What to Delegate to an AI Ops Agent First (and What to Keep)
What to Delegate to an AI Ops Agent First (and What to Keep)
The most common question from owners considering an AI ops agent isn't "can it work?" — it's "where do I start without betting the business on it?" The wrong answer is to pick tasks by job title ("it does marketing now"). The right answer is to pick tasks by blast radius: how bad is the worst case, and how easily can you undo it?
Here's the delegation ladder that answer produces, and a 30-day plan to climb it.
Rung 1: Read-only work (day 1)
Start with tasks where the agent can only look, summarize, and recommend:
- Inbox and message triage: classify, prioritize, digest
- Report compilation: pull the numbers, assemble the summary
- Monitoring: watch dashboards, renewals, deadlines, and flag anomalies
- Research: vendor comparisons, competitive scans, meeting prep
Worst case at this rung is a wrong summary — annoying, never catastrophic. You get immediate value (mornings back) and, more importantly, you get evidence: a daily record of the agent's judgment you can grade before trusting it with anything sharper.
Rung 2: Drafts, not sends (week 2)
Next, let it produce work that a human reviews before it leaves the building:
- Email replies drafted in your voice, queued for your approval
- Invoice and proposal drafts
- Social posts and content, staged not published
- Calendar suggestions, not confirmed bookings
The agent does 90% of the labor; you keep 100% of the authority. Track the edit rate. When you're approving drafts without touching them for a couple of weeks, that's your promotion signal.
Rung 3: Bounded writes (weeks 3–4)
Now allow autonomous action inside explicit limits:
- Publishing routine content on a fixed schedule
- Updating internal records and CRM fields
- Sending templated, low-stakes messages (appointment reminders, receipt confirmations)
- Filing, organizing, archiving
Two conditions make this rung safe: every action must be either reversible or verified after the fact, and every action must be logged. An agent that acts autonomously but can't show its work hasn't earned this rung, whatever its accuracy.
What to keep gated — possibly forever
Some things stay behind a human gate not because the agent can't do them, but because the worst case can't be undone by an apology:
- Moving money: payments, refunds, pricing changes
- Commitments to customers: scope, timelines, SLAs
- Hiring, firing, and anything touching people's livelihoods
- Public statements in your name beyond routine content
- Deleting anything that isn't backed up
The gate can be lightweight — a one-tap approval on your phone — but it exists, and the agent knows it exists. Good agents are designed to prepare these decisions beautifully and refuse to make them.
The 30-day rollout, condensed
Week 1: read-only triage and reports. Week 2: add drafts, grade the edit rate. Weeks 3–4: promote the two or three tasks with the cleanest record to bounded autonomy, and write down what stays gated. At day 30 you have something most automation projects never produce: a documented trust boundary, earned with evidence instead of granted on faith.
Delegation to an agent isn't a leap. It's a ladder, and the first rung is deliberately close to the ground.
The book covers this ladder in full — including the promotion criteria, the demotion stories, and the gates that never open. If you want the 30-day plan as a worked playbook rather than a blog post, that's where it lives.