When people picture AI review, they usually picture a binary: a reviewer looks at an answer and either blesses it or bins it. Real review is richer than that, because real judgments are rarely all-or-nothing. An answer can be mostly right with one weak claim. It can be sound but need context before a client sees it. It can be salvageable with one more pass. A review system that only offers approve or reject flattens all of that into two buttons, and the nuance gets lost exactly where it matters.
This is why sign-off in Qonera has three real outcomes, not two. A named reviewer can approve the work, annotate it, or send it back for revision, and each of those is a different professional judgment with a different consequence. Together they make the review step work the way senior review has always worked in good firms, rather than the way software defaults tend to flatten it.
Approval is the strongest statement in the workflow: a named supervisor has examined the work, weighed the flagged claims against the evidence, and decided it is fit to rely on. The approval is recorded with the reviewer’s identity and a timestamp, which is what separates it from the ambient sense that someone probably looked. Months later, the firm does not have to reconstruct who checked this. The record says so.
That named-ness is doing real work. Sign-off that could be anyone’s costs nothing to give; sign-off that carries your name is given differently. It changes the reviewer’s attention, and it gives the client and the firm a specific accountable person rather than a process abstraction.
Annotation covers the large middle ground where the work is right but the reviewer knows something the answer does not say. This figure is correct but preliminary. This clause reads adverse but is standard for the sector. This conclusion holds unless the deal structure changes. Annotations let the reviewer attach that judgment to the work itself, where the next reader will actually see it, instead of losing it in a chat thread or a hallway remark.
This is often where the most valuable senior input lives. The model produced a defensible answer; the reviewer added the context that makes it usable. The record keeps both, which means the firm’s expertise compounds inside the workflow instead of evaporating around it.
Sending work back for revision is the third outcome, and its value is that it is directional rather than terminal. The reviewer is not discarding the work; they are saying what is wrong and routing it for another pass: tighten this claim, check this source, answer the question that was actually asked. The loop is recorded too, which matters more than teams expect. Work that was sent back and then approved carries a different history than work approved on sight, and the trail shows the scrutiny happened.
All three outcomes land in the same tamper evident audit trail, alongside the evidence and the flagged claims the reviewer was looking at when they decided. That is the difference between a review step and a review culture: the workflow captures not just what was decided but the deciding itself.
The same structure is what Article 14 of the EU AI Act calls meaningful human oversight: a person with the authority to interpret, override, or reject the system’s output, not a rubber stamp at the end of a pipeline. Approve, annotate, or send back is that authority made concrete. The reviewer is not decorating the AI’s decision. They are making the firm’s.
This article is for general information only and does not provide legal advice. Organisations should consult qualified legal counsel about how Article 14 and the EU AI Act apply to their specific systems, workflows, and obligations.
Multi-model stress testing, Conflict Heatmap, tamper-evident audit trail, and structured sign-off, built for teams who need defensible AI output.