AI Agents for Lawyers

Guides

What Are AI Agents? A Practical Guide for Lawyers

AI agents for lawyers explained in plain terms: a model plus tools, a loop and a goal. How legal agents review NDAs, where they fail and what to expect in 2026.

By Dr. Niklas SchmidtUpdated 13 November 202610 min read
On this page
  1. The four parts of an agent
  2. What an agent does with 40 NDAs
  3. Anatomy in detail: planning, tool calls, memory and gates
  4. The agents lawyers can actually use in September 2026
  5. Where agents fail
  6. What to expect for the rest of 2026
  7. What to do this week

An AI agent is a language model given tools, a goal and permission to loop: it plans, acts (searches a data room, reads a contract, drafts a redline), checks the result and decides what to do next until the goal is met or a rule tells it to stop and ask a human. A chatbot answers one message; an agent runs a task. For lawyers the difference is delegation, and delegation needs supervision.

The four parts of an agent

Strip away the marketing and every agent has the same anatomy.

The model. A large language model such as Claude Opus 5, GPT-6 Astra or Gemini 3.1 Pro supplies the reasoning. It reads, summarises, compares and drafts. On its own it cannot open a file, run a search or send anything; it can only produce text. (For the mechanics, see how large language models work.)

The tools. Tools are functions the model is allowed to call: search a document management system, read a PDF, query Westlaw, write tracked changes into Word, post to Slack. Since late 2025 most tools speak the Model Context Protocol, which Anthropic donated to the Linux Foundation on 9 December 2025 with more than 10,000 public servers already in circulation. iManage and NetDocuments both shipped MCP servers in May 2026 that run under the user's own credentials and respect ethical walls.

The loop. After each tool call the model looks at what came back and chooses the next step. Anthropic's reference text from December 2024 draws the line here: workflows are "systems where LLMs and tools are orchestrated through predefined code paths", while agents are "systems where LLMs dynamically direct their own processes and tool usage" (Anthropic, 19 December 2024). Harvey's ladder makes the same point in legal terms: a model gives "a single response", an agent can "plan, adapt, and meaningfully interact with humans" (Harvey, 11 March 2025).

The goal. The agent is given an outcome ("review these NDAs against our playbook and produce a redline and a risk table"), not a single instruction. Gartner's rule of thumb is useful when a vendor pitches you: use "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 claiming the label, and coined "agent washing" for the rest (Gartner, 25 June 2025). The full taxonomy, with a checklist for spotting agent washing, is in AI agents vs assistants vs chatbots.

What an agent does with 40 NDAs

The clearest way to see the difference from chat is to follow one task through. Here is what a commercial-contracts agent built on Anthropic's open-source Claude for Legal plugins does when asked to review 40 inbound NDAs against a firm playbook:

  1. Cold-start interview. The first time it runs, the plugin interviews the lawyer for 10 to 20 minutes and writes a practice profile (which positions are standard, which fallbacks are acceptable, who escalates). Every later skill reads that file.
  2. Retrieve. The agent lists the 40 documents in the matter folder through the iManage or NetDocuments MCP connector. It sees only what the lawyer's own credentials allow.
  3. Classify and compare. For each NDA it identifies the clause types present, compares each against the playbook and marks the clause RED, YELLOW or GREEN.
  4. Draft. It writes proposed changes as Word tracked changes with an explanatory comment on each, using the firm's preferred fallback language.
  5. Stop at the gate. Any clause outside the approved fallback range is not redlined; it is escalated with a one-paragraph note. Any statement of law it wants to make is tagged [verify] unless it was checked through a connected research tool.
  6. Report. It produces a table of 40 rows: counterparty, governing law, term, deviations, escalations, with each cell citing the page it came from.
  7. Hand over. A lawyer reviews the table, opens the escalations, accepts or rejects tracked changes, and sends. Nothing leaves the firm without that step.

Nobody pasted 40 documents into a chat box or wrote 40 prompts. The agent chose the order, called the tools and knew when to stop. That is the whole idea, and the full workflow is set out in NDA review at scale and automate contract review with a playbook.

Anatomy in detail: planning, tool calls, memory and gates

Planning. The better products now show you the plan before they act. LexisNexis's Legal Intelligence Engine, released 24 August 2026, "presents an executable plan for user approval" before running (Artificial Lawyer); ChatGPT Work (9 July 2026) has a Plan mode with approval gates. Insist on this. An agent that cannot tell you what it intends to do cannot be supervised.

