There is a strain of thinking in AI product design that treats full autonomy as the goal. The ideal AI, in this view, would handle everything without human intervention. The user sets a strategy, the AI executes indefinitely, and the results just appear.
We think this is wrong. Not as an engineering problem - it is technically achievable. It is wrong as a product design decision. For marketing, where every piece of communication carries the weight of a brand's reputation, full autonomy is not a feature. It is a liability.
The approval architecture
Cleo can draft an email to ten thousand subscribers. It can generate a week of social content. It can adjust advertising spend across campaigns. It can publish blog posts. These are high-stakes actions with real business consequences. Every one of them goes through a review queue where the human sees exactly what will happen and explicitly approves.
This is not a safety net. It is the core interaction model. The AI handles the creative and analytical work that takes hours when done manually. The human applies judgment, makes final edits, and decides what goes live. Each party does what they do best.
Why speed is not the metric
The criticism of human-in-the-loop is always about speed. "If I have to approve everything, what is the point of the AI?" The point is that the AI compressed four hours of work into four minutes of review. The bottleneck was never the approval step - it was the creation, analysis, and optimisation that preceded it.
A marketing manager who spends thirty seconds reviewing and approving an AI-drafted email campaign has gained back the two hours they would have spent writing it. The approval step is not friction. It is the moment where they exercise the judgment they were hired for.
Trust through control
Human-in-the-loop builds trust faster than any transparency mechanism or confidence score. When users know they have the final say on everything that goes live, they engage more freely with the AI. They experiment more. They push the system further. They use it for higher-stakes tasks. Paradoxically, giving the human a veto makes them more willing to delegate.
We have seen this pattern repeatedly. Users who start cautiously, reviewing every detail, gradually shift to quick-scanning and approving as trust builds. The review queue adapts to this - high-stakes actions get prominent treatment, routine approvals are streamlined. The human stays in the loop, but the loop gets faster.
The design of approval
How you present the approval matters as much as the approval itself. We show the complete output, not a summary. We show where it will be published. We show who will receive it. We make the approve and reject actions clear and prominent. We allow inline editing so the user can adjust without starting over.
The approval experience is a first-class product surface, not an afterthought checkbox.
- Cleo's Team