AI Agents for Lawyers

Glossary

Agentic AI (Definition for Lawyers)

Agentic AI means software where a language model plans and carries out multi-step work with a degree of autonomy. A five-level autonomy ladder, legal examples and cautions.

By Dr. Niklas SchmidtUpdated 13 November 20264 min read
On this page
  1. The autonomy ladder
  2. Why every vendor says it
  3. Three cautions specific to lawyers
  4. How to read the word in a pitch

Agentic AI is the umbrella term for software where a language model plans and carries out multi-step work with tools and a degree of autonomy, instead of answering one prompt. It is an adjective, not a product category, and it covers everything from a Word assistant that pauses for approval to a scheduled agent that works overnight. The useful question is how much autonomy, and where the gates sit.

The autonomy ladder

"Agentic" describes a spectrum. Placing a product on it tells you more than the label does.

LevelWhat the system doesWho approvesLegal examples (as of September 2026)
0. ChatAnswers one prompt and waitsYou, every turnClaude, ChatGPT or Gemini in chat mode
1. AssistedProposes tracked changes in your document; you accept themYou, per changeMicrosoft Legal Agent for Word (preview since 30 April 2026); Spellbook
2. Supervised agentPlans and runs a multi-step task, pausing at defined gatesYou, at each gateLexis Legal Intelligence Engine "presents an executable plan for user approval" (24 August 2026); Harvey agents "flag moments requiring user confirmation before proceeding"
3. Scheduled or headlessRuns unattended on a trigger and queues results for reviewYou, after the factLegora aOS: overnight counterparty redlines "reviewed, flagged and responded to by the time lawyers arrive" (7 May 2026); Claude for Legal's 13 scheduled agents
4. Open-ended autonomousPursues a goal for hours with no fixed gatesNobody, until it stopsPerplexity Computer, "capable of running for hours or even months" (25 February 2026); no legal product belongs here

Sources for the quotations: Artificial Lawyer on the Lexis engine, Harvey's Agent Builder announcement, Artificial Lawyer on Legora aOS and Perplexity.

Harvey's own ladder is worth knowing because other vendors borrow it: a model gives "a single response"; a model system chains preset calls; an agent can "plan, adapt, and meaningfully interact with humans"; a workflow is "one or more agents that collectively produce a meaningful, specific work product" (Harvey, 11 March 2025). Harvey's Workflow Agents and its Agent Builder (early access 5 May 2026, alongside 500+ pre-built agents) sit at level 2. The single building block is defined at AI agent; the chained kind is a workflow agent; the Lexis engine is examined in the Lexis+ with Protégé review.

Why every vendor says it

Because the money does. Harvey, whose March 2026 round was announced as capital to scale agents across law firms and enterprises, raised a further $550m at a $15.5bn valuation on 9 September 2026. Legora calls its platform an "agentic operating system". LexisNexis CTO Greg Dickason describes the Legal Intelligence Engine as "an agentic harness". Relativity renamed RelativityOne to Relativity aiR on 16 September 2026 with agentic AI included in platform pricing. Every one of these is a real product. The word still needs checking.

Gartner's Anushree Verma put it plainly: "Many use cases positioned as agentic today don't require agentic implementations." Gartner counted roughly 130 vendors with genuinely agentic products among thousands claiming the term, called the rest "agent washing", and predicted that over 40% of agentic AI projects will be cancelled by the end of 2027 (Gartner, 25 June 2025).

Adoption trails the marketing. The Thomson Reuters Institute found fewer than 20% of law firms and departments using agentic AI widely, about half planning to, and only 47% of legal professionals saying agents should be used for legal work at all, against 22% who said they should not (Thomson Reuters Institute, 9 April 2026). Deloitte's legal-department survey (April–May 2026) reported 61% piloting agentic AI and 10% with it fully embedded. One TR respondent's view deserves weight: "Agentic AI removes oversight too far. I prefer prompting and reviewing results over machines having autonomy to act without concrete review."

Three cautions specific to lawyers

  1. Liability follows autonomy. Clifford Chance's briefing "Agentic AI: the liability gap your contracts may not cover" (9 February 2026) flags authority to bind, hallucinated commitments, indemnities and insurance as the open points (Clifford Chance). An agent that emails a counterparty has, in practice, made a representation on your letterhead.
  2. Regulators now use the word. The State Bar of California's Board of Trustees approved updated Practical Guidance covering generative and agentic AI on 14 May 2026, and the CCBE's Technical Guide on AI tools and models (adopted 27 March 2026, published 23 April 2026) addresses agentic systems for European lawyers. "Agentic" is no longer a term the professional rules ignore.
  3. Access is the risk, not intelligence. Super.com's general counsel Michele Lee: "If you give it access to start deleting files, that's concerning" (Law.com, 19 May 2026). Level 3 systems need a person at the end of the queue; see human-in-the-loop.

How to read the word in a pitch

Ask the vendor to place the product on the ladder above, name every gate and show the log. If the answer is "level 2 with gates at send, sign and delete", proceed to a pilot. If it is "fully autonomous", ask what happens when it is wrong and who pays. The AI Lab for Lawyers course spends its fourth module on where those gates go, using n8n, Make and Zapier rather than a vendor demo. The buying framework is in AI vendor due diligence for law firms.

Frequently asked questions

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

An AI agent is a thing: one model with tools and a goal that decides its own next step. Agentic AI is the adjective for systems built that way, covering single agents, chains of agents and platforms such as Legora's aOS or Lexis's Legal Intelligence Engine. Vendors use the adjective because it can describe almost any product with a loop in it.

Is agentic AI safe to use in a law firm?

At levels 1 to 3 of the autonomy ladder, with gates at send, sign and delete and a human at the end of every queue, yes, and the leading products are built that way. Open-ended autonomous systems that pursue a goal for hours without fixed gates have no place near client files as of September 2026.

Why do vendors call everything agentic?

Because that is where investment and pricing have gone: Harvey raised $550m at a $15.5bn valuation in September 2026 and Relativity now bundles agentic AI into platform pricing. Gartner calls the rebranding of chatbots and scripted automation as agents agent washing and counted only about 130 vendors with genuinely agentic products.

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