What supervised autonomy costs you in review time, and what it saves — an honest accounting

Supervised autonomy — an AI agent drafts, a human approves, then it executes — is not free. Every approval step is a real interruption in someone's day, a context switch, a moment where a technician has to stop what they're doing and evaluate something instead of just letting it run. It's worth being honest about that cost instead of pretending review is weightless, because the entire argument for keeping a human in the loop only holds up if the thing it buys is worth more than the thing it costs. Sometimes it isn't, and the design decision is figuring out which is which, action by action.

What the review step actually costs

• Technician attention pulled away from other work every time a draft needs a decision, even a fast one
• Latency between 'the fix is ready' and 'the fix runs,' which matters more on some problems than others — a printer offline overnight costs nothing to wait an hour on, a service outage does
• The cognitive cost of reviewing well enough to actually catch a bad draft, not just rubber-stamping because the last twenty drafts were fine
• Queue buildup if approvals pile up faster than one on-call technician can clear them, which is a staffing and policy problem more than a technology one

What it buys back

• A second set of eyes on anything destructive or irreversible before it executes, catching the draft that's technically plausible but wrong for this specific client's environment
• A defensible audit trail where the AI's draft and the human's approval are two separate, linked events — not one blended entry that obscures who actually decided
• The ability to catch a systematic drafting error early, across a handful of reviewed cases, before it's had a chance to execute unsupervised across hundreds
• Client trust that survives an audit, a compliance review, or an incident postmortem, because the record shows a human decision at every consequential step, not just an AI's self-report

The honest accounting isn't 'autonomy good' or 'review good' — it's that review is a cost you pay to buy a specific thing, catching bad drafts before they execute, and that trade is worth making on some actions and not worth making on others.

This is exactly why policy in Nexus is graduated rather than all-or-nothing: a tenant operating in Advisory mode pays the full review cost on everything, a tenant in Assisted mode has let specific low-risk, non-destructive capability classes graduate toward less friction, and consequential actions don't graduate regardless of how clean the track record looks, because the cost of one bad irreversible action doesn't average out against a hundred good routine ones. The honest version of 'how much review time does supervised autonomy cost' is that it depends entirely on how narrowly you've scoped what gets to skip review — and that scoping is a policy decision each tenant makes deliberately, not a default we set once and stop thinking about.