When AI Agents Act, Where Should the Human Check Go?
EXPERT OPINION
By Pouya Shafabakhsh
As legal AI moves from producing answers to completing multi-step work, end-stage review may come too late. Human approval should sit immediately before actions that create material or hard-to-reverse consequences.
Reprinted with permission from the September 24, 2026 edition of the “Legaltech News” © 2026 ALM Global Properties, LLC. All rights reserved. Further duplication without permission is prohibited, contact 877-256-2472 or asset-and-logo-licensing@alm.com.
“Human in the loop” sounds reassuring until the human arrives after the important step.
That timing problem is becoming more relevant as legal AI moves from answering questions to carrying out multi-step work. Agentic systems can search, organize, compare, draft, route, and in some settings update or trigger parts of a workflow. The more steps a system can complete on its own, the less useful a final review becomes if an earlier action was difficult to reverse.
The practical question is not whether a human should remain involved. It is where human approval belongs.
Agentic AI Changes the Timing of Supervision.
With a traditional drafting tool, the control point is obvious. The system produces text. A lawyer reviews it before using it.
An agentic workflow can be different. The system may first gather information, decide which sources matter, classify documents, create a draft, route it to another tool, and prepare a next action. Each step may depend on the one before it.
Recent legal technology releases show that this is no longer theoretical. LexisNexis has described agentic workflows that move from research through drafting and review. DISCO has introduced agentic research for multi-step fact investigation in litigation. Other legal platforms are adding agents that work inside matter spaces and carry context forward.
These tools can reduce repetitive work. They also make one broad instruction—“keep a human in the loop”—too vague to be an operating control.
Not Every Step Needs the Same Gate.
A human approval step before every machine action would defeat much of the point of automation.
An agent should not need a lawyer's permission to open every document it is already authorized to read or to format a table. But the same logic does not necessarily apply when the system is about to change a matter record, send something outside the team, commit a classification that affects downstream review, or trigger another action that is costly to unwind.
The control should follow consequence.
That means separating actions that are easy to reverse from actions that materially change the state of the work.
Put Approval Before Consequence, Not After Completion.
A practical approach is consequence-gated review.
The idea is simple: define a small number of action types that require human approval before the agent crosses the line. Everything else can continue within the approved workflow.
The gate is not based on whether AI is involved. It is based on what the next action would do.
Reversible Internal Steps
Low-consequence actions can usually proceed without interruption when they stay inside approved systems and permissions.
Examples might include searching an authorized document set, organizing results, creating an internal draft, suggesting issues for review, or preparing a comparison that has not yet been relied on.
These steps still need ordinary quality controls, but forcing a manual approval between each one adds friction without adding much protection.
Material Internal Decisions
Some internal actions deserve a checkpoint because they can shape later work even if nothing has left the firm.
Examples may include finalizing a review classification, changing a priority designation, selecting a source set for a material analysis, or moving an item into a workflow that others will treat as completed.
The human check should occur before the decision becomes an input that the rest of the process assumes is correct.
External or Hard-to-Reverse Actions
The strongest gate belongs before actions that create an external effect or are difficult to undo.
That could include sending a communication, transmitting information to a third party, changing an external-facing record, initiating a filing-related process, or taking another action that the firm has defined as requiring human authorization.
The exact list will vary by organization and matter. The point is to decide it before the agent reaches the step.
A Gate Needs More Than a Button.
An approval screen that simply asks “Continue?” is not meaningful supervision.
The reviewer needs enough context to make the decision. The gate should show what the agent is about to do, what information it used, what material exceptions or uncertainty it identified, and what will happen if the reviewer approves.
The system should also record who approved the action and when. That does not mean keeping every internal reasoning trace. It means preserving a simple decision record around the action the organization considered consequential.
The same approach can help with exceptions. If the agent reaches a condition outside its approved scope, the workflow should stop and escalate rather than improvise.
Design the Gate Before Deployment.
The worst time to decide where human approval belongs is after an agent has already been connected to a live workflow.
Before deployment, the team can map the workflow, identify the few steps with real consequence, and place approval immediately before them. That keeps human review focused where judgment matters most.
It also makes automation easier to scale. Lawyers do not have to watch every step, and technology teams do not have to treat every action as equally risky.
Agentic legal AI will continue to become more capable. The answer is not to surround every action with manual friction. It is to be precise about authority.
A human can be “in the loop” and still be too late.
A better control asks a more useful question: before this agent takes the next consequential action, has the right person seen enough to decide whether it should proceed?
Pouya Shafabakhsh is an IAPP Certified AI Governance Professional (AIGP) and Principal Forensic AI Auditor at Radsam Academy of AI. His work focuses on litigation AI governance, auditability, and defensibility. He is not a lawyer. This article is for general informational purposes only and does not constitute legal advice or a statement that any control described here is required in every matter.
Reprinted with permission from the September 24, 2026 edition of the “Legaltech News” © 2026 ALM Global Properties, LLC. All rights reserved. Further duplication without permission is prohibited, contact 877-256-2472 or asset-and-logo-licensing@alm.com.
Author: Pouya Shafabakhsh Co-Founder, CAIO & Principal Forensic AI Auditor, Radsam Academy of AI Sovereign Governance. Independent AI governance and forensic-audit analysis. Educational information only; not legal advice.
Would you be opposed to finding your firm's AI gaps before opposing counsel or insurance companies do? I invite you to assess your firm using our AI Risks Calculator. Safely, securely, privately, and confidentially—with no need to enter any information, not even your email address—you can get a comprehensive AI Risks and Liabilities report in just five minutes.
If you have a sensitive file and want to protect it against AI hallucinations, bias, algorithmic discrimination, and AI-assisted risks through an Air-Gapped Shadow AI Audit before court, or if you are involved in a legal dispute related to AI and need a Judicial Forensic AI Audit, at Radsam, we are willing to assist you whether through our human-based services or by providing our in-house solution, the Sovereign Sanctuary Vault. In these cases, I invite you to fill out our Pre-Qualifying Assessment Form. We will review your information, qualifications, and capacity, and will contact you within two business days.
Comments