Onboard an AI agent the way you onboard a new hire
An AI agent is not a smarter search box. It is software that takes a task, decides the steps, uses your systems to carry them out, and reports back. In other words, it does work. And anything that does work in a company needs what every new hire gets: a role, an account, limits, a manager and, one day, a way out.
Before the first day
Write the job description. One sentence on what the agent is for, such as "reconciles supplier statements with our ledger every Monday", and a list of what is not its job. An agent with a vague mandate will always find something to do. It will not always be what you meant.
Give it its own account. Never a borrowed login or a shared service password. In its own name, every action it takes can be traced to it, and it can be stopped without locking out a person.
Grant the rights the role needs, and no more. Read access to the statements and the ledger; the right to propose entries, not to post them. Widen its rights later, once it has earned them.
The first weeks
Name a manager: a person who reviews its work, answers its questions and is responsible for what it does. An agent with no manager is not autonomous. It is unsupervised.
Start with a probation period. Everything it proposes is checked by a person, and every correction is noted. The corrections are the real training: they show where the instructions are unclear and where data is missing.
Measure it like any colleague: how much it finishes, how often it is corrected, how often it asks for help. When corrections fall, give it more room. When they rise, find out why before anything else.
Leaving
Agents get replaced too: a better one comes along, a process changes, a contract ends. Plan the exit from the start. One switch that removes its access everywhere, its open tasks handed to a named person, and its record kept, because questions about its decisions will come after it is gone.
Why the comparison holds
Companies already know how to bring a stranger into the organisation safely: hiring, onboarding, supervision, review, departure. These practices were built over decades for people. They fit agents almost unchanged, and they are far easier to explain to an auditor than a new framework invented for the machine.
Treat the agent as a colleague and you will ask the right questions before it starts. Treat it as a feature and you will find the answers out later.