How to automate startup operations with AI agents

How to automate startup operations with AI agents is three habits in a row: schedule the boring jobs, gate every write that touches the outside world, and clear Approvals the same day. Skip any one of those and you get either surprise sends or a pile of waiting cards nobody reads.

This is not a prompt gallery. It is an ops pattern you can run in Nautis with Routines, Agent Autonomy, the Approval Center, and the AI Activity log. You can copy the same pattern in other stacks if you keep the gates honest.

Jobs worth automating first

Automate chores that already have a human checklist and a boring output. Do not automate judgment you have not written down.

Prefer catalog routines with a Read-only label while you learn the system. Custom instructions come after you trust permissions.

Preconditions before the first schedule

If your sources of truth are still forked across chat and spreadsheets, agents will amplify the fork. Stabilize the attention and execution loops first, then automate.

1. One home for tasks, one for deals, one for the cash figure you quote
2. A morning surface people already open
3. A written rule for who may approve external writes
4. A named Owner or Admin who will clear Approvals daily

Nautis Routines live under Chief of Staff. Only organization Owners and Admins can add, edit, run, pause, resume, or remove a routine. Anyone with Chief of Staff read access can see the list and run history.

Catalog first, custom later

Routines are standing orders for your AI agents. You pick a job from a catalog, or describe your own, and it runs on a schedule or when something happens in your workspace. After each run you get a summary, and anything the routine wants to change still follows your approval rules.

Add from Explore

Open Chief of Staff, then Routines, then Explore. Categories include Briefings, Meetings, Sales, Fundraising, Finance, Team, Operations, Growth, Research, Email, and Custom. Each card shows when it runs (or Triggered), a Read-only label if it changes nothing, and an estimated credit cost.

Configure name, settings, cadence (Every day, Weekdays, Weekly, or Monthly), time in the organization timezone, and how you hear back (In-app notification, Email summary, Chief of Staff chat). Read the Permissions box: what it Can read, what it Can write, and whether writes asks first or runs automatically.

Use Try it once before you rely on the schedule. That test still asks first on writes and uses credits.

Triggered routines

Some jobs fire on events such as a new lead or a logged decision. They show Triggered instead of a clock. There is no schedule to set, and Try it once is not available; the next matching event runs them.

Built-ins

Organizations start with built-ins such as Morning Briefing, Evening Wrap-up, Monday Weekly Digest, and Meeting Prep Nudge. You can customize timing or pause them. Untouched built-ins do not consume a plan slot; a built-in you change or pause does.

How to automate startup operations with AI agents without surprise sends

The scary failure mode is not a bad summary. It is an email or record change that left without a human look.

Set autonomy on purpose

Under Chief of Staff, Settings, Agent Autonomy, each capability is External or Internal with Ask every time or Run without asking. Ask every time means you see the action first and nothing happens until you approve. Run without asking means it runs on its own and is logged in AI Activity, with undo within 10 minutes where supported. Deleting stays locked to Always asks.

Start External writes on Ask every time. Narrow Run without asking only after the same template has been approved cleanly several times.

Clear Approvals daily

If a routine write asks first, the run shows Waiting for approval and a banner points to the Approval Center. Approve or Reject the same day. Unattended requests expire. A stack of stale cards is not automation; it is deferred risk.

For a policy table by action type, see the human-in-the-loop AI agents guide. Product detail for standing jobs is in the Routines help center.

Plan caps, credits, and pause behavior

Your plan caps how many routines you can have at once: 5 on Launch, 10 on Voyage, 15 on Fleet. That is a total, not a monthly allowance. Paused routines still count toward the limit. Removing a routine frees a slot.

Every run uses AI credits except where a card shows about 0. A run that cannot start because credits ran out counts as a failed run. After 3 failed runs in a row, a routine pauses itself and shows Paused after failures with the last error. Fix the cause, then Resume. Do not hammer Run now hoping the error disappears.

If the person a routine Runs as leaves the workspace, the routine pauses until an Owner or Admin clicks Run as me and resume.

Audit what actually happened

AI Governance is a running record of what your AI agents did: the action, its status, which agent took it, the model it used, and when. Open Chief of Staff, Settings, AI Activity, or the standalone AI Governance page. Filter by Auto-ran, Approved, Rejected, or Pending. Owners and Admins only.

Once a week, skim Auto-ran rows. If something ran that should have asked, flip that capability back to Ask every time. The AI Governance help center walks through the log fields and autonomy controls.

A practical weekly rhythm

Keep a human morning pass on blockers and approvals so automation feeds attention instead of replacing it. The daily founder briefing habit is the companion, not a competing product pitch.

Two operator stories (examples)

Example: read-only month

Nia turns on Meeting Prep Nudge and a read-only pipeline digest only. For four weeks every write stays on Ask every time. She clears Approvals during coffee. When a follow-up draft is wrong once, she Rejects it and edits the routine settings. No investor sees a bad send.

Example: autonomy flipped too early

Chris sets several External capabilities to Run without asking on day two so cards stop piling up. A routine updates a CRM stage incorrectly over a weekend. The AI Activity log shows Auto-ran. He reverts External writes to Ask every time and treats Approvals as a daily chore again. Speed returned after trust, not before.

What "boring" looks like after thirty days

A healthy automation month is quiet. Approvals are emptied daily. Auto-ran rows are rare and expected. Credit spend maps to a short list of jobs you can name without opening the app. Nobody asks in Slack whether the morning digest ran, because the notification or chat delivery is trusted.

If people still ask whether the digest ran, fix delivery settings or the routine's Run as person before you add more catalog cards. Silence without trust is just another kind of noise.

Mistakes that look like progress but are not

• Writing a vague custom routine that cannot ask follow-up questions, then blaming the model
• Letting Approvals sit until Friday
• Counting paused broken routines against the plan cap forever instead of fixing or removing them
• Measuring success by number of routines instead of hours of fake work removed
• Allowing contractors to change autonomy without an Owner present
• Automating outreach before the CRM is the system of record

Thirty-minute setup checklist

1. Open Routines and note your plan slot count
2. Add one Read-only catalog routine with Try it once
3. Confirm Agent Autonomy: External on Ask every time
4. Put a daily calendar block titled Clear Approvals
5. Bookmark AI Activity and filter Auto-ran each Wednesday
6. Write which Owner covers Approvals when you travel
7. Add a second routine only after the first week stays boring

Automate startup operations with AI agents until the scary part is rare and the boring part is scheduled. Gates and a daily clear beat a clever prompt every time.