Use your first 30 days to learn how decisions flow, then prove one thing a week: that you can read the workspace, review held decisions correctly, raise a problem the right way and improve one part of the operation. The plan below sets a goal, an output and a check for each week, so your manager sees progress without asking.
What is a 30-day plan for in this kind of job?
AI operations work is judged on decisions, not activity. You will review what agents propose, decide which ones go ahead, catch the ones that should not and help the team improve the agent's brief and sources. A 30-day plan makes that visible early, while your manager is still forming a view of you.
It also protects you. New starters in this work often get dropped straight into a review queue with no context. A written plan gives you a reason to ask for the policy library, the escalation routes and the red lines before you approve anything. For the mindset behind the plan, see how to make the most of your first 30 days in an AI operations job.
How do you fill it in?
Agree the plan with your manager in your first two days. For each week, write one goal, the outputs you will produce and how both of you will know it worked. Use your employer's language: if they call a held decision a "pending approval", write that.
Keep outputs small and real. A one-page map of the escalation routes beats a long report. If your role is remote-first, add one written update a week, because nobody sees you at your desk.
The 30-day plan template
| Week | Goal | Outputs | How you prove it | Questions to ask |
|---|---|---|---|---|
| 1 | Understand the operation | Map of agents, their capabilities and their depth; list of red lines; escalation routes | Manager agrees the map is right | Which decisions can only a person make? Who signs off a change to an agent's limits? |
| 2 | Review under supervision | [Number] held decisions reviewed alongside a colleague, each with a reason | Colleague agrees with your call on [most / all] cases; disagreements discussed | What does a missed red line look like here? |
| 3 | Review on your own | Your own queue for [hours or cases]; a log of every amend, reject and escalation | Spot-check by your lead shows no missed red line | Where does the decision record live, and who reads it? |
| 4 | Improve one thing | One proposed fix to a brief, source or limit, with the cases that show the need | Fix is reviewed by whoever owns the agent | What would you like me to own from next month? |
What should you be able to prove by day 30?
Three things. First, that you know where the lines are: what agents may do alone, what is held for a person, and what is refused. Second, that your review decisions are right and explained. Third, that you notice patterns, not just single errors. One good proposal backed by five cases tells your manager more than fifty clean approvals.
Do not try to automate anything in month one. If you find a change worth making, write it up for the agent's owner. Changing an agent's limits is normally someone else's decision, and respecting that early earns trust.
A filled sample
Here is the plan for a new operations analyst we will call Amara, joining a mid-sized lender's collections team, which uses agents to draft payment-plan offers:
| Week | Goal | Outputs | How she proves it |
|---|---|---|---|
| 1 | Understand the operation | Map of three agents: one reads accounts (depth 1), one drafts payment-plan letters (depth 2), one books plans within limits (depth 3). Red lines: no contact with customers flagged as vulnerable without a person. | Her lead, Callum, corrects two items and signs off the map |
| 2 | Review under supervision | 60 held payment-plan drafts reviewed beside a colleague | Agreement on 57; three disagreements discussed and logged |
| 3 | Review on her own | Her own queue three mornings a week | Spot-check of 40 cases, no missed red line |
| 4 | Improve one thing | Proposal: the letter agent's source on hardship rules is out of date; six held cases show it | Agent owner accepts and updates the source |
The week-four proposal is the strongest output. It shows Amara read the decision record, found the cause rather than the symptom, and sent it to the right owner.
Questions people ask
What if my employer already has an onboarding plan?
Use theirs and add the "how you prove it" column. Most onboarding plans list tasks; few say how you will show you did them well.
Should I share the plan before I start?
You can mention it at offer stage, but agree the details in your first week, once you know the systems. If you are still job hunting, the guide to getting your first AI operations job and the wider AI agent jobs and careers guide come first.
What if I make a wrong call in week three?
Log it, say so and show what you changed. A reviewer who reports their own error is more useful than one who hides it, and the decision record will show it anyway.