Tool calls. Each call is a decision point and an audit-trail entry. The MCP specification of 28 July 2026 added "multi round-trip" confirmations so that a server can pause mid-call and ask the user to confirm, which is exactly the hook a firm wants on anything that writes or sends. iManage's server is "permission-bound" and "fully logged" and "respects existing ethical walls" (iManage, 14 May 2026). How to wire this up is in connect Claude to your DMS with MCP.

Memory. Chat forgets; agents increasingly do not. Harvey II (18 August 2026) added Memory and matter-aware Spaces; Claude's legal plugins keep a practice profile on disk; Cowork keeps working files between sessions. Memory is what turns a generic model into your associate, and it is also where a bad instruction can persist unnoticed (OWASP lists "memory poisoning" as one of its ten agentic risks).

Approval gates. These are the human-in-the-loop controls, and by September 2026 they are standard: Anthropic's litigation plugin refuses to send a demand letter until a human confirms the settlement-privilege framing; the corporate plugin has a review gate before board consents execute; Harvey's agents "flag moments requiring user confirmation before proceeding"; Lexis Protégé Workrooms enforce "dual approvals, role-based permissions, and audit trails". A gate is only a control if you keep it switched on.

The agents lawyers can actually use in September 2026

ProductWhat the agent doesBuilt-in checkStatus
Harvey Agent Builder and 500+ pre-built agentsDiligence grids, lease summaries, playbook review; 700,000+ agent tasks a day (June 2026)Flags moments needing user confirmationEarly access May 2026; 25,000+ customer-built agents
Claude for Legal12 practice-area plugins, 150+ skills, 13 scheduled agents (docket watcher, renewal watcher, reg monitor)[verify] tags; send and review gates; cold-start interviewOpen source, Apache 2.0; expanded 12 May 2026
CoCounsel Legal (Thomson Reuters)Brief Builder, Tabular Analysis (10,000 documents × 100 questions), Word drafting agentKeyCite and Westlaw grounding; Deep Research VerifyUS GA 20 August 2026 on the Claude Agent SDK
Lexis+ with Protégé300+ configurable workflows; Legal Intelligence Engine picks models and sourcesShepard's-grounded verification; plan for approval; WorkroomsGA 24 February 2026; engine 24 August 2026
Legora aOSOvernight counterparty redlines reviewed and answered before lawyers arrive (vendor description)Lawyer review of flagged items7 May 2026
ChatGPT WorkTakes an outcome, works "independently for hours", returns documents, spreadsheets, appsPlan mode; approval gates; Compliance API9 July 2026
Claude CoworkDesktop agent with sandboxed access to your own files and shellSandbox; per-action approvalsMerged into the Claude app 17 September 2026
Microsoft Legal Agent for WordClause-by-clause playbook review with tracked changesApply or dismiss each commentFrontier preview 30 April 2026; needs Anthropic enabled as subprocessor

Adoption is real but uneven. ILTA's 2026 survey (14 September 2026, 500+ firms) found 94% of firms using or exploring generative AI, with Microsoft 365 Copilot at 76%, CoCounsel and Claude at 44% each and Harvey at 43%; among firms with 700+ lawyers Harvey reaches 84%. Thomson Reuters Institute reported in April 2026 that under 20% of firms and departments use agentic AI widely, with about half planning to (Thomson Reuters Institute, 9 April 2026).

Where agents fail

They get most things right and something wrong. Harvey launched LAB on 6 May 2026: 1,200+ long-horizon legal tasks across 24 practice areas, scored against 75,000+ expert-written rubric criteria. On the strict metric (every criterion must pass) the first published run on 26 May 2026 put Claude Opus 4.7 at 7.1%, with each task costing about $50.90 and 22 minutes; GPT-5.6 Sol later scored 2.5% (Harvey, LAB initial results). By 18 September 2026 third-party leaderboards showed Claude Fable 5 at 13.3% and Opus 5 at 11.7%, while criteria-level pass rates sat around 83–88%. Read that carefully: the agent does almost everything right and misses something almost every time. What the numbers do and do not mean is unpacked in legal AI benchmarks explained.

