Skip to content

How Do You Onboard a New AI Employee?

Onboard a new AI employee in 30 days: a written job, a context pack, its own key, a week of drafts only, supervised work, then a day-30 keep-or-cut call.

Reviewed by WaveBrain Operations TeamReviewed 8 min read

The short answer

You onboard a new AI employee by giving it a written job, a short context pack and its own key, then widening its freedom in stages over about 30 days: one week where it only drafts, one or two weeks of supervised work on low-risk tasks, then spot checks. At day 30 you decide to keep, expand, rewrite or retire it, based on the record of what it did.

Most owners "hire" an AI employee in an afternoon. They connect a tool, paste a long prompt and let it run. Then something odd happens on day three, nobody is sure what the agent was told, and the whole experiment gets switched off.

A human new hire wouldn't be treated that way. They'd get a role, a manager, someone to show them how things are done and a few weeks before anyone trusts them with the bank account. This guide turns that into a 30-day onboarding plan for AI agents that non-technical owners can follow, whether the new worker answers support tickets, drafts marketing posts or writes code.

What does onboarding an AI employee mean?#

Onboarding an AI employee means preparing everything it needs to do one job well, then letting it earn independence in stages instead of all at once. It covers the written job, the facts about your business it can't guess, the accounts it uses, who checks its work and how you'll know what it did.

The human version is often done badly. Gallup reports that only 12% of U.S. employees say their company does a good job of onboarding, and that its effectiveness "is largely contingent on the manager's active involvement in the process." The same is true for an AI worker, with one difference: an agent can't walk over to a colleague and ask what you meant. Whatever it needs to know has to be written down.

If you're still deciding whether you need an AI employee at all, start with our explainer on what an AI employee is.

What should you prepare before an AI employee's first day?#

Prepare four things before it does any real work: a job card, a context pack, its own access and a named owner. Each one takes minutes, and skipping any of them is the usual reason a new agent goes off track in its first week.

  1. A job card: half a page saying what it does, what it must never do, which tools it uses and what a finished task looks like.
  2. A context pack: the facts a new hire would pick up in their first week (see the next section).
  3. Its own access: a separate login or API key for each tool, with the smallest scope that works, so its actions are easy to tell apart from yours and easy to revoke.
  4. An owner: one person who reads its work, answers its questions and can pause it.

For the access step, our guide on how to give an AI agent access to tools safely has a table of what to give and hold back for common jobs.

What goes in an AI employee's context pack?#

A context pack is a short written briefing with the facts an agent needs every time it works. Keep it to things it would otherwise get wrong: your business, your rules, your voice and where to hand off.

SectionWhat to includeExample
The businessWhat you sell, to whom, prices it may quote"Plans start at $29/month. Never promise discounts."
PoliciesThe rules that decide its answersRefund policy, shipping times, what counts as urgent
Voice3 to 5 real examples of good past workTwo customer replies and a post you liked
Never doHard lines, written plainly"Never email the whole list. Never change prices."
Hand-offWho gets what it can't handle"Billing disputes go to Maria."

Coding agents already have a standard home for this. Claude Code reads a CLAUDE.md file at the start of every session, and can read a repository's AGENTS.md instead. AGENTS.md describes itself as "a README for agents" and says it's used by over 60,000 open-source projects. Anthropic's advice for these files works for any context pack: aim for under 200 lines, write instructions concrete enough to verify, and add a line when the agent "makes the same mistake a second time."

One warning from the same docs: Claude treats these files "as context, not enforced configuration." A context pack shapes what an agent tries. It doesn't stop it. Hard limits belong in the tool's permissions.

How do you run an AI employee's first 30 days?#

You run the first 30 days in four stages, and the agent only moves to the next stage when the current one goes well. The dates are a guide: a low-risk job can move faster, and a job that touches money or customers should move slower.

StageWhenWhat the agent may doWhat you do
1. Drafts onlyDays 1 to 7Prepare work, but a person sends, publishes or deploys itRead every draft and fix the context pack
2. SupervisedDays 8 to 14Act on its own for low-risk tasksCheck every finished task the same day
3. Spot checksDays 15 to 30Act on its own for its whole job, except actions needing approvalCheck a sample each week
4. Day-30 decisionDay 30Keeps its current access until you decideKeep, expand, rewrite or retire it

Anthropic recommends "extensive testing in sandboxed environments, along with the appropriate guardrails" before trusting an agent, and describes good agents as ones that "pause for human feedback at checkpoints." The drafts-only week is that sandbox for a small business: real tasks, no real consequences.

