An AI agent arrives knowing a great deal about the world and nothing about your organisation. It has read more than any new hire ever will, and it does not know which customer is handled differently, which approval is a formality and which is not, or why the March figures are always restated. The model is not what it lacks. What it lacks is the knowledge your people carry, and most of that knowledge has never been written down.
That makes onboarding an agent a knowledge transfer job, and an unusual one. A new colleague absorbs the organisation by sitting in it: they overhear, they ask, they get corrected. An agent absorbs nothing it is not given. Everything it will rely on has to move out of people's heads and into something written, owned and checkable. Organisations that treat deployment as a configuration task find this out in the first week, when the agent does exactly what it was told and the result is still wrong.
This note sets out what that transfer involves, and the question that tests each part of it.
The knowledge that matters is the knowledge nobody wrote down
Every team runs on a written process and an unwritten one. The written one says invoices over a threshold need a second approver. The unwritten one says that one supplier's invoices always need a second look, that a particular client's requests go to a named person whatever the ticket says, and that a report is never sent on the last working day of the quarter. People learn the unwritten process in their first month and stop noticing they know it.
An agent will follow the written process faithfully and break every unwritten rule, because it cannot know them. The first piece of onboarding work is therefore not technical. It is sitting with the people who do the job today and asking what they do that the procedure does not say: the exceptions, the judgement calls, the things they check without being asked. That list is usually longer than the procedure, and it is the most valuable thing the deployment will produce.
Write the role before you write the prompt
A new hire gets a job description, a manager and a set of systems they may use. An agent needs the same, stated more precisely, because it will not infer the boundaries a person would. What is this agent for? What may it read, what may it change, and what must it never touch? Who does it answer to, and who reviews its work?
A narrow role is not a limitation on value. An agent with one clear job and a short list of permitted actions can be checked, trusted and widened. One asked to be generally helpful across a department cannot be meaningfully reviewed, because nobody can say what correct behaviour would have been. The role definition is also where most of the safety work happens: the harness decides what an agent remembers, touches and must never do, and the harness can only enforce a role somebody wrote.
Memory is a record, not a recollection
An agent that starts every task from nothing is a tool. One that carries knowledge between tasks starts to behave like a colleague, and that persistent memory is where most of the value and most of the risk sit. Whatever the agent remembers, it will act on, so a wrong fact in memory is not a single mistake. It is the same mistake made every time the fact is relevant, with full confidence.
Treat agent memory as you would any other operational record. Each entry should say where it came from, when it was added and who approved it. What the agent learns from its own work should be proposed, not adopted: a person confirms it before anything relies on it. And memory should be readable by the people responsible for the agent, in plain language, so that they can find and correct what it believes. Knowledge that has been learned and approved is an asset. Knowledge the agent picked up and nobody has seen is a liability waiting for the right question.
Knowledge goes stale on a schedule nobody set
Policies change. Prices change. A supplier is replaced, a team is reorganised, a threshold is raised. When a person's knowledge goes out of date, someone usually tells them, or they notice. An agent keeps applying the old rule until somebody changes what it was given, and nothing in its behaviour signals that the rule has moved.
Every source an agent relies on needs an owner and a review date, the same as any controlled document. When a policy changes, the change process should include the question of which agents read it. And retiring knowledge matters as much as adding it: an agent with access to both the current and the superseded version of a procedure will sometimes use the wrong one, and will cite it convincingly.
The first weeks are supervised, and the supervision is designed
Nobody hands a new hire full authority on their first day. They shadow, they draft, someone checks their work, and their scope widens as they show they can be trusted with it. The same sequence works for agents, but it has to be built in, because an agent will not ask whether it is ready.
Start the agent on drafts that a person reviews and sends. Move to actions that need approval before they run. Widen its permissions only on evidence: a record of what it proposed, what was changed before it went out, and why. The corrections people make in those first weeks are the best source of the unwritten knowledge the first interviews missed, so capture them as proposed additions to the agent's knowledge rather than fixing each output by hand and moving on.
Escalation is part of the job description
A good colleague knows when to stop and ask. An agent has to be told. The most expensive agent failures are rarely the ones where it could not do the task. They are the ones where it did a task it should have handed over: a complaint that needed a person, a request outside its role, a case where the information was ambiguous and the agent chose an answer.
Define the conditions under which the agent stops, who it stops for, and what it hands over: what it was asked, what it found, what it was about to do and why it did not. A handover that a person can act on in a minute is worth more than any amount of autonomy. And route escalations somewhere that is read, with an owner, or the agent will have learned to ask for help from nobody.
What it changes for your people
Bringing an agent into a team changes what the experienced people in that team do. The person who knew the exceptions becomes the person who writes them down, reviews what the agent proposes to learn, and decides when its role can grow. That is skilled work, and it should be recognised as such, not added to an existing job as an afterthought. It is also the moment an organisation finds out how much of what it knows lives in a few heads, which is a risk worth knowing about whether or not an agent is involved. We have written before about making that knowledge measurable; onboarding an agent is often the first time anyone has to.
None of this is about the model. Every serious model will do the work once the knowledge, the role, the memory discipline, the supervision and the escalation path are in place, and none of them will do it well without. The organisations getting value from agents in production are the ones that treated deployment as onboarding.
We run that onboarding with you: interviewing the people who do the work today, writing the role and its boundaries, setting up memory that is owned and reviewed, designing the supervised first weeks and the escalation route, and training your team to own all of it afterwards. If you are about to put your first agent to work, talk to us before it starts.