The most dangerous word in an AI governance process is low. Low can mean two completely different things. It can mean the system is boring after someone looked at it. It can also mean nobody found the dangerous part because nobody asked for enough evidence.
Proportionality is necessary. A governance process that sends every trivial tool through the same review as a high-impact automated decision will collapse under its own seriousness. The cost is not just time. When everything is escalated, escalation stops meaning anything. The scarce resource in governance is not forms. It is serious attention.
The failure mode appears when proportionality becomes a discount applied before knowledge exists. A team describes a tool as assistive, internal, or human-reviewed. The form has a risk box. Nothing obviously bad appears in the first pass. The easy move is to treat missing concerns as absence of concerns. That is how low risk becomes a story told at intake rather than a property discovered by review.
Better proportionality works in the opposite direction. It does not start by asking whether anyone can find a reason to escalate. It asks what would have to be true for this system to avoid the heavier review. The lighter path has to be earned. The system has to be bounded. The data it sees has to be understood. The actions it can take have to be reversible, supervised, or blocked at the point of commitment. The owner has to be named. The review point has to sit before harm, not after the log file explains it.
Uncertainty is not guilt, but it needs a default. The default should be stronger oversight. Not because every unknown hides a disaster. Because the system has not earned the right to be treated as routine. A blank field is not evidence. “Human in the loop” is not evidence unless the human has the right information at the right time and can stop the action before it matters. A comparison to an earlier reviewed case is not evidence unless the important parts are actually the same.
This changes what the governance work is. The hard part is not inventing one more tiering model. The hard part is deciding what counts as proof. What claims can a team make about a system? What has to sit behind those claims? Which changes break reliance on an earlier review? Which gaps can be tolerated temporarily, and which ones force a more serious conversation?
This makes proportionality a safety mechanism rather than a relief valve. The point is not to reduce the number of reviews. The point is to stop spending the same review effort on unlike things. Attention saved on cases that have proved they are ordinary can be spent on cases that are unusual, unclear, or consequential.
The sentence I would build around is simple: the absence of a known escalation trigger is not evidence of low risk. A governance process that forgets this will drift toward optimism. The work is to make the optimistic path prove itself.