Some actions should never leave approval, even after day 30: payments, refunds, deletions, deploys, price or budget changes and emails to many people.

How do you know when an AI employee is ready for more freedom?#

An AI employee is ready for the next stage when its recent work needs no real corrections and it hands off what it should. A simple rule: move up after a week in which you'd have sent every draft unchanged, or close to it.

Watch for three things during each stage:

  1. Corrections: how many tasks you had to fix, and whether the same mistake came back after you added it to the context pack.
  2. Hand-offs: whether it passed billing disputes, angry customers or unclear requests to the right person instead of guessing.
  3. Reporting: whether every action it took shows up in its record. An agent that does work you can't see isn't ready for more freedom, however good the work looks.

If the same mistake survives two context-pack fixes, the job is probably too broad. Narrow it before you widen its access.

What should you decide at the end of onboarding?#

At day 30 you make one of four decisions, based on the record of what it did rather than how it felt. Read its last two weeks of work, the corrections you made and any failures, then pick:

  1. Keep: it does the job well. Move it to your normal weekly review of AI employees.
  2. Expand: it does the job well and has room for one more task. Add one, and run that task through stages 1 and 2 again.
  3. Rewrite: the work is uneven. Rewrite the job card and context pack, and repeat stage 2.
  4. Retire: it costs more checking than it saves. Revoke its keys and remove its tools.

Write the decision on the job card with the date. The next time someone asks why the agent can do something, the answer is there.

Where does WaveBrain fit in onboarding?#

WaveBrain gives you the record each onboarding stage depends on. It's a hosted activity log that your agents report their finished actions to, so you can see everything a new AI employee did without opening every tool it touched.

Agents connect through WaveBrain's MCP server, which gives them a log_action tool and instructions to call it after any action with an effect outside the chat, such as a post, an email, an ad change or a deploy. Each new agent can log under its own agent name and a display name for the bot, so a new hire's work is easy to filter. Scripts and no-code workflows log with one REST request. The dashboard suggests a separate API key per machine or agent, so you can revoke a new agent's key without touching the rest.

During the supervised stage, Slack is the easiest way to watch. Pick a channel and WaveBrain posts each action there as it happens, with who did it, what changed and whether it worked. You can send everything, only failures, or only sensitive actions such as deploys and payments, and Slack is included on every plan.

Be clear about what it isn't. WaveBrain isn't a permission system: it doesn't block, approve or sandbox anything, so the drafts-only stage still needs your tools' own settings. It records only what agents report, so an agent that never calls log_action leaves no entry. It doesn't score your agents for you. It isn't LLM tracing and doesn't record prompts, responses or token costs. It isn't a tamper-proof audit trail, and it doesn't filter secrets from the text agents write on the server, so entries should read like commit messages. See what gets logged for the full list.

The Free plan includes 1,000 actions a month, 7-day history and 2 API keys. Pro is $12 a month for 50,000 actions, 90-day history and 50 API keys. You can start free with no credit card.

Frequently asked questions#

How long does it take to onboard an AI employee?#

Plan on about 30 days for a job that acts in your accounts: a week of drafts only, one or two weeks of supervised work and two weeks of spot checks. Low-risk jobs can move faster.

Can I skip the drafts-only week?#

You can for jobs where every action is easy to undo, such as internal notes. For anything that reaches customers, money or production, the drafts-only week is where you catch the gaps in your instructions.

What is a context pack for an AI agent?#

It's a short written briefing with your business facts, policies, voice examples, hard limits and hand-off contacts. Coding agents keep the same kind of briefing in a CLAUDE.md or AGENTS.md file.

Should a new AI employee use my login?#

No. Give it its own login or API key with the smallest scope that works. Its actions stay separate from yours, and you can revoke it without locking yourself out.

Does WaveBrain onboard agents for me?#

No. WaveBrain records the actions your agents report, so you can review each stage and make the day-30 decision. The job card, context pack and permissions are still yours to set.

Sources#