AI Employee Onboarding Checklist: Your First 30 Days

Most teams bring on an AI employee and treat it like installing an app. Download, connect, leave. That is the fastest way to get burned — wrong replies in customer threads, access scopes nobody reviewed, a team that stops trusting the output before week two is over.

AI employee onboarding is an operations problem, not a technical one. The model works. The question is whether you built the rails. This checklist covers the first 30 days: what to define, when to review, and how to expand scope without breaking anything that already works.

If you want the definitions first, read What is an AI Employee?. If you want the training angle — the companion to this checklist, covering how to teach specific skills — read How to Train an AI Employee. This post picks up where that one leaves off: the operational setup around the teaching.

1. Day 1–3: Define the scope

Do not start with "handle support." Start with one job that repeats. Write it in one sentence.

Good scopes: "Reply to shipping and stock questions on WhatsApp, flag refund requests to a human." "Triage inbound Gmail into three labels: lead, support, and spam, then draft replies for the first two."

Bad scopes: "Help the team." "Be more productive." Any scope where the output cannot be checked by a human in under two minutes.

Write the scope down in a shared doc. Put it somewhere the whole team can see. That document becomes the single source of truth for every decision that follows — and it stops scope creep before it starts.

Three things to lock in during this window:

• One channel (WhatsApp, Slack, or Gmail — pick one)
• One job (not three variations of "help with everything")
• One person who owns the review

Use real tools as the backbone. Google Docs for policy, Notion for process notes, Slack for internal handoff. The AI employee reads from these; it does not guess from memory.

2. Day 4–7: Grant access deliberately

This is the step where most teams skip the thinking and regret it later.

Read every OAuth scope. If a tool asks for full Gmail send access plus Calendar write plus Drive read on a single consent screen, ask yourself whether the job you defined in step one actually needs all three. It probably doesn't.

The safer path: start with shared-channel access. An AI employee that works inside WhatsApp or Slack — alongside your team, not inside your accounts — has a smaller blast radius. It can read messages, draft replies, and hand off to humans without ever touching your Gmail or Google Workspace login.

If the job genuinely needs account-level access, grant one scope at a time. Start with read-only. Add write access only after the review cycle proves the system stays inside the lane you drew.

This is also where you check the privacy architecture of whatever product you are using. If the vendor cannot explain where your data lives and who can see it, stop. Onboard something else.

3. Day 8–14: Set the review cadence

Review mode is not a checkbox. It is the operating model for the first two weeks.

Run it daily. Every morning, the AI employee drafts output from the previous day. A human reads every draft before anything goes live. This takes 15–30 minutes — less if the scope is narrow.

Track three things per draft:

• Correct? Yes or no.
• On brand? Yes or no.
• Needed a human anyway? Yes or no.

Put the answers in a Google Sheet. After one week, you will see the pattern. The same three reply types fail over and over. Fix those. The rest are fine.

The goal is not zero mistakes. The goal is that every mistake happens in review, not in a live thread.

4. Day 15–21: Write the escalation rules

Escalation rules are the difference between an AI employee that earns trust and one that burns it.

Write them as a short table:

The list should fit on one screen. If you need a flowchart, you have too many rules.

Assign one person per shift as the escalation owner. If a thread hits the handoff rule at 9pm and nobody sees it until 9am, the rule doesn't matter. Put the escalation channel somewhere visible — a dedicated Slack channel, a WhatsApp group, a Gmail label that triggers a notification.

5. Day 22–28: Graduate from review to auto-send

Only automate the categories that have passed several reviews with zero edits.

Start with the easiest ones: shipping status, opening hours, basic pricing, order confirmation. These are high-volume, low-risk, and the answers rarely change. Switch them to auto-send and keep watching.

Leave the messy stuff — refunds, complaints, negotiation, anything with subjective judgment — in review mode. Some categories will never graduate. That is fine. The point is not 100% automation. The point is freeing the team to spend their attention on the cases that need it.

A good live setup by day 28 looks like:

• One channel active with auto-send on the safe categories
• 20–30 real examples in the knowledge base
• 5–10 escalation rules that everyone knows
• One owner per shift
• A daily summary posted to the founder or team lead

That is a working AI employee. Not a demo. Not a pilot. A system doing real work with guardrails that hold.

6. Day 29–30: Run the month-one retrospective

Do not skip this. The team has spent a month in the trenches. They know what works and what does not.

Sit down for 30 minutes with the reviewer and the escalation owner. Ask three questions:

1. Which categories can we add to auto-send next month?
2. Which escalation rules are firing too often — and should become auto-send instead?
3. What is the one piece of context the system still does not have that would cut handoffs in half?

Write the answers down. They become the scope expansion for month two.

This is also the moment where an AI employee starts to feel like an AGI employee. The same memory, tone, and escalation rules are now carrying across channels. WhatsApp, Slack, Gmail, Google Workspace — different surfaces, same operating model. The system shipped for one job in month one; in month two, it can take a second job without starting from zero.

What to do if it breaks

Three failure modes show up in the first 30 days.

Scope creep in week two. Someone adds a second channel or a fourth job before the first one is stable. Cut it back. One job, one channel, one reviewer. Ship that before anything else.

The escalation rules are too vague. If a handoff triggers and nobody knows who owns it, the rule does not work. Every escalation rule needs a named owner and a response-time target. "Hand off to support" means nothing. "Hand off to Sarah in #escalations, target 15 minutes" is a real process.

Auto-send was turned on too early. If you are catching mistakes in live threads instead of review, switch auto-send off and go back to step four. This is not a failure of the AI employee. It is a failure of the graduation process. Fix the process and try again next week.

The pattern that works is boring: narrow scope, deliberate access, daily review, clear handoffs, slow graduation. That is how an AI employee becomes useful. It is also how the team learns to trust it.

If you want the companion read on teaching specific skills, start with How to Train an AI Employee. If you want the big-picture definition, read What is an AI Employee?. And if the checklist lands for you, hire Perla and read more on Capabilities.

Frequently asked questions

How long does it take to onboard an AI employee?

The first useful output can come in under a week if the job is narrow and the rules are clear. A full 30-day cycle — review, tighten, expand — gets you to a stable deployment where auto-send is safe and the team actually trusts the output.

What is the single biggest mistake during onboarding?

Giving the AI employee full access to everything on day one. Start with one channel and read-only access to one source of truth. Expand access only after the review cycle shows the system stays inside its lane.

How many people need to be involved?

One owner and one reviewer are enough for the first month. Adding more people before the rules are stable creates conflicting feedback. Appoint one person who reads the daily output and one who can tighten the rules.

What if the AI employee makes a mistake in the first week?

That is what review mode is for. No message goes out without a human check until the same category has passed review several times without edits. A mistake caught in review is proof the process is working, not a reason to stop.

Is this the same thing as onboarding an AGI employee?

It is the same process, and the line between AI employee and AGI employee blurs once the system carries memory and rules across channels. What matters for onboarding is whether the same instructions work in WhatsApp, Slack, and Gmail without separate setup per channel.