Direct answer
What makes human review meaningful in a legal AI workflow?
Meaningful review needs a review standard, a named reviewer, budgeted time and a retained record. It must also let the reviewer reject or overturn the output.
Start with the decision, not the disclaimer
“Human in the loop” is often written as a safety statement. It becomes an operating control only when the firm defines the decision the person is making. Are they checking extracted facts, verifying quotations, confirming that every required issue was considered, or approving a recommendation?
The Solicitors Regulation Authority's warning notice on misuse of AI focuses on inaccurate information and client confidentiality. Those risks cannot be answered by a generic instruction to check the output.
Write the review standard before the test
For each material output, name the source the reviewer must open and the checks they perform. A contract-review workflow might require every quotation to match the source, each flagged clause to be traced to the document and each required playbook issue to receive either a finding or an explicit “not found”.
Set the failure route at the same time. Missing evidence, conflicting text or an unsupported conclusion should stop the workflow or move it to a named escalation path.
Name a reviewer who can challenge the output
The reviewer needs enough subject knowledge, authority and context to disagree with the system. The ICO's guidance on individual rights in AI systems explains why nominal review can fail when a person cannot meaningfully assess or overturn an output.
Assigning review to “the team” is not enough. The operating record should show who reviewed this output and who owns the standard when the tool, matter type or risk changes.
Why should checking time count as workflow time?
Review should not be an unmeasured overhead. Record how long qualified checking takes, how many corrections are made and how often work escalates. A faster first draft can still increase total effort if it creates a longer or more demanding verification step.
Retain enough evidence to learn
Keep the review date, reviewer, checks completed, corrections and final disposition. The purpose is not paperwork for its own sake. Repeated corrections show whether the problem sits in the source data, extraction, instructions, model behaviour or review design.
Where personal data is involved, the firm's assessment and recordkeeping should also follow its data-protection duties and professional advice. The workflow record does not replace a DPIA, legal advice or regulatory accountability.
How can a firm check whether it has designed human review?
If the firm cannot state the review standard, assign a capable reviewer, budget the time and show what happens on failure, it has not designed human review yet. It has added a person to the diagram.
For the wider context, read our legal AI workflow answers. Explore Margo by Margo Legal, currently in development, or use the Workflow Value Workshop to examine one repeated workflow.