorney-client relationship is formed, handling submitted information confidentially, and—in some jurisdictions—disclosing that the visitor is talking to an AI. Well-implemented deployments route every qualified lead to a human and log conversations for review. - Glossary page: https://legalai.help/term/ai-intake-chatbot/ ### Contract analysis AI AI tools that read contracts to extract key terms, flag risks and deviations from preferred positions, compare drafts against playbooks, and summarize obligations—across single agreements or entire contract portfolios. These tools combine language models with extraction and comparison logic to answer questions such as which agreements lack a limitation-of-liability cap or where change-of-control consent is required. They are used in due diligence, contract review, and post-signature obligation management. Extraction accuracy varies with document quality and clause complexity, so material findings—especially deal-blocking ones—are verified against the underlying text. - Glossary page: https://legalai.help/term/contract-analysis-ai/ ### Document automation Software that assembles legal documents from templates and structured inputs—answers to a questionnaire populate clauses, names, and terms to produce a consistent draft. Increasingly augmented with generative AI for drafting flexibility. Classic document automation is deterministic: approved template language plus client data yields predictable output, which is why it has long been trusted for wills, leases, and corporate documents. Generative AI adds the ability to draft or adapt language beyond the template, trading some predictability for flexibility. Many firms combine the two—templates for settled language, AI for bespoke provisions—while keeping attorney review before anything is signed or filed. - Glossary page: https://legalai.help/term/document-automation/ - Reference: https://en.wikipedia.org/wiki/Document_automation ### E-discovery AI AI used in electronic discovery to identify, classify, and prioritize documents in litigation and investigations—finding responsive material, flagging potential privilege, and surfacing key facts across collections far too large for page-by-page human review. Machine learning has been used in e-discovery longer than in most legal domains, beginning with technology-assisted review for responsiveness classification; newer generative tools add summarization, timeline building, and natural-language querying of document sets. Courts have generally accepted properly validated AI-assisted review workflows. Defensibility rests on documented process—sampling, validation, and quality control—rather than on any particular tool. - Glossary page: https://legalai.help/term/e-discovery-ai/ - Reference: https://en.wikipedia.org/wiki/Electronic_discovery ### Legal AI Artificial intelligence applied to legal work—research, document review, drafting, contract analysis, e-discovery, and client intake. Most current legal AI is built on large language models adapted with legal data sources and retrieval. Legal AI spans purpose-built platforms from legal technology vendors as well as general-purpose assistants used with legal judgment. Current tools are strongest at first drafts, summarization, extraction, and finding relevant material across large document sets, with the supervising professional responsible for the final product. Bar authorities have generally concluded that existing duties—competence, confidentiality, supervision, candor—govern AI use rather than prohibiting it. - Glossary page: https://legalai.help/term/legal-ai/ ### Legal research AI AI tools that answer legal questions and find authority by searching case law, statutes, and secondary sources, then generating a synthesized, cited response—typically using retrieval-augmented generation over a curated legal database. These tools pair a language model with an authoritative legal corpus so answers are grounded in retrievable sources rather than the model's memory. They speed issue-spotting and first-pass research, but they can still mischaracterize holdings or miss controlling authority, so cited sources must be read, not just collected. Confirming that every cited case exists and says what the tool claims remains the researcher's professional responsibility. - Glossary page: https://legalai.help/term/legal-research-ai/ ### Technology-assisted review (TAR) An e-discovery workflow in which reviewers code a sample of documents and a machine-learning model learns from that coding to classify or rank the rest of the collection, sharply reducing manual review. In common TAR workflows the model continuously learns from reviewer decisions and serves up the documents most likely to be responsive—an approach known as continuous active learning—until little relevant material remains unfound. TAR predates generative AI and relies on classification rather than text generation, which makes its accuracy measurable through statistical sampling. Courts have repeatedly accepted TAR when the process is transparent and validated, and negotiating TAR protocols with opposing parties is now routine in large matters. - Glossary page: https://legalai.help/term/technology-assisted-review/