Legal AI arrived with its own vocabulary, mostly undefined. This glossary covers the 28 terms a solo or small-firm lawyer meets when evaluating AI-native practice management - plain English, with the practice angle vendor pages skip. Terms run alphabetically.
The terms, A to Z
Agentic workflow - A multi-step process an agent runs end to end: gather inputs, act in sequence, hand the result to a person for review. Opening a matter from a signed engagement letter is an agentic workflow; answering one question is not.
AI agent - Software that uses a language model to plan and execute tasks, not only to generate text. It reads the situation, decides the next step, and acts through connected tools. See AI agents for law firms.
AI-assisted firm - A firm where lawyers use AI inside individual tasks - drafting, summarizing, research - while the workflow still runs on people retyping data between systems. Most firms in 2026 are here.
AI-native firm - A firm whose operations are built around agents doing routine work while lawyers supervise and decide. The distinction is architectural, not a count of AI tools purchased. The longer treatment: what is an AI-native law firm.
Approval loop (human-in-the-loop) - A control that stops an agent before a consequential action and waits for a person to sign off. This is where professional responsibility lives: the AI prepares, the lawyer approves. Referent applies it to every client-facing or high-risk action.
Audit trail - A log of who or what did each thing, and when. When agents act on matters, the trail is how you supervise after the fact - and show that you did.
Data isolation - Keeping each firm’s data in its own separated workspace so one firm’s matters never surface in another’s results. A first-round diligence question for any tool that touches client files.
E-billing readiness - Time and activity captured in structured, coded form - matter, task, timekeeper, narrative - clean enough for a client’s billing system to accept without rejections. Agents help by assembling the record as work happens, not at month end.
Grounding - Constraining a model’s answer to specific source material you supply - a matter file, a contract, a statute - rather than its general training data. Grounded output can be checked against the source.
Hallucination - A fluent, confident statement by a model that is false: an invented citation, a wrong date, a nonexistent clause. The failure mode that makes lawyer review non-negotiable.
Intake automation - Handling a new inquiry with software from first contact to opened matter: capture, qualification, engagement, payment. Usually the first workflow small firms automate, because unanswered inquiries are lost revenue.
ISO 27001 - The international standard for running an information security management system. Certification means an accredited auditor verified the vendor’s security processes, not only its intentions.
LLM (large language model) - The model class behind modern AI assistants: trained on large volumes of text, it predicts language, reasons over documents, and drives agents. GPT, Claude, and Gemini are LLMs.
Matter context - Everything the firm knows about a matter - parties, documents, correspondence, deadlines, billing - available to an agent in one place. Without it, AI gives generic answers; with it, the answer knows which case you mean.
Matter lifecycle - The path a matter travels from inquiry to close: intake, engagement, active work, resolution, billing, archive. AI-native systems attach agents to specific stages of that path.
MCP (Model Context Protocol) - An open standard, originated by Anthropic in November 2024, connecting AI assistants to external tools and data through one interface. Supported by Claude, ChatGPT, Gemini, Perplexity, and open-source agents such as OpenClaw and Hermes.
MCP client - The assistant side of an MCP connection: the AI app that discovers a server’s tools and calls them. Claude or ChatGPT acts as an MCP client when you add a connector.
MCP server - The tool side of the connection: a service that exposes data and actions for MCP clients to use. Referent runs one at mcp.referent.law, so an assistant you already use can read and act on your matters with the same permissions, approvals, and audit trail as the app - details at /mcp/.
Persistent memory - An agent’s ability to retain knowledge across sessions instead of starting blank each time. Useful for standing preferences; also a confidentiality surface to manage.
Prompt - The instruction you give a model. In agent systems, prompts harden into standing policies the agent follows on every run, so precision matters more than clever phrasing.
RAG (retrieval-augmented generation) - The system retrieves the relevant documents first, then has the model answer from them - how software grounds answers in your files rather than the model’s memory.
Sandboxing - Running an agent’s actions inside a restricted environment so a mistake, or a malicious instruction hidden in a document, cannot reach beyond it. Ask vendors what their agents cannot touch.
Self-hosting - Running software on hardware you control, as with open-source agents like OpenClaw or Hermes. Maximum control; the operator owns security patching, confidentiality, and supervision. In a law firm, the operator is you.
Skill (agent skill) - A packaged capability an agent can execute: a script, a procedure, a checklist. Some agents write their own - Hermes Agent, released February 2026, documents a skill after completing a complex task and reuses it later.
SOC 2 - A US attestation framework in which an independent auditor examines a vendor’s security controls. Type 1 assesses design at a point in time; Type 2 assesses operation over months. Request the report during vendor diligence.
System of action - Software that executes work rather than only storing it: drafts the reply, assembles the bill, opens the matter, subject to approval. The category line is blunt: legacy software records the work - Referent executes it, and the lawyer signs.
System of record - The authoritative database where client, matter, and billing data lives. Traditional practice management is a system of record: it stores what happened. Agents can work alongside one over API, or the data can migrate.
Zero data retention (no-training commitment) - A vendor promise that your data is not kept beyond processing and never used to train models. Get it in writing; confidentiality duties extend to every tool that touches client data.
Where should you go next?
Start with what is an AI-native law firm for the argument behind the vocabulary, then AI agents for law firms. The security page shows the trust terms as shipped controls. The rest of the atlas lives at /ai-native/; the Cohort 2 waitlist is at /apply/.
Frequently asked questions
What is the difference between an AI-assisted and an AI-native law firm?
An AI-assisted firm uses AI inside individual tasks - a lawyer drafts with it, summarizes with it, researches with it - while the workflow itself still runs on people moving data between systems. An AI-native firm is built the other way around: agents execute the routine workflow, and lawyers review and approve at defined checkpoints. The difference is architectural, not a matter of how many AI tools the firm has bought. Most firms in 2026 are AI-assisted; AI-native is the destination this atlas describes.
Which of these terms matter most when evaluating legal software?
Start with the trust vocabulary: approval loop, audit trail, data isolation, and zero data retention. Those four determine whether an AI product is compatible with your confidentiality and supervision duties. Then ask about SOC 2 and ISO 27001 status, because they show whether an independent auditor has verified the vendor's claims. Feature lists change monthly; these fundamentals do not.
Do lawyers need to understand MCP?
A working knowledge helps, because MCP is becoming the standard way AI assistants connect to practice tools. You do not need the technical details - only the model: a server exposes your firm's data and actions, a client such as Claude or ChatGPT connects to it, and the permissions and logging of the underlying system should carry over. If a vendor offers an MCP server, ask whether authentication, permissions, and audit logging match what you get in the app.
Is it safe to run an open-source agent like OpenClaw or Hermes on client matters?
They are capable general-purpose frameworks, and running them locally keeps data on your own hardware. But out of the box they have no approval workflow, no audit trail, and no legal-specific guardrails, and self-hosting means you own security patching, confidentiality, and supervision personally. Your professional duties do not shrink because the tool is free or local. Treat them as power tools for experimentation, and be deliberate before pointing them at confidential files.