How to run an Approval Center for AI agents

An Approval Center for AI agents is the queue where background writes wait until a person decides. Nautis help uses that name; in the product you open Chief of Staff, then Approvals. Owners and Admins Approve, Reject, or Cancel request. This post is about running that queue as an operating habit: who owns it, which response clocks you choose, and how to decide without turning every card into a debate. It is not a second tour of Agent Autonomy dials. Operating guidance only, not legal advice.

Make Pending a staffed work queue

Pending should not age into surprise. Someone named opens Approvals on a cadence you choose, reads Show details when the stakes are real, decides, and then spot-checks AI Activity so Approved and Auto-ran rows match expectations.

Illustrative posture (composite only): one Owner clears Approvals twice daily; a deputy covers travel days; nobody batch-clicks Approve during standup.

Batch Approves without Show details are how wrong invoices, wrong stage moves, and wrong emails ship. If the queue is long, pause sources first, then clear. Throughput without reading is not a win.

If Approvals is only opened when something already went wrong, you are doing incident response. Put Approvals next to the morning briefing habit, not after inbox zero. The badge is a work queue. Clear it while context is fresh so Reject reasons still make sense.

Staff the queue before you widen autonomy

Write three names on a Note:

Those times are your operating choice, not a Nautis default. Help does not prescribe how many clears per day. Pick clocks your team can keep, write them down, and revise them when the queue changes.

Only Owners and Admins can decide approvals and change Agent Autonomy. If your primary clearer lacks that role, fix permissions before you depend on the cadence.

Illustrative staffing (composite only): COO as primary, CEO as escalation, no third "everyone watches the badge." Badge blindness is how Expired rows appear under History.

Publish the staffing Note where the team already looks for Chief of Staff links. New admins who gain Approvals rights must read your SLA board before their first Approve.

Keep Autonomy conservative while the staffing Note is new. Ask every time on External capabilities means cards land in Approvals for background agents; Run without asking logs in AI Activity instead. Deleting stays on Always asks.

Do not rotate the primary clearer every day. Rotation without a handoff Note creates Expired rows. When coverage changes, write the dates on the staffing table the same hour you update calendars.

Pending SLA board you define

Treat Pending like a ticket queue with clocks you set. Nautis does not ship a default SLA for how long a card may wait. Unattended requests expire on their own; that expiry is not a recommended decision path.

Illustrative reader-chosen board (composite only; replace the ages with your own):

Switch from Pending to History to audit Approved, Rejected, Expired, and Cancelled with filters. Count Expired weekly. Repeated Expired customer-facing cards mean your staffing or clocks failed, not that the product cleaned up for you. Fix staffing or pause the source routine before you change Autonomy.

Illustrative SLA add-on (composite only): customer-email drafts decided before lunch; internal Note drafts wait for a second daily pass. Night and weekend coverage needs an explicit rule: either the deputy clears, or External stays on Ask every time with a staffed Monday sweep.

Decision rubric: Approve, Reject, Cancel request

Approve confirms with product copy that the agent will execute now and that the action can have real effects. Use it when Show details matches an outcome you would type yourself.

Reject declines the action. Use it when the content is wrong, the timing is wrong, or the capability should not have proposed this. After repeated Rejects from one routine, pause that routine rather than training yourself to click Reject forever.

Cancel request withdraws the item; anyone can cancel. Use it when the underlying need vanished (meeting cancelled, duplicate card) and there is nothing to execute.

Illustrative reject (composite only): a routine proposes a discount email the founder never authorized. Reject, pause the routine, fix the brief.

If two Owners disagree on a card, Reject or leave Pending only long enough to escalate; do not Approve to end the argument. The execution is real. Capture the disagreement in the decision Note and fix the routine brief so the card does not return tomorrow in the same form.

For email-specific ask-before-acting detail, see AI that asks before acting. For the Chief of Staff overview, see What is an AI chief of staff.

Teach deputies the Approve confirmation line so they expect real effects. If they are unsure, Reject is safer than a hopeful Approve. You can always send a corrected message as a human after Reject.

Check AI Activity after you clear Approvals

Approvals is for deciding. AI Governance AI Activity is for reviewing what already happened. After you clear the queue, open AI Activity (Owners and Admins) and filter Approved and Auto-ran. Confirm the actions you just allowed show as Approved, and that Auto-ran rows are only for capabilities you intentionally set to Run without asking.

If a Pending card disappeared into Expired, read History, then tighten your clocks or the staffing Note. If Auto-ran External rows appear and you did not mean them, set those capabilities back to Ask every time under Agent Autonomy.

Routines that stop on asks first show Waiting for approval and may banner into Approvals. Clearing those banners belongs on the same cadence as the rest of the queue.

Briefing Do it proposals and chat or voice asks that require approval also land in the same Approvals list for Owners and Admins. One queue beats three mental inboxes labeled briefing, chat, and routines.

Healthy signals for an Approval Center for AI agents

A healthy queue shows:

• Pending near the empty state by the end of your chosen workday.
• History dominated by Approved or intentional Rejected, not Expired.
• AI Activity free of surprise External Auto-ran rows.
• One primary clearer who can name which routine creates most cards.

When Pending spikes, slow intake before you add more clickers: pause noisy routines, keep External on Ask every time, and reject customs that generate cards nobody reads. Speed returns after the queue is boring again.

Measure with three numbers on the staffing Note on a weekly rhythm you choose: median time-to-decision for External cards, Expired count, and Auto-ran External count in AI Activity. If Expired or surprise Auto-ran rises, do not celebrate a low Pending number earned by waiting for expiry.

What to write when you Reject or escalate

When you Reject or escalate, leave a one-line reason in your team Note or ticket: wrong recipient, bad dollar amount, needs legal, duplicate. Future you will read AI Activity and wonder why a card died. Approvals History shows Rejected, but it will not reconstruct your private reasoning unless you wrote it down.

Illustrative reason line (composite only): "Reject - discount not on the approved price list; pause sales routine until brief updated."

Approve can stay silent when Show details was enough. High-risk Approves (wire instructions, contract language, customer commitments) deserve a same-day Note anyway so finance and counsel can audit intent.

Connected AI apps that write into the workspace still follow the same autonomy rules and can create Approvals or AI Activity rows. Your clocks apply to those cards too; do not exempt them because the prompt started in another product.

An Approval Center for AI agents is running when decisions are timed by you, staffed by name, and logged, and when Autonomy changes wait until that sentence stays true for several quiet weeks.

Revisit the staffing table when you add a co-admin or when routines multiply. Scale with clearer ownership and tighter intake, not with more people Approving without Show details. When your weekly numbers stay green, you earned a separate Owners-and-Admins meeting about Autonomy changes.

Keep that weekly review on the calendar even when Pending stayed empty. Empty is a result worth confirming in AI Activity.