The right mental model
An AI employee behaves like a capable junior who has read everything you gave them, remembers all of it, works every hour, and has no instinct for when something is off. Every strength and every failure in the first month follows from that combination.
Teams that onboard one this way succeed. Teams that treat it as software that was purchased rather than a colleague that was hired spend the month confused about why it keeps doing exactly what it was told.
What works immediately
High-volume, well-specified, low-stakes work is productive from the first week: classifying and routing inbound messages, drafting first-pass replies, summarising calls and threads, extracting structured data from documents, answering questions that are genuinely documented somewhere.
These share a property — a human can verify the output in seconds. That is what makes them safe to hand over before trust is established.
What doesn’t work yet
Anything requiring undocumented context fails, and fails confidently. The exception a manager approved once, the customer everyone knows to handle carefully, the rule that was changed verbally in a meeting — none of that exists to a system trained on your written material.
It doesn’t know what everybody knows. That gap is the whole first month of work.
The month is mostly documentation
The realistic first-month plan is: run the AI employee on a narrow scope, review everything, and write down each piece of tribal knowledge it turns out to be missing. The corrections are the deliverable. Most of what you build is the written substrate your business never had.
This is why the exercise pays off even when the initial scope is small — the documentation produced makes every subsequent human hire faster too.
Set the scope deliberately
Pick one workflow, one channel, and one clearly bounded decision. Expanding scope before the corrections have flattened out just multiplies the review burden and disguises whether it is actually working.
What good looks like at day thirty
A narrow workflow running with occasional corrections, a written record of the knowledge that was missing, and a clear escalation path that fires on the right cases. Not autonomy, and not headcount replacement — a reliable junior on one job, which is the only foundation the later stages can be built on.