These are the hands — the tools the AI is allowed to use — and every one of them sits behind a single rulebook. Each action the AI wants to take is sorted by fixed rules: safe actions just run, consequential actions are written up and wait for a human's yes, and destructive actions are refused outright — no approval unlocks them, ever.
The AI suggests; a fixed rulebook decides. Which tier an action falls into is determined by code the AI cannot talk its way around, and anything the rulebook doesn't recognize is automatically treated as maximum-risk.
Looking things up and easily-undone housekeeping run immediately: list files, read a spreadsheet, check calendars, summarize a recording. Looking is free; consequences are not.
Anything with consequences — filing a ticket, holding a meeting slot, publishing a report — is drafted in full and set aside. You see a plain-English summary and click approve or decline.
Deleting data, wiping systems, moving money: refused outright. Every command is also scanned for dangerous payloads, so a destructive action can't hide inside an innocent-looking one.
The detail that matters: even after you approve an action, the system re-checks it at the moment it fires — the classification is done fresh, and the conscience gets a second look. Approval is never a blank check.
Around fourteen connected services — mail, calendar, file storage, spreadsheets, slides, code hosting, meetings, note vaults, remote servers and more — all sit behind the same bridge. Adding a new service never creates an exception: the new tool inherits the tiers, the conscience, and the permanent record on day one.
The AI earns independence the way a new hire does: by track record. Every yes and no you give is recorded per type of task. A task type can graduate to running automatically only after you've approved it consistently, over time — and even then with limits, an undo button, and a list you control.
The only thing that widens the AI's independence is your actual record of approvals — not activity metrics. A busy quarter can't accidentally hand over a power nobody signed off on.
Only tasks that can be cleanly reversed are ever eligible to run on their own. Anything irreversible stays behind a human's yes, permanently.
Meeting notes filed, daily briefs archived, a broken build diagnosed with the cause found and a tested fix attached — each arriving as a one-click decision, not another chore.
Fixed rulebookConscience checkProposal queueIndependent second opinionPermanent recordApproval portal