AI workers: a digital colleague is onboarded, not installed

By giving it an onboarding path, exactly as you would somebody joining, and by knowing where the analogy stops. An agent handed a standing workload needs the same four things as a new hire: access granted by somebody, a written scope, a person who reviews its work for the first weeks, and a place where what it learns is recorded.

The colleague vocabulary circulates widely, and it is worth more than people think provided it is taken seriously rather than as brochure imagery. If you are after the definition, the use cases and the return-on-investment calculation, Cobalt’s guide to AI workers does that job. What follows is about what the word commits you to once you have used it.

What the onboarding analogy brings

It brings a method your company already knows, and that is its main merit.

Nobody has to explain to an agency what a joining process is. You prepare the access before day one, you tell the person what they do and what they do not, you give them a reference point, you review their work for a few weeks, and you widen it when things go well. Applied to an agent, that sequence avoids the two most common mistakes: opening everything on day one, or opening reading only and wondering six months later why nothing has changed.

The human onboarding figures are worth recalling, because they show how neglected this stage is. An employee takes around twelve months to reach full productivity according to Gallup, and only 12% of employees think their company did a good job onboarding them. Structured onboarding is associated with 82% higher new-hire retention, and 33% of new hires leave within ninety days. A company that fails at onboarding a human will fail at onboarding an agent too, with the difference that the agent will not resign: it will keep producing mediocre work indefinitely.

Where the analogy breaks

On two points, and those are precisely the two that decide the outcome.

An agent learns nothing by osmosis. A new hire learns half the job in the corridors: they overhear a conversation, they watch somebody handle a difficult client, they work out that you never put that profile in front of that client. None of that reaches an agent. What it knows about your company is exactly what somebody took the trouble to write down, so onboarding means writing down what nobody ever needed to write.

That is tedious in the first weeks and it is the only work that really counts. The good news is that it happens as you go rather than as a project: each time an answer is corrected, the correction is written once and holds for every use afterwards. The bad news is that an agent which does not record what it is taught wastes everybody’s time twice, once discovering it and once the next time round.

An agent does not earn latitude by performing well. This is the least intuitive point. With a person, autonomy widens on its own through accumulated trust, and that works because a human knows when they are leaving their remit and hesitates. An agent does not hesitate. What it may not do on day one must therefore not become permitted because the first three months went well: widening is a decision taken by somebody, on a date, and written into the product rather than into an instruction. A rule written in natural language is worked around by another sentence in natural language.

The first thirty days, concretely

Four decisions, taken before day one, and none of them is technical.

What it can reach, and with whose rights. The useful answer is never “a dedicated account”: an agent works with the rights of the person talking to it, failing which it becomes a path around your permissions. Start with the systems you own, and leave your clients’ for later.

What it does, in what order. Tasks whose result is faster to verify than to produce first, which is the criterion that settles the first delegation. Preparing an interview from the file, saying who has been waiting ten days for an answer, pulling together what has been said with a client over six months.

What it never does. Writing to a candidate, a client or a partner without a person having read and released the message. That line does not move after three months, and that is what makes it hold.

Where what it learns gets written. Somewhere the company owns and several people can read, not in the history of one colleague’s conversation. That is the rule we apply to Balt, and it has a consequence worth trying: what one person corrects once, everybody has the next time round. What an agent learned working with somebody belongs to the company and does not leave with that person, which is one of the few real differences between an agent and a tool.

Should it get a name and a face?

A name, yes, because you have to call it something. A face and a biography, no, and that is the limit of the vocabulary used throughout this article.

The word “colleague” is useful internally: it forces decisions about access, scope and ownership that no tooling word forces. It becomes harmful the moment it goes outside. A candidate who receives a message signed with a first name and a photo, then finds out they were talking to a system, does not remember the technical feat; they remember being misled on a point where it cost nothing to be straight. So we disclose systematically, in every case where the agent addresses somebody outside, for economic as much as regulatory reasons: the legal trigger is direct interaction, ours is simpler.

The practical line is easy to hold. Inside, you speak of it as a colleague because that produces the right decisions. Outside, it says what it is in the first sentence, and it carries the company name rather than an identity borrowed from a person who exists.

Who is accountable for it?

A named person, and this is the question people forget because it does not look like a technical one.

An agent with no owner ends up exactly like a tool with no owner: still connected, read by nobody, and impossible to stop without calling a meeting. The role amounts to little: reviewing its work in the first weeks, deciding when the scope widens, and answering when somebody asks why it wrote that.

That accountability becomes plainly necessary once the agent starts working without being asked, on a recurring task triggered by a time or an event. There, nobody watches the run, and attaching it to a person is the only thing stopping a routine from outliving whoever created it.

And after six months?

The risk changes nature, and becomes the opposite of the day-one one.

A well-onboarded agent accumulates: notes, preferences, learned rules, exceptions recorded one Tuesday in March. Part of that stock becomes false without warning, because a client changed contact or an internal rule was dropped. An agent reasoning on stale material returns a credible, wrong answer, which is why what it forgets matters as much as what it keeps.

Onboarding therefore does not stop on day thirty, it simply changes object: you stop adding and start removing. It is less gratifying, it takes an hour a quarter, and it is the difference between a colleague who improves and a colleague who ages.