How to match AI autonomy to module risk

Match autonomy to module risk by scoring where a write lands before you flip any dial to Run without asking. A capability that emails a customer is not the same risk as one that tags an internal note, even when both sit on the same Agent Autonomy screen. Owners and Admins set those dials under Chief of Staff, Settings, Agent Autonomy, and every automatic or approved action is audit-logged in AI Activity. This post is a risk-banding habit for modules, not a second tour of every autonomy definition.

Match autonomy to module risk before you flip a dial

Open Agent Autonomy and read each capability as a bet on a module, not as a speed setting. Ask every time means you see the action first as a card in chat or voice, or as an Approvals entry for background agents, and nothing runs until someone decides. Run without asking means it runs on its own, lands in AI Activity, and can be undone within 10 minutes where the product supports undo. Deleting records stays on Always asks and cannot be loosened.

Illustrative starting rule (composite only): "Nothing that leaves the company or moves money language leaves Ask every time in month one."

If you match autonomy to module risk, you decide the band first, then choose the dropdown. Vibes and inbox fatigue are not a band. Keep the first version of the table to one page. If the sheet needs a legend longer than the bands, you are over-specifying. Three bands and a default autonomy column are enough to match autonomy to module risk in practice. Name the Admin who owns the band table. Without an owner, every cofounder resets dials after a busy week and the log fills with surprises. Put the owner next to the table title and review the sheet every Monday for ten minutes.

Score the module, not the model's confidence

Risk comes from blast radius: who sees the change, whether you can reverse it, and whether a wrong write creates legal, cash, or trust damage. The model sounding sure in chat does not lower that radius.

Score each module or surface you care about with three questions:

1. Does a wrong write leave the company (email, message, post, connected app outbound)?
2. Does a wrong write change cash, pricing, invoices, expenses, or equity records people will rely on?
3. Can a non-Admin undo or correct the change in minutes without a forensic search?

Yes to 1 or 2 usually means High. Yes only to quieter internal structure with easy correction means Medium or Low. Write the score in a shared note next to the capability name so the next Admin does not re-litigate from memory.

Illustrative score line (composite only): "Expenses category edits - High - wrong merchant categories confuse books; keep Ask every time."

Connected AI apps that write into the workspace follow the same Agent Autonomy rules as native agents, including AI Copilot and connected-app writes. Those writes are logged too. Do not invent a softer band for a vendor logo. When two modules share one capability label, score the riskier module and keep that band. Do not average High and Low into Medium as a compromise. Averaging hides the outbound write inside a safer sibling.

Three risk bands and what they imply

Use a simple table so founders argue about bands, not about feelings in the moment.

Illustrative banding (composite only): customer reminder email is High; moving a lead between two named stages on a stable board is Medium; adding an internal topic tag is Low.

Bands are operating policy. The product still shows External or Internal on each capability. External means the action touches the outside world; Internal stays inside Nautis. Treat External as at least High until you have evidence otherwise. Internal is not automatically Low. For Expenses and other finance records, prefer Ask every time while categories and merchants are still settling. A wrong auto-filed receipt is quieter than a wrong customer email, yet it still poisons exports your bookkeeper trusts.

Let External and Internal set the floor

Help text labels capabilities External or Internal. That label is the floor for your band, not the ceiling. An Internal write that changes invoice fields can still deserve High because cash reports depend on it.

Keep External on Ask every time while Reject reasons still mention wrong recipient, wrong tone, or wrong board. Widening External to Run without asking before the Approvals queue is calm just moves mistakes into Auto-ran rows you read after the fact.

Illustrative floor (composite only): Internal task creation moves to Run without asking after fourteen days with no Reject on that capability; External email stays Ask every time.

Always asks for deletes remains outside the banding exercise. Do not plan a project to automate deletes. The product locks deletes on Always asks. If undo is available for an Auto-ran Internal action, still treat the band seriously. A ten-minute undo window is not a reason to put customer outbound on Run without asking. Undo support varies by action; High-band policy should not depend on it.

When you change a dropdown, new actions follow the new setting. Past AI Activity rows stay as history. Tell the other Admin in the same note where you keep band scores. Record the date you last changed a High-band dial. Stale High settings that silently moved to Run without asking during a crunch are a common audit surprise.

Tighten one action without loosening a whole capability

Some capabilities include Per-action overrides so one action can differ from its siblings. Use Filter actions to find the risky step instead of flipping the parent capability to Run without asking for everything.

Illustrative override (composite only): a capability that both drafts and sends keeps send on Ask every time while a low-risk internal draft step runs without asking after review.

Overrides are how you match autonomy to module risk inside a bundled tool. Without them, teams either freeze every action or auto-run the dangerous sibling by accident.

After you set an override, approve or reject from Approvals as usual. The activity log records what ran. Owners and Admins are the only roles who change autonomy or decide those cards.

Clear Approvals the same day you accept more risk

Raising a band without staffing Approvals creates Pending pileups and Expired rows. Open Chief of Staff, Approvals, and treat Pending as a same-day queue whenever High-band capabilities are active.

Illustrative queue rule (composite only): "High-band Pending cleared before lunch; Medium before end of day; nothing Critical left overnight."

After you Approve or Reject, skim AI Activity filters for Auto-ran, Approved, and Rejected. Auto-ran should match what you intentionally widened. Rejected should teach better prompts or a tighter band, not a shrug. Write Reject reasons that name the module risk, not only the prompt. "Wrong customer email on High-band outbound" teaches the next Admin faster than "didn't like it." Keep those phrases short enough to scan in Approvals history.

If connected apps add write volume, re-score those modules the week you connect them. Connect Claude, ChatGPT and other AI tools explains how connected writes wait on autonomy and show up in the log. Do not assume a new app inherits a Low band because the pilot felt helpful in chat.

Routines that propose outward writes also feed Approvals when asks-first applies. Pausing a noisy routine often shrinks risk faster than loosening autonomy so the queue "keeps up."

Recheck bands when the company changes shape

Revisit the band table when you add a finance workflow, open a customer channel, hire someone with Admin, or connect a new AI app. Module risk moves when the blast radius moves.

Illustrative trigger (composite only): first time agents may draft invoice reminders, Expenses and Invoices related capabilities return to High and Ask every time for two weeks of supervised Approvals.

Keep the written band table next to your autonomy habits. For reading the log itself, use your AI Activity review routine; for staffing the queue, keep Approvals owned. This post stays on matching the dial to the module's risk band.

Chief of Staff holds Approvals, Agent Autonomy, and the path into AI Activity. AI Governance is the same log with shortcuts to Approvals and Autonomy. Use those pages for control surfaces; use the band table so every Admin makes the same call when a new capability appears. Ship the band table before you connect a second AI app. Adding tools while every capability is still on improvisation multiplies Pending cards without a shared definition of High.