Human in the loop is not a control unless decision rights are defined

Ask an organisation how it controls AI-generated multilingual content and a common answer is that there is a human in the loop. It is offered as a complete response, and it rarely is. A person positioned next to a process is not a control. A control is a defined intervention with stated authority, a trigger, a decision and a record. Without those, the reviewer is a witness.

Multilingual review is distinctive because the reviewer may be the only person in the chain who can read the output. If their role is undefined, the organisation has no reliable account of what was checked.

Authority is the part that is usually missing

Start with the simplest question. If the reviewer decides the output is unacceptable, what happens? In some operations the answer is that they flag it, someone discusses it and the deadline wins. That is not review. It is consultation.

A reviewer with a control function needs an explicit right to stop publication, to require rework or to route the item elsewhere. That authority has to be written down and known to the people whose deadlines it will occasionally disrupt. If nobody has ever seen the reviewer use it, that is worth investigating.

Triggers decide what review is for

Reviewing everything is the fastest way to review nothing properly. When the same level of scrutiny is applied to internal help text and to regulated product information, attention drifts to whatever is in front of the reviewer rather than whatever carries the exposure.

Trigger conditions come from content classification. Which content classes require review, in which markets, at which confidence levels, and which may proceed without it. Once triggers are explicit, the reviewer knows why an item reached them, which changes how they read it.

Competence is specific, not general

Reviewer competence is often treated as a language question. It is usually a domain question. Fluency does not qualify someone to judge whether a translated contractual clause carries the same obligation, whether a claim remains permissible in that market or whether a term matches the approved product vocabulary.

Governance should state what the reviewer is deciding and what qualifies them to decide it. Where the decision spans language and domain, it may need two people, and the framework should say who resolves disagreement between them.

Evidence is what survives the reviewer

A review that leaves no trace cannot be relied on later. The minimum useful record is what was reviewed, against which criteria, by whom, what was changed and what was accepted. It does not need to be elaborate. It needs to exist somewhere other than the reviewer's memory and a chat thread.

Many systems already generate fragments of this in their translation and content systems. The gap is rarely tooling. It is that nobody has stated which fragments count as evidence.

Exceptions and escalation are the real test

Even well-run operations sometimes bypass their own process. A launch moves, a market regulator asks for something quickly, a campaign lands late. The question is whether the bypass is recognised as an exception or simply happens.

Defined governance names who may authorise an exception, on what grounds, for how long and what is recorded when they do. It also names the escalation route when the reviewer and the requesting team cannot agree. Both are unglamorous and both are the difference between a control that holds under pressure and one that only holds when nothing is urgent.

What to check in your own operation

• Can a reviewer stop publication, and has anyone done so in the past year?
• Is it written down which content classes require review and which do not?
• Is it clear what the reviewer is deciding, rather than simply reading?
• Would you be able to reconstruct a review decision six months afterwards?
• Is there a named authoriser for exceptions and a route for disagreement?

If several answers are uncertain, the human in the loop is doing useful work but is not yet functioning as a control. That is a governance gap rather than a staffing problem, and it can be addressed by defining the decision rather than adding reviewers.