AI Agents for Lawyers

Glossary

Workflow Agent — Definition for Lawyers

A workflow agent is an AI agent that runs a fixed sequence of legal steps with approval gates. Definition, Harvey and Lexis examples, and five design rules.

By Dr. Niklas SchmidtUpdated 13 November 20264 min read
On this page
  1. Workflow agent versus free-form agent
  2. Who sells them, as of September 2026
  3. Five design rules
  4. Where this leaves you

A workflow agent is an AI agent whose steps are fixed in advance: intake this document, extract these terms, compare them against this playbook, draft this output, stop for approval. The model does the reading and writing inside each step; the sequence, the tools and the checkpoints are decided by whoever built it. That predictability is why most legal AI agents sold today are workflow agents, not free-roaming ones.

Workflow agent versus free-form agent

Anthropic drew the line in December 2024. 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 ladder of 11 March 2025 puts the workflow above the agent: an agent can "plan, adapt, and meaningfully interact with humans", while a workflow is "one or more agents that collectively produce a meaningful, specific work product" (Harvey). The useful test is who decides the next step.

Workflow agentFree-form agent
Who decides the stepsThe builder, in advanceThe model, at run time
Typical legal taskNDA review against a playbook; lease summaries; a disclosure schedule"Find out whether this target has undisclosed litigation"
Where the human sitsAt gates written into the designWherever the model chooses to ask
Failure modeWrong in the same way every time, so testableWrong in a new way, so surprising
FitsHigh-volume, repeatable workInvestigation and open-ended research

Gartner's rule from 25 June 2025 is worth quoting back at a vendor: "AI agents when decisions are needed, automation for routine workflows and assistants for simple retrieval". The same release predicts over 40% of agentic AI projects cancelled by end-2027 (Gartner). The wider taxonomy is in what are AI agents and agentic AI.

Who sells them, as of September 2026

Harvey ships 500+ pre-built agents and opened Agent Builder to early access on 5 May 2026; customers had built 25,000+ of their own by March 2026, and the agents "flag moments requiring user confirmation before proceeding" (Harvey, 9 March 2026). A&O Shearman sells four Harvey-built agents to clients on subscription, including loan review and fund formation (A&O Shearman, 6 April 2025). See the Harvey review for pricing.

Lexis+ with Protégé launched in the US on 24 February 2026 with 300+ configurable workflows and a no-code builder (LawSites). Its Legal Intelligence Engine (24 August 2026) "presents an executable plan for user approval" before running: the workflow principle applied to a free-form request (Artificial Lawyer).

Gemini Enterprise Workflow Builder reached general availability on 3 September 2026 (Google Cloud release notes). Anthropic's open-source Claude for Legal repository does the same in code: 13 scheduled agents, such as renewal-watcher, run fixed routes with hard gates (GitHub).

OpenAI's Agent Builder, launched 6 October 2025, closes on 30 November 2026; Relay.app shut to paying customers on 14 September 2026. A workflow outlives its platform only if you can export it.

Five design rules

  1. Fix the route. One job, one input, one output per step. If a trainee could not follow the steps as a checklist, the task is not ready.
  2. Gate anything that leaves the building. Anthropic's litigation plugin will not send a demand letter until a human confirms the settlement-privilege framing; Lexis Workrooms require dual approvals. Put the human in the loop at the send step, not at the end of a report nobody reads.
  3. Make every output cite its source. Anthropic's corporate plugin promises "every cell cited" in its diligence tables and tags unchecked citations [verify]. Copy that standard.
  4. Test on a fixed set before scaling. Harvey's LAB benchmark scores the best frontier model at 13.3% for passing every criterion on a long legal task (18 September 2026), and in Harvey's first run drafting without review cost points while verify-and-revise behaviour gained them (Harvey, 26 May 2026). Twenty real matters scored by a lawyer beat any demo.
  5. Keep an exit. Export the workflow definition and run history as a file before building anything else on a vendor's canvas.

Module 4 of the AI Lab for Lawyers course builds workflow agents of this kind in n8n, Make and Zapier, gate by gate.

Where this leaves you

If the work is repeatable and high volume, a workflow agent is the right shape and the cheapest to supervise. Write the route as a numbered checklist, pick the gate, and build that step first. The contract review playbook walks through one from intake to redline.

Frequently asked questions

What is the difference between a workflow agent and an AI agent?

Both use a language model with tools. In a workflow agent the sequence of steps, the tools used at each step and the points where a human must approve are decided in advance by the builder. In a free-form agent the model decides the next step itself at run time. Workflow agents are predictable and testable; free-form agents handle open-ended tasks but are harder to supervise.

Are Harvey Workflow Agents real agents or just automation?

They sit between the two. Harvey describes a workflow as one or more agents that together produce a specific work product, with the model doing the reading and drafting inside each step and confirmation prompts at defined moments. Under Gartner's June 2025 rule, that is automation for routine work with an agent inside, which is exactly what most legal tasks need.

Can a small firm build a workflow agent without a developer?

Yes, for bounded tasks. No-code builders such as n8n, Make, Zapier and Lexis Protégé's workflow builder let a lawyer chain intake, extraction, comparison and drafting steps with an approval stop. The hard part is not the software but writing the route as a checklist and choosing a test set of real documents to score the output against.

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