AI Agents for Lawyers

Guides

AI Agents vs Assistants vs Chatbots: The Difference for Legal Work

AI agent vs chatbot vs assistant, explained for lawyers: a five-tier taxonomy with examples, risk profile and supervision needs, plus a checklist for spotting agent washing.

By Dr. Niklas SchmidtUpdated 13 November 20269 min read
On this page
  1. The five tiers at a glance
  2. Tier 1: chatbots, the same model in its safest form
  3. Tier 2: assistants and copilots, the model inside your application
  4. Tier 3: workflows, the fixed script that runs every time
  5. Tier 4: agents, the model chooses the path
  6. Tier 5: multi-agent systems, delegation all the way down
  7. Is it really an agent? A checklist for vendor pitches
  8. Which tier for which task

The difference between a chatbot, an assistant and an agent is who decides the next step. In a chatbot you do, one message at a time. In an assistant the software suggests and you accept. In a workflow a fixed script decides. In an agent the model decides, calling tools and looping until the goal is met. Risk, supervision and price all rise in that order.

The five tiers at a glance

TierWho decides the next stepWhat it can do in a firmLegal examples (September 2026)Risk profileSupervision needed
1. ChatbotYou, one turn at a timeSummarise, draft, explain, brainstorm from what you pasteChatGPT chat, Claude chat, Gemini app, Mistral VibeConfidentiality of what you paste; hallucinationRead and verify every output
2. Assistant / copilotSoftware suggests, you acceptRedline in Word, draft replies in Outlook, answer questions about an open documentMicrosoft 365 Copilot, Legal Agent for Word, Spellbook, Claude for Word, Lexis CreateSame, plus silent acceptance of wrong editsSpot-check accepted changes; keep tracked changes
3. WorkflowA script you or the vendor configuredRun the same multi-step process every time: intake, triage, extraction, tabular reviewLexis+ Protégé's 300+ workflows, Harvey Workflows, Legora Tabular Review, n8n, Zapier, MakeErrors repeat at scale; bad inputs propagateAudit logs; sample outputs; test on changes
4. AgentThe model, within limits you setTake an outcome and work for minutes or hours: research, redline, draft, file into DMSChatGPT Work, Claude Cowork, Lexis Legal Intelligence Engine, Harvey Agent Builder agents, Legora aOS, Claude for Legal pluginsUnintended actions; prompt injection; data leaving scopePlan-for-approval; gates on write/send; full trace
5. Multi-agentAn orchestrating model delegating to othersParallel sub-tasks across a data room or a litigation bundle; long-running monitorsPerplexity Computer, Claude Managed Agents cookbooks (diligence-grid, docket-watcher), Freshfields' multi-agent workflows on VertexCascading failures; unclear accountabilityEverything in tier 4 plus per-agent scoping and kill switches

The tiers are not a ranking. A tier-3 workflow that extracts 40 fields from 10,000 leases is worth more to most firms than a tier-4 agent that improvises. Gartner's rule from June 2025 still holds: "AI agents when decisions are needed, automation for routine workflows and assistants for simple retrieval" (Gartner).

Tier 1: chatbots, the same model in its safest form

A chatbot is a language model behind a text box. It cannot open your DMS, cannot send email and cannot remember beyond the conversation unless the product adds memory. Its risks are exactly two: what you paste in (consumer tiers of ChatGPT, Claude and Gemini may use it for training unless you opt out, and the Upper Tribunal has said pasting a client letter into a public tool puts it "in the public domain") and what it makes up. Both are manageable with a business-tier account and the discipline in the confidentiality checklist and verification protocol.

The chat tier is also where most lawyers still work. Clio's 2025 Legal Trends Report found 79% of legal professionals using AI while use of legal-specific tools fell from 58% to 40% as people defaulted to general chatbots. There is nothing wrong with that; a well-prompted chat session with a matter's documents uploaded is a legitimate way to work. It is simply not an agent.