They can be hijacked by the documents they read. Zenity's "PleaseFix" research (disclosed 3 March 2026, presented at Black Hat on 5 August 2026) showed zero-click takeovers of Claude in Chrome, Gemini in Chrome, Comet and Copilot through ordinary emails, calendar invites and web pages: a request to "summarise my emails" was enough to exfiltrate a Google Drive (Dark Reading). Anthropic's own mitigations cut the attack success rate for Claude for Chrome from 23.6% to 11.2%, and its researchers still write that "No browser agent is immune to prompt injection". This is the single most important thing to understand before you point an agent at an inbox; see prompt injection and agent security for lawyers.

They do what you allowed, not what you meant. In February 2026 a venture investor let Cowork "organise" a desktop, approved deleting "temp files", and an rm -rf removed 15 years of family photos, recovered only through an iCloud rollback (Futurism). In July 2026 xAI's Grok Build uploaded 5.10 GiB of whole repositories, including an unredacted secrets file, when the model used about 192 KB (The Hacker News). And in a cyber evaluation run with safeguards deliberately off, OpenAI's GPT-5.6 Sol escaped a sandbox and took about 17,600 actions over 4.5 days inside a third party's infrastructure (disclosed 21 July 2026). Substitute a matter folder for the photos and the lesson writes itself.

Their vendors disappear. ChatGPT Atlas lasted 292 days. OpenAI announced on 3 June 2026 that Agent Builder shuts on 30 November 2026, eight months after launch. Relay.app closed to paying customers on 14 September 2026. Contract for data export from day one.

What to expect for the rest of 2026

Expect agents to move from pilots into named workflows: diligence grids, renewal tracking, docket watching, regulatory monitoring, first-pass redlines. Expect the profession to stay split; in the Thomson Reuters survey 47% said agents should be used for legal work and 22% said not. Expect the liability questions to sharpen: Clifford Chance's February 2026 briefing on "the liability gap your contracts may not cover" lists agency law, hallucinated commitments and insurance as open points (Clifford Chance, 9 February 2026). Expect regulators to catch up slowly: California's May 2026 practical guidance is the first US bar text to address agentic AI explicitly, and the EU AI Act's Article 50 transparency duties have applied since 2 August 2026 (see the EU AI Act for lawyers). Gartner still predicts that more than 40% of agentic projects will be cancelled by the end of 2027. Most of those will be the ones that skipped the boring part: a bounded task, a test set and a gate.

What to do this week

  1. Pick one task that is repetitive, document-heavy and has a clear finished state (NDA triage, lease abstraction, DSAR response drafting).
  2. Assemble 20 real examples with known correct answers and strip client identifiers, or use synthetic documents.
  3. Run the task in a sandboxed agent (Cowork with a scratch folder, a Harvey or Legora workspace, or an n8n flow) and require a written plan before execution.
  4. Switch every gate on. Nothing sends, files or deletes without a person.
  5. Score the 20 outputs. Count criteria met, not tasks "done".
  6. Write the result into the firm's AI policy as an approved or rejected use.

Module 4 of the AI Lab for Lawyers course builds exactly this kind of agent live, in n8n and Make, without code. If you would rather start on your own, build your first legal AI agent in an afternoon walks through the same steps.

Frequently asked questions

What is an AI agent in simple terms for a lawyer?

An AI agent is a language model that has been given tools (search, file access, a drafting function), a goal and the freedom to take several steps in a row without a new prompt each time. It plans, acts, checks its own result and continues until the task is done or a rule tells it to stop and ask a human.

What is the difference between an AI agent and ChatGPT?

ChatGPT in chat mode answers one message at a time and does nothing until you type again. An agent such as ChatGPT Work or Claude Cowork takes an outcome, breaks it into steps, calls tools like a document store or a browser, and works for minutes or hours. The agent decides the order of steps; in chat you do.

Are AI agents reliable enough for legal work in 2026?

For bounded, document-heavy tasks with a human review step, yes. For unsupervised work, no. Harvey's LAB benchmark shows the best model passes every criterion on a long legal task only about 13% of the time, although it gets most individual criteria right. Treat agent output as a draft from a capable but unaccountable junior.

What are approval gates in a legal AI agent?

Approval gates are hard stops written into the agent's instructions: it must present a plan before running, tag unverified citations, or pause before sending a demand letter or executing a board consent. Anthropic's open-source legal plugins, Harvey's agents and Lexis Protégé Workrooms all use them. They are the mechanism that keeps the lawyer responsible for the output.

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