“Looks good to me”: why human approval is AI’s real safeguard

Article 14 of the European AI Act requires ‘effective’ human oversight of high-risk AI. At Maestro, nothing is agreed without ‘Looks good to me’: not a regulatory constraint added later, but the product's central mechanism from day one.

Article 14 of the European AI Act requires ‘effective’ human oversight for high-risk AI systems: understanding, monitoring, detecting anomalies and being able to intervene (ia-act.org/comprendre/supervision-humaine-article-14). At Maestro, this requirement is not a checkbox added afterwards: nothing is agreed without your ‘Looks good to me’, starting with the first document produced.

The real risk is not AI making mistakes

The documented risk is not the AI error itself; it is approval becoming a mechanical stamp. On Hacker News, developers say it plainly: ‘you either become a bottleneck (defeats the point) or you rubber-stamp it (creates the debt)’ (trjordan, news.ycombinator.com/item?id=45424824). Another: ‘engineers who are using AI to write and review the code, rubber stamp it, and push it into prod’ (shaftway, news.ycombinator.com/item?id=49301064). Approval that costs nothing to give protects nobody.

Why approving 40 files is impossible, and approving a document easy

Reading the code produced by an agent team line by line does not fit into a day, still less for someone who does not code. Approving a two-page document in everyday French, explaining what the product will do before a line of code exists, fits into ten minutes. This is a deliberate shift: approval happens before production, not after, and concerns the intention rather than its technical execution.

The ‘Looks good to me’ gate: approving an intention, not code

Every stage in Maestro, from the brief and specification to the architecture and task breakdown, stops before proceeding and waits for your explicit agreement. Refusing is not a process failure; it is normal operation: the document returns for revision with your comments until it truly suits you. Nothing moves behind your back while you read something else.

What the law says, in one minute

The AI Act does not require a human to reread every line produced by AI; it requires a human to retain a real ability to understand, monitor and intervene in decisions. That is exactly the difference between stamping a document you have not read and rejecting one you have read and find unsuitable. A report cited by SQ Magazine also notes that the greatest gains from AI agents concern tasks with clear success criteria and recoverable errors (sqmagazine.co.uk/ai-agent-autonomy-statistics), precisely what a well-placed approval gate imposes.

Refusing is an act of direction

A leader rejecting a poorly phrased brief or a specification missing a case is not wasting time: they are directing. It is the same act as a conductor stopping a rehearsal to repeat a bar, not wasted time but the very definition of the role. This role requires no particular technical skill, only the ability to say no while something is wrong and to do so precisely enough that the next revision corrects the right place. We discuss it more broadly in directing an agent team without coding.

What changes in practice

A short document in business language that you can actually judge is worth more than a security report nobody reads. It is the same logic behind publishing defeats in our own benchmarks: bound what can drift at the point where a human can still influence the outcome.