How to pick supervision tags for each AI job

AI supervision tags, in this post, means a labeling scheme you invent on a whiteboard for each AI job: read-only, wait-for-approval, or run-on-rules. Nautis does not ship a supervision-tags feature, a tag picker, or those three names as product controls. What Nautis ships is Agent Autonomy (Ask every time or Run without asking per capability), the Approval Center, and AI Activity. Use your labels so teammates share a language before Owners touch the Autonomy dropdowns. This is operating guidance. It is not legal advice.

AI supervision tags as your labeling scheme

Keep the phrase AI supervision tags for the team vocabulary only. Write the three labels on paper. Refuse to talk as if the product UI contains them.

1. Read-only (your label): the job may look up workspace data and draft text that a human still copies by hand. No agent send. No durable write you would regret.
2. Wait-for-approval (your label): the job may prepare a real action, but a human must Approve before it runs.
3. Run-on-rules (your label): the job may execute only where Owners already set Run without asking for that capability, with logs you review afterward.

If someone wants "mostly autonomous except when scary," keep Wait-for-approval until the scary list is written as concrete actions (investor email, invoice changes, ownership edits, social posts, deletes). Your AI supervision tags stay useful because the set is small. A dozen pastel sticker names recreate confusion.

Illustrative facilitation (composite only): a founder runs twelve minutes of tagging, photographs the wall into Notes, and forbids private shadow lists. The labels never appear in Nautis settings; they appear next to the coffee machine.

Say the disclaimer once in the kickoff: "We are inventing AI supervision tags for ourselves. The product still only shows Ask every time and Run without asking." That sentence prevents a new hire from hunting a Tags tab that does not exist.

Score each job before you assign your label

Score four axes from 1 (low) to 3 (high), then apply the rule under the table. This scorecard is yours, not a Nautis screen.

Rules for your AI supervision tags:

• Any axis at 3 on blast radius or reversibility: maximum label Wait-for-approval until a written exception exists.
• Sum 4 to 6: default Read-only.
• Sum 7 to 9 with no axis-3 blast or reverse: Wait-for-approval is normal; Run-on-rules only for narrow Internal tidy jobs after calm logs.
• Sum 10 to 12: keep Wait-for-approval; do not invent a fourth label.

Illustrative scoring for "draft investor update from Briefing": blast 3, reverse 3, freshness 2, owner 1. Sum 9 with two threes. Your label: Wait-for-approval.

Illustrative scoring for "summarize closed tasks into Briefing": blast 1, reverse 1, freshness 2, owner 1. Sum 5. Your label: Read-only.

If two leads score differently, average is not allowed. Debate the divergent axis. Default to Wait-for-approval until you converge.

Connect your labels to Agent Autonomy (the real product control)

Open AI Governance (checked 31 Oct 2026) or Chief of Staff Settings, Agent Autonomy. Each capability is labeled External or Internal and has a dropdown. Help states Ask every time as: "You see the action first (a card in chat or voice, or an entry in Approvals for background agents) and nothing happens until you approve." Help states Run without asking as: "It runs on its own and is logged in AI Activity. You can undo within 10 min where supported." Deleting is locked to Always asks: help says some actions cannot be automated and deleting always asks.

Map without pretending the dropdown shows your stickers:

• Read-only jobs should not need External capabilities on Run without asking.
• Wait-for-approval jobs should keep relevant External capabilities on Ask every time so cards land in Approvals.
• Run-on-rules jobs require an Internal capability already set to Run without asking, a named log owner, and a promotion date on your wall.

Chief of Staff (checked 31 Oct 2026) holds the Approval Center and Agent Autonomy for Owners and Admins. A card from your own chat or from Do it is yours to decide whatever your role. Skip rebuilding the full Approvals walkthrough in this workshop; decide the label, then clear Waiting for approval the same day. The workshop output is the wall. The Autonomy screen remains the control plane.

Do not tell the company a job is Run-on-rules while the capability still says Ask every time. The sticker and the dropdown must match, or the sticker is theater.

Label by job family on your wall

Walk jobs you already run. Keep AI supervision tags on the job name even when models change.

Briefing lookups and Ask Commodore digests: Read-only unless Do it would change something. When Do it would create a change card, treat that path as Wait-for-approval unless autonomy already allows it to run.

Outbound email, investor follow-ups, calendar invites: Wait-for-approval while mail and calendar capabilities stay on Ask every time.

CRM stage changes, task creation, note filing: Wait-for-approval early; promote to Run-on-rules only for narrow Internal actions after AI Activity looks boring.

Overnight summary Routines (checked 31 Oct 2026): Read-only for summaries into Briefing; Wait-for-approval if a routine wants to send or rewrite customer records. Routines still follow approval rules after each run.

Memory policy lines: never Run-on-rules. Humans confirm verified memory with Liftoff habits such as Looks right and Not us on false profiles.

Ownership ledgers, payroll, and bank portals: keep Wait-for-approval or human-only in early workspaces. Conservatism is allowed.

Owners should change at most two capabilities per week while the wall is new. Rapid Autonomy flips make your AI supervision tags look decorative even when the stickers are honest.

Illustrative wall after one workshop

Illustrative six-person wall (composite only). Columns are yours; Autonomy note points at real Nautis controls.

Refuse shadow tags in private docs. If a lead's personal list says Run-on-rules while the wall says Wait-for-approval, the wall wins. AI supervision tags are a company vocabulary layered on top of Autonomy, not a second product.

Review your labels on a fixed cadence

Labels rot when blast radius changes and the sticker stays green. Review monthly, and immediately when you connect a new External channel.

In AI Activity, filter Auto-ran, Approved, Rejected, and Pending. If Auto-ran rows touch customers while the wall still says Wait-for-approval, either the label is wrong or a dropdown flipped. Fix the mismatch the same day.

When you promote Wait-for-approval to Run-on-rules, write the promotion date, the Internal capability name, and undo expectations on the wall. Demote after one bad Auto-ran. Owners should change at most two capabilities per week while the wall is new.

Illustrative onboarding line: "Read our supervision wall before you ask why a card is waiting. Those names are ours. In Nautis, sends wait when the capability is Ask every time; auto actions require Run without asking and show up in AI Activity."

AI supervision tags finish the job when a new teammate can predict what needs their yes tomorrow without believing the stickers are a Nautis menu. The product controls remain Ask every time, Run without asking, Approvals, and AI Activity. Your labels are how humans talk about those controls before anyone opens Settings.

Keep a quarterly archive of wall photos so promotions from Wait-for-approval to Run-on-rules leave a paper trail. If Auto-ran incidents rise after a promotion, demote the label and the dropdown together. Search intent for "AI supervision tags" is thin compared with autonomy and approval queries; treat this article as a labeling playbook that still points at the real Agent Autonomy help, not as documentation of a fictional Tags module.

Photograph disagreements too. When the room cannot pick a label, leave the sticky as Wait-for-approval and schedule a fifteen-minute follow-up with the Approver named on the scorecard. Unlabeled jobs default to Wait-for-approval until scored. That default keeps AI supervision tags from becoming optional decoration while Agents keep moving.