Every team needs an AI employee

On 4 November I am at Banking Exchange in Frankfurt for a fireside chat with Max Schwarz about internal AI orchestration and how its return reaches customers (session details on my speaking page). Customer value is also where this post starts: with a question I was asked a few weeks ago, and what it means for anyone who manages a team.

The question from GenAI Berlin

At GenAI Conference Berlin in October, someone in the audience asked me roughly this: if every company ends up running AI agents, what is the point? Do we just end up with a lot of companies full of agents running around?

The answer I gave on stage is the one I still hold. A large share of the work people do today is repetitive. Repetitive work can be handed to agents. When it is, the business process that contained that work gets cheaper to run, and a company whose processes cost less has a choice about the difference. It can keep it as margin, or it can lower its prices, or it can give customers more for the same price.

Competition decides which. One company that cuts its process costs can keep the savings for a while. Once its competitors have done the same, the savings move into prices and into the product, because that is where companies compete. That is where the customer value comes from: the same service for less, or a better one for the same money.

And companies do not become interchangeable. Every company has had email and spreadsheets for decades, and owning them stopped being an advantage long ago. What still separates companies is how well they use them. With agents, that means how well each company has taught its agents its own processes, and that teaching is mostly management work.

What I mean by an AI employee

I have written about this twice already: what an AI employee is, and what happened when twenty teams onboarded one, which I also presented at WeAreDevelopers World Congress. The short version, because the rest of this post depends on it.

An AI employee holds a role in a team. It has its own accounts in the tools the team already uses, and work reaches it in two ways. A person asks it, in a direct message or by tagging it in a thread. Or an event starts the work: an alert fires, someone writes in a support channel, a Jira ticket gets a comment, a pull request is opened, a wiki page changes. It also works on its own initiative. It looks for problems nobody reported, proposes a fix and writes it up for the team.

A chatbot or a coding assistant waits until someone opens it and types. The work belongs to whoever typed, and whatever it learned from them stays with that one person. An AI employee is part of how the team runs. Its work is visible to the whole team, and a correction from any team member improves it for all of them.

Every department, not only engineering

Engineering got there first because our work already lives in tools an agent can read: pull requests, tickets, logs. That head start is the only reason. In June I listed teams that onboarded without any rollout programme: lawyers checking contract language against earlier decisions, marketing pulling campaign data together without waiting for an analyst, business leads who want meeting action points recorded the way their team actually decides, and customer support receiving a diagnosis with each complaint instead of a raw ticket.

The work in every department has the same two parts. One part needs judgement, relationships and someone who is accountable. The other part is the four-hundredth instance of something that has been done before: the same question answered in a channel, the same report compiled, the same check before something goes out. The second part is where an agent belongs.

So my position is that in 2026 and 2027, every team should have at least one AI employee alongside its human ones. A team without one is paying people to do work a well-briefed system could do, and those people are not doing the work only they can do.

Plan the team with both

For a manager, the change starts at planning. Today, when a team is overloaded, the request is another headcount. The first question I would ask is what the overload consists of. If a large share of it is repetitive and can be described, it goes to the AI employee, and the next human hire is for the work that needs a human.

This is not a plan for fewer people. I wrote in September that the aim is to automate as much repetitive work as possible and replace no one. What changes is how the people on the team spend their week.

It also changes what a manager prepares before anyone joins. The process has to be written down: what it does, why it exists, how it is executed including the exceptions, and what a good result looks like. A new human hire can survive a vague brief, because they ask colleagues and fill in the gaps over a few months. An agent with a vague brief improvises, and improvises differently each time. The written role pays off twice, because the next human hire gets the same document on their first day.

Week one is mediocre, for both

Nobody expects a new hire to be good in their first week. They do not know the systems yet, or which stakeholder needs a heads-up before a change, or which alert is noise. They get better because colleagues correct them, specifically and soon after the mistake.

An AI employee starts the same way. In its first weeks it answers questions slightly off, escalates things it could have handled and misses context the team takes for granted. A team that judges it in week one and gives up is making a call it would never make about a person.

The improvement loop is the same as well. When a person gets something wrong, a colleague tells them. When the agent gets something wrong, a team member flags the response and writes a sentence about what was wrong. The system turns that into a proposed change to its own knowledge, and a person approves it before it takes effect. In June I described teams seeing several visible changes in its behaviour every day.

Two differences are worth naming. The agent does not forget a correction after a holiday, and a correction one person gives improves it for everyone it works with. The human brings what the agent lacks: knowing when a rule should be broken, and being the one who answers for it.

Managing an AI employee is a real job

If every team has an AI employee, someone in each team owns it, the same way someone owns each new person’s onboarding. In practice that means four things, and none of them are new to anyone who has managed people.

A named owner. When the agent makes a bad call, a person answers for it. The agent cannot.

A written scope. What it handles, what it escalates and what it must not touch. Its access matches that scope. An agent holding more permission than its task needs is what OWASP calls excessive agency, and combined with prompt injection the damage reaches as far as its credentials do.

Regular review, the way you would hold a one-to-one. What it handled, what it escalated, where it was corrected most, and what each run costs. That last one only works if every run is attributed to a process and a team.

Feedback as a team habit. If only the owner corrects it, it learns one person’s view of the job.

The manager’s skills transfer directly. The one real difference is that an agent’s performance record is complete, which is more than any manager has ever had for a person.

The objections I hear

We tried this with RPA. RPA automated steps that never vary, and broke on every exception someone had not encoded in advance. An agent can be given the reason a process exists and make a bounded judgement on a case nobody anticipated. That widens automation from work that never varies to work you can describe, which I covered in more detail in September.

Why not give everyone a chat assistant? Because the work stays with whoever opened the chat. Nothing happens when an alert fires at 3 AM and nobody is typing. A correction one person makes reaches nobody else. Personal assistants make individuals faster, but the cost of a business sits in its processes, and a process does not change because one person in it got faster.

Agents make mistakes. So do new hires, which is why nobody gives a new hire production admin rights in their first week. Scope the access, put a human approval in front of anything irreversible, and review what it got wrong.

See you in Frankfurt

If you are at Banking Exchange in Frankfurt on 4 November, my fireside chat with Max Schwarz is at 11:35. Ask me the GenAI Berlin question again, or tell me which team in your company should get its AI employee first.