Tier 2: assistants and copilots, the model inside your application

An assistant lives where the work is: Microsoft 365 Copilot in Word and Outlook (76% of firms in ILTA's September 2026 survey), Microsoft's Legal Agent for Word (Frontier preview from 30 April 2026, which despite its name works as an assistant: it reviews a contract against a playbook and offers each redline for you to apply or dismiss), Spellbook, Claude for Word, Lexis Create. The software proposes; you dispose.

The risk is subtle. Because accepting a suggestion takes one click, assistants invite the "cognitive surrender" that Scientific American identified in May 2026 as the reason lawyers keep filing fake citations despite every warning. The control is procedural: keep tracked changes on, review the redline as you would a trainee's, and never accept all. The Word Legal Agent review and Copilot for lawyers go into the settings.

Tier 3: workflows, the fixed script that runs every time

Anthropic's definition is the cleanest: workflows are "systems where LLMs and tools are orchestrated through predefined code paths" (Anthropic, December 2024). Someone, you or the vendor, decided the steps in advance. The model does the reading and writing inside each step; it does not choose the steps.

This is the tier that has quietly delivered most of the measurable results in legal AI. Lexis+ with Protégé launched on 24 February 2026 with 300+ configurable workflows and a no-code builder. Legora's Tabular Review reads up to 100,000 documents against a fixed set of questions. Thomson Reuters' Tabular Analysis runs 10,000 documents against 100 questions. A three-lawyer German firm runs intake, onboarding and deadline capture on self-hosted n8n with a local model for €20–30 a month. In each case the process is the same every time, which is precisely why it can be audited, tested and trusted. The workflow agent entry and the n8n for law firms guide show how to build one.

The danger of workflows is that a mistake repeats at scale, silently. Test every change on a held-out set of documents before it goes live.

Tier 4: agents, the model chooses the path

Now the model directs its own process. Give ChatGPT Work (launched 9 July 2026) an outcome and it works "independently for hours", returning documents, spreadsheets or slides, with a Plan mode and approval gates in between. Claude Cowork (merged into the main Claude app on 17 September 2026) does the same on your own desktop files inside a sandbox. Lexis's Legal Intelligence Engine (24 August 2026) is described by its own CTO as "an agentic harness" that picks models, sources and skills and "presents an executable plan for user approval" before running (Artificial Lawyer). Legora's aOS (7 May 2026) promises overnight counterparty redlines "reviewed, flagged and responded to by the time lawyers arrive" (Artificial Lawyer).

The contrast within one vendor's line-up is the clearest illustration:

Same vendorChat or assistantAgent
OpenAIChatGPT chat: one reply per messageChatGPT Work: outcome in, deliverables out, hours later
AnthropicClaude chat and Claude for WordClaude Cowork and Claude for Legal plugins with file, shell and connector access
LexisNexisProtégé workflows: fixed, configurable stepsLegal Intelligence Engine: chooses models, sources and skills itself
HarveyAssistant and Vault Q&AAgent Builder agents and 500+ pre-built agents that "flag moments requiring user confirmation"

What changes is the blast radius. The same model that is harmless in chat can, with write access, delete a folder (a Cowork user lost 15 years of photos in February 2026 after approving "delete temp files") or be hijacked by a poisoned email (Zenity's zero-click "PleaseFix" attacks on Claude in Chrome, Comet and Copilot, disclosed March 2026). Supervision therefore has to move from reading outputs to controlling actions: a written plan before execution, a gate on anything that writes, sends, files or deletes, and a trace you can read afterwards. What are AI agents explains the anatomy; agent security for lawyers covers the controls.

Tier 5: multi-agent systems, delegation all the way down

At the top of the ladder an orchestrating model spawns sub-agents. Perplexity Computer (25 February 2026) is marketed as a "digital worker" that runs sub-agents for hours or longer. Anthropic ships five of its legal agents as headless Managed Agents "cookbooks" (diligence-grid, docket-watcher, launch-radar, reg-monitor, renewal-watcher) with sub-agent delegation. Freshfields has built multi-agent workflows on Google's Vertex platform alongside its firm-wide Claude deployment.

