AI Agents for Lawyers

Glossary

Human-in-the-Loop — Definition for Lawyers

Human-in-the-loop means an AI system pauses for a person to review or approve before it acts. Where the gate belongs in legal workflows and why regulators ask for it.

By Dr. Niklas SchmidtUpdated 15 November 20263 min read
On this page
  1. Where the gate belongs
  2. Why regulators and bars keep asking for it
  3. Designing an approval that is not a rubber stamp
  4. Where this leaves you

Human-in-the-loop describes an AI system designed to stop at defined points and wait for a person to review, correct or approve before it continues. In legal work it is the control that separates "the agent drafted it" from "the agent sent it". The model reads, extracts and proposes; a lawyer decides whether anything leaves the firm, changes a record or commits a client.

Where the gate belongs

The principle is simple and the placement is everything. A checkpoint at the end of a long automated run, where a lawyer skims a summary of what has already happened, is not oversight. A checkpoint immediately before an irreversible step is. The rule of thumb used throughout this site: any action that sends, files, signs, pays or deletes waits for a person.

Step in the workflowGate?Reason
Reading and classifying incoming documentsUsually noErrors are internal and correctable
Extracting dates, parties or clauses into a tableSpot-checkWrong data propagates; sample and verify against source
Drafting a reply, redline or memoReview before useNothing leaves the firm until a lawyer has read it
Sending an email or letter to a client or counterpartyYes, alwaysExternal, irreversible, may waive privilege or bind the client
Entering a deadline in the calendar or docketing systemYesA wrong date can cause a missed limitation period
Filing with a court or registryYesProfessional responsibility rests with the signing lawyer
Deleting, moving or overwriting filesYesData loss and preservation duties

Workflow agents built in tools such as n8n make this explicit with an approval node; chat-based agents increasingly ship a plan-then-approve step. Either way, the question to ask of any tool is the same: what can it do without asking you?

Why regulators and bars keep asking for it

Professional rules do not use the phrase, but they require the substance. A lawyer remains responsible for work product whoever, or whatever, drafted it, and must supervise the tools and people who contribute. The analysis in ABA Formal Opinion 512 rests on duties of competence, confidentiality and supervision that apply regardless of the technology; European bar guidance, covered in Austrian and German bar guidance, takes the same line on final responsibility.

The EU AI Act goes further for high-risk systems. Article 14 requires that such systems be designed so that natural persons can effectively oversee them, including understanding their limits, interpreting their output and being able to override or stop them. Most legal-practice tools are not high-risk under the Act; the EU AI Act guide explains the classification. The concept is still the right design standard.

Designing an approval that is not a rubber stamp

An approval click is only as good as what the reviewer sees. A checklist for any gate:

For each action awaiting approval, the reviewer sees:
[ ] What the system proposes to do, in one sentence
[ ] The exact text or data it will send, file or change
[ ] The source passage it relied on, with a pinpoint reference
[ ] Who will receive it or what record will change
[ ] A reject option that is as easy as approve, with a reason field
[ ] A log entry recording who approved, when, and what they saw

Keep the number of gates small. A lawyer asked to approve every step soon approves without reading, which is worse than an honest gap because it looks like control.

Where this leaves you

For each AI workflow in the firm, write down every action it can take and mark which ones happen without a person. Any external or irreversible action on that list is a design flaw to fix before the next run. Module 6 of the AI Lab for Lawyers course covers verification and keeping human responsibility when agents do the routine work.

Frequently asked questions

What is the difference between human-in-the-loop and human-on-the-loop?

In a human-in-the-loop design the system cannot proceed past a checkpoint until a person approves. In a human-on-the-loop design the system acts on its own while a person monitors and can intervene or stop it. For legal work, anything that leaves the firm or changes a record should be in-the-loop; on-the-loop is acceptable only for low-stakes internal steps such as sorting or tagging.

Does a human-in-the-loop step make AI output safe to rely on?

Only if the review is real. A partner clicking approve on fifty agent actions in a row is a rubber stamp, not oversight. The gate works when it is placed at a small number of decisive points, when the reviewer sees the source material behind each proposed action, and when rejecting is as easy as approving. Volume is the enemy of genuine review.

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