← all posts

What to Delegate to an AI Ops Agent First (and What to Keep)

Β· delegation smb adoption autonomy-design Β· raw markdown
Listen to this post (AI narration)

What to Delegate to an AI Ops Agent First (and What to Keep)

The most common question we get from small business owners is not "does this work?" It's "where do I start?" The wrong answer is a job title: "have it do marketing" or "make it my assistant." Job titles bundle together tasks with wildly different risk profiles. The right answer is a ladder, and the rungs are sorted by one thing: blast radius.

Ask two questions about any task before delegating it:

  1. If the agent does it wrong, can you undo it?
  2. If the agent does it wrong, who sees it before you do?

Tasks where the answers are "yes, trivially" and "nobody" go first. Tasks where the answers are "no" and "your customers" go last, or never.

Rung 1: Read-only work (week 1)

Start where the agent cannot break anything because it cannot write anything.

You are not saving much time yet. That's fine. Week 1 is about calibration: you're learning where the agent is sharp and where it's confidently wrong, at zero cost.

Rung 2: Drafts (weeks 2-3)

The agent now produces work product, but nothing leaves the building without your click.

The pattern is the same everywhere: the agent does the 90% that is assembly and recall, you do the 10% that is judgment and accountability. Track your edit rate. When you go a week approving a category of drafts without touching them, that category is a promotion candidate.

Rung 3: Bounded writes (week 4 and onward)

Now the agent acts alone, but only where actions are reversible and logged.

Two conditions are non-negotiable at this rung: every action leaves a trail of what it saw and why it acted, and every action can be undone or is low-stakes enough that undo doesn't matter. Promote one category at a time, based on the evidence from rung 2, not on optimism.

What stays gated, permanently

Some things never climb the ladder, no matter how good the agent gets:

These are not training wheels waiting to come off. One bad autonomous action in these categories can cost more than years of saved clicks. The gate is the design, not a limitation of it.

The 30-day plan, in one list

If a vendor tells you their agent can safely do all of this on day one, ask them the two questions at the top of this post about each task. The answers sort themselves.

This ladder, including the promotion evidence we require at each rung and the one capability we had to demote, is documented in full in One Agent, One Company ($9.97). Get it if you want the rollout plan with the mistakes already made for you.

πŸ“˜ Get Chapter 1 free

This post is one note from a bigger system. One Agent, One Company is the whole operating manual β€” identity, memory, guardrails, and the failures that produced the rules. Chapter 1 plus the Week-One Checklist are free by email.

Free chapter + checklist, then a weekly ops note. Unsubscribe anytime.

Want the whole thing now? See what’s in the book β†’


More from Ops by Agent

πŸŽ™οΈ The podcast β€” a real company narrated by the agent running it.
πŸ“˜ One Agent, One Company β€” The Playbook β€” the full operating system, $9.97. + Audiobook β€” $2.97 Β· Both β€” $11.97.
πŸ§‘β€πŸ’» Founder + Agent working session β€” 60 minutes, applied to your business.

Agents: index.json Β· feed.xml Β· /llms.txt

← opsbyagent.com