AI Agents for Lawyers

Glossary

AI Agent (Definition for Lawyers)

An AI agent is a language model with tools, a goal and permission to take several steps on its own. Definition for lawyers, an NDA triage example, failure data and four questions to ask.

By Dr. Niklas SchmidtUpdated 13 November 20265 min read
On this page
  1. Chatbot, assistant or agent: the one-line test
  2. Worked example: an NDA triage agent
  3. Where agents fail, with numbers
  4. Four questions before you deploy one

An AI agent is a large language model that has been given tools, a goal and permission to take several steps without a fresh prompt each time. It reads, decides, calls a tool (search a document store, draft a redline, post to a queue), inspects the result and chooses what to do next. A chatbot answers one message. An agent runs a task, and someone must supervise it.

Chatbot, assistant or agent: the one-line test

The word is applied to everything from a search box to an autonomous deal reviewer, so the definition needs an edge. Anthropic's engineering team drew it in December 2024 and the industry still uses their wording: workflows are "systems where LLMs and tools are orchestrated through predefined code paths", agents are "systems where LLMs dynamically direct their own processes and tool usage" (Anthropic, 19 December 2024). Harvey's version for lawyers: a model gives "a single response", while an agent can "plan, adapt, and meaningfully interact with humans" (Harvey, 11 March 2025).

The practical test is who decides the next step at run time. If the model does, it is an agent. If you do, it is a chatbot; if a fixed script does, it is automation.

ChatbotAssistant (in Word or Outlook)Agent
Who chooses the next stepYou, one message at a timeYou; it suggests, you acceptThe model, until done or gated
ToolsNone, or web searchThe document in front of youDocument stores, search, drafting, email, scheduling
Runs forSecondsSecondsMinutes to hours
Examples (September 2026)Claude or ChatGPT in chat modeMicrosoft 365 Copilot, Spellbook in WordChatGPT Work, Claude Cowork, Harvey agents, LegalOn Triage
Needs a log and gatesNoRarelyAlways

Gartner's rule for buyers: "AI agents when decisions are needed, automation for routine workflows and assistants for simple retrieval". Gartner also counted only around 130 genuinely agentic vendors among thousands using the label and named the rest "agent washing" (Gartner, 25 June 2025). The fuller taxonomy is in AI agents vs assistants vs chatbots.

Worked example: an NDA triage agent

This is the task most in-house teams automate first. An inbound NDA arrives and the agent:

  1. Extracts the parties, term, governing law and the positions your playbook cares about (mutuality, residuals, non-solicitation, carve-outs).
  2. Classifies the document against the playbook: within tolerance, negotiable, or unusual.
  3. Acts according to the class. Within tolerance: it drafts an approval note. Negotiable: it produces a tracked-changes redline with a reason for each change. Unusual: it stops and escalates to a named lawyer with a two-line summary.
  4. Waits at the gate. Nothing leaves the firm. The redline lands in a review queue and a person sends it.
  5. Logs every extraction, decision and tool call, so the reviewer can see why a clause was flagged.

None of this is hypothetical. LegalOn's Triage agent, launched 10 February 2026, "auto-approves low-risk contracts or escalates" the rest, alongside Playbook, Intake, Translate and Draft agents; the company reported 9,000+ customers by August 2026 (LegalOn, 10 February 2026). Anthropic's open-source commercial-legal plugin does the same job into Word tracked changes and adds a 90-day renewal tracker; it begins with a "cold-start interview" that writes your playbook into a profile every later skill reads (Claude for Legal repository, 18 September 2026). Steps 4 and 5 separate an agent you can defend from one you cannot. The volume version of this workflow is in NDA review at scale.

Compare the chatbot route: paste the NDA into a consumer chat window and ask for problems. You get one opinion, no playbook, no redline, no log, and a client document sitting in a tool whose terms may allow training. The Upper Tribunal was blunt in UK v SSHD [2026] UKUT 81 (IAC): pasting client letters into public AI tools puts them "in the public domain".

Where agents fail, with numbers

Autonomy is what makes agents useful and what makes them dangerous.

  • They finish imperfectly. On Harvey's LAB benchmark, which requires every expert rubric criterion to pass, Claude Opus 4.7 scored 7.1% on the initial run, at about $50.90 and 22 minutes per task (Harvey, 26 May 2026). The 18 September 2026 leaderboard puts Claude Fable 5 at 13.3% (llm-stats). Criteria-level pass rates sit around 83–88%: most of the work is right and something is wrong somewhere.
  • They do what they were told, literally. In February 2026 a venture investor let Claude Cowork "organise" a desktop, approved a clean-up of temporary files and lost 15 years of photos to an rm -rf, recovered only through an iCloud rollback (Futurism, February 2026). Substitute a matter folder.
  • They can be hijacked by what they read. Zenity's "PleaseFix" research (disclosed 3 March 2026, presented at Black Hat on 5 August 2026) showed a single poisoned email turning a request to summarise an inbox into exfiltration of Gmail and Drive contents across Claude in Chrome, Gemini in Chrome, Comet and Copilot Edge (Dark Reading). Anthropic's own position, after cutting attack success to about 1% with Opus 4.5: "No browser agent is immune to prompt injection" (Anthropic, 24 November 2025). The controls are in the agent security guide.

Four questions before you deploy one

  1. What can it read, and does that include an inbox or a shared drive you do not control?
  2. What can it write, send or delete, and is every external action gated by a person?
  3. Where is the log, who can read it, and how long is it kept?
  4. What is the test set, and what score on it would make you stop?

Module 4 of the AI Lab for Lawyers course builds an agent of this shape in n8n and Make, gates included. For the long-form treatment start with What are AI agents?; the adjective has its own entry at agentic AI.

Frequently asked questions

What is an AI agent in one sentence?

An AI agent is a large language model that has been given tools, a goal and permission to take several steps in a row, deciding after each step what to do next, until the task is finished or a rule tells it to stop and ask a person. The autonomy is the defining feature and the reason it needs supervision.

Is ChatGPT an AI agent?

In ordinary chat mode, no: it answers one message and waits. ChatGPT Work, launched on 9 July 2026, is an agent: it takes an outcome, plans, calls tools and works for hours, returning documents or spreadsheets, with plan mode and approval gates. The same product family can be a chatbot or an agent depending on the mode you use.

What is an example of an AI agent in legal work?

NDA triage. An inbound NDA is extracted against the firm playbook, classified as within tolerance, negotiable or unusual, then approved, redlined with tracked changes or escalated to a named lawyer, with nothing sent externally until a person releases it. LegalOn's Triage agent (February 2026) and Anthropic's open-source commercial-legal plugin both work this way.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion