The difference between the AI suggesting and the AI doing

A product lead at a services company told me about a change that happened without a decision behind it. Their assistant started out proposing actions. It would draft the email, suggest the record update, recommend the refund, and a person would review and approve before anything happened. Over a few iterations, to make the feature feel faster and more capable, the approval step got trimmed for more and more action types, until the assistant was taking many of them directly. Nobody had held a meeting to decide that the AI should act without review. It drifted there, one convenience at a time, and the last safe checkpoint quietly disappeared.

There is a real line between an AI that suggests and an AI that acts, and it is one of the most important lines in the whole design. On one side, the model proposes and a human commits, so a mistake stops at a person before it reaches the world. On the other side, the model commits directly, so a mistake reaches the world before anyone sees it. Moving from the first to the second is a genuine increase in risk, and it deserves to be a decision, not a drift.

The approval step is where mistakes stop

When a human approves an AI’s proposed action, that step does quiet, essential work. It catches the wrong refund, the misdirected email, the update to the wrong record. The model can be confidently wrong, and the approval is the place where confident wrongness meets a person who can say no before it costs anything. Remove that step and every confident mistake becomes a real action with real consequences.

The trouble is that the approval step is also friction, and friction is exactly what teams optimize away when they want a feature to feel powerful. Each removal seems small. Approving low-risk actions feels unnecessary, so they get automated. Then slightly higher-risk actions, because the low-risk ones went fine. The boundary creeps upward without anyone deciding where it should stop, until actions that genuinely needed a human are happening without one.

What makes this dangerous is that the drift is invisible in the good case. As long as the model is right, removing approvals just makes the feature snappier. The cost only appears on the wrong action, which is precisely the case the approval step existed to catch.

Decide the line on purpose, by consequence

The fix is to place the human gate deliberately, based on how bad it is to get an action wrong, and to make crossing the line a decision rather than a drift.

What helps:

Gate by consequence, not by convenience. Actions that are hard to undo, or that affect money, customers, or data, belong behind human approval regardless of how confident the model is.
Keep low-cost actions fast, high-cost actions reviewed. Automating the reversible, low-stakes actions is reasonable. Automating the irreversible ones is where you need a person.
Make automation of an action an explicit choice. Moving an action type from proposed to automatic should be a decision someone makes and records, not something that erodes over iterations.
Preserve a way to review after the fact. Even for automated actions, keep a record and a path to catch and reverse the ones that were wrong.

The teams that get this right hold the line where the consequences justify it. They let the AI move fast on the cheap, reversible actions and keep a human in front of the expensive, permanent ones. The gate is placed on purpose, and it stays where they put it.

How we approach it at Density Labs

In the AI Opportunity Assessment, our fixed two-week, $2,500 engagement, we map which actions the AI proposes and which it takes directly, and we weigh each against its consequences. When high-cost actions have drifted into full automation, we flag it and design the gate before it ships. The approval step is the cheapest safety control an AI feature has, and the easiest to erode without noticing.

Suggesting an action and taking it are different levels of risk. Decide where the human gate belongs by how bad the mistake would be, and make crossing that line a choice.