Multi-agent systems are where accountability blurs. OWASP's Top 10 for Agentic Applications (9 December 2025) lists "cascading failures" and "insecure inter-agent communication" as distinct risks, and one study found that a single compromised MCP server in a five-server setup produced a 78% attack success rate. Anthropic's Managed Agents architecture answers this by keeping credentials in a vault so that "Claude-generated code" never sees tokens. Very few law firms need tier 5 today; the ones that do have engineers.

Is it really an agent? A checklist for vendor pitches

Gartner counted only about 130 genuinely agentic vendors among thousands using the word and named the practice "agent washing". Before you sign, ask:

  1. Can it take more than one step without a new prompt from me? If not, it is a chatbot or assistant.
  2. Does it choose which tools to call, or run a sequence someone configured? If the latter, it is a workflow. Say so in the contract and price it accordingly.
  3. Does it show me a plan before it acts, and can I edit the plan? Lexis and OpenAI now do this; anyone can.
  4. Can it pause and ask me mid-task? The MCP specification of 28 July 2026 added multi-round-trip confirmations precisely for this.
  5. Is every tool call logged, and can I export the log? Ask what the log omits; Anthropic's Compliance API, for example, does not capture thinking blocks or MCP server configurations.
  6. Can I stop it, and what happens to half-finished work? A kill switch and a rollback story are the difference between a product and a demo.

A vendor that answers "no" to questions 1 and 2 is selling a workflow, which may be the better buy. A vendor that answers "no" to 3 to 6 is selling an agent you cannot supervise. The full due-diligence questionnaire is in AI vendor due diligence for law firms.

Which tier for which task

Use a chatbot for thinking, drafting from supplied material and explaining. Use an assistant for redlining and correspondence inside Word and Outlook. Use a workflow for anything you do the same way more than ten times a month: intake, extraction, first-pass review, monitoring. Reserve agents for open-ended tasks with a clear finished state and a human at the gate: a diligence grid across an unfamiliar data room, an overnight first response to a counterparty's mark-up, a docket watch that drafts the client note. Reserve multi-agent systems for when a single agent demonstrably runs out of context or time.

The Thomson Reuters Institute found in April 2026 that under 20% of firms use agentic AI widely and that 47% of legal professionals think agents should do legal work at all; the sceptics' line, "I prefer prompting and reviewing results over machines having autonomy to act without concrete review", is a perfectly defensible tier-2 policy. Module 4 of the AI Lab for Lawyers course walks each tier in turn, from a custom GPT to an n8n workflow to a gated agent, so that you can see the supervision burden change with your own eyes before you choose one for the firm. The human-in-the-loop entry is the shortest statement of the principle that runs through all five tiers.

Frequently asked questions

What is the difference between an AI agent and an AI assistant?

An assistant works inside your task and waits for you: Copilot in Word suggests a redline, you accept it. An agent is handed an outcome and takes its own sequence of steps, calling tools such as a document store, a search engine or a drafting function, and only returns when the goal is met or a rule makes it pause. The assistant augments a lawyer's action; the agent replaces a chain of them.

Is agentic AI the same as generative AI?

No. Generative AI is the capability to produce text, images or code from a prompt; every chatbot has it. Agentic AI adds tools, a goal and a loop, so the model can act on the world and decide its own next step. All agents are built on generative models, but most generative AI products are not agents, whatever the brochure says.

Which is safer for a law firm, a chatbot or an agent?

A chatbot with a business-tier data policy is safer because it cannot do anything except write text you then read. An agent can search, edit, send and delete, which is where confidentiality leaks, prompt-injection attacks and unintended actions happen. The trade-off is that agents do far more work per hour, so the answer is gated agents for bounded tasks, not a ban.

How can I tell if a legal AI product is really an agent?

Ask whether it can take several steps without a new prompt, whether it chooses which tools to call, whether it can pause and ask you, whether it shows a plan and a log, and whether you can stop it mid-run. If the honest answer is that it runs a fixed sequence you configured, it is a workflow, which is often exactly what you want.

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