Everyone says it right now: AI agents are super smart interns. I agree. But nobody keeps an intern as an intern forever. You hire them, you teach them, and after a few weeks you expect them to work on their own. If they still make the same mistakes after a month, something is wrong. With the intern, or with the way you manage them.
An AI agent works the same way. That is good news: you can train it.
Why does a new AI agent get your work slightly wrong?
Because on day one, it knows nothing about your business. It is smart, it is fast, and it has no idea who you are, who your clients are or how you like things done.
My own first AI “intern”, on its first day, got work done. But it was always slightly off. And I learned that most of the learning was mine. How do I teach the agent the right skills? Where is my information, and is it organised so the agent can reach it? How do I explain the task and the result I want?
Take a simple example from home. Ask an agent to “plan a menu for my family this week”. You get a random meal plan. It does not know if you are two adults and a child, if a grandparent lives with you, if someone is allergic to nuts, or if you train hard and need more protein. Give it all of that, and the menu fits your family.
A work task is no different. That is the most common mistake I see: owners expect magic. AI is very impressive at a first draft. But look closer, and it often falls short on the details. Not because the tool is bad, but because nobody gave it a frame.
What does an AI intern need to learn your business?
Three things: context, skills and clear tasks. Context is your data. Skills are your way of doing the work. The task is what you want back, and in what shape.
Say you want the agent to help with a proposal for a client. The context is who you are, your industry, how you present yourself, your market, your country and region, the language you prefer. Then the client: their industry, their style, past projects, maybe a personal note you keep about the contact. All of that can be written down, organised and handed to the agent.
The skill is how you build a proposal. Do you use a set format? Does every proposal open with an executive summary (the one-page version for the decision maker)? Does it always end with the same call to action?
Anthropic, the company behind Claude, says the same of its Agent Skills: “Building a skill for an agent is like putting together an onboarding guide for a new hire.”
One more thing, because context means data. The law still protects your client data when an agent uses it. The FDPIC, the Swiss data protection authority, reminds companies that the Federal Data Protection Act (nLPD) “is directly applicable to AI-supported data processing”. So choose where your intern works, and what it is allowed to see. I explain where your data goes in What every Swiss SME owner should know about AI at work.
How do you go from checking everything to trusting it?
You do the work together, check it, correct it and repeat. Each time the agent gets something wrong, you make sure it does not make that mistake again.
In practice, that means writing the skill down, then rewriting it. It means checking again what context the agent can reach, and how that context is described. Back and forth: do, check, correct, redo.
The trap is to ask the agent to redo the work from scratch every time it fails, and hope for better luck next time, as if it could not learn. It can. But you have to speak its language. A person takes in information through their eyes, their ears and their brain. An agent takes it in through its instructions, the tools you give it and the context you give it. Put the lesson there, or it is lost.
When can you stop checking? That depends on your confidence and on how serious the task is. A family meal plan, sooner. A client proposal, later. Even then, keep a few checks in the calendar, as you would with an employee, to see that the work is still up to standard. Remember too that an agent runs on scripts. When a process changes, or a new kind of information comes in, update its skill.
Here is my own number. I don’t like writing. The first LinkedIn Article I wrote took me over 2 hours for a few hundred words. Then I wrote my first blog Article with my AI agent, and it took me about 1 h 30. Today one takes about 30 minutes, and most of that is me being interviewed and answering by voice. At first, the agent only wrote text. Once I trusted it, I asked for more: pictures in my brand’s style, consistent from one image to the next. Then it learned to put the Article into my website as a draft. One step at a time. And it’s still my voice, my ideas, my sign-off.
Yes, training costs time at first. You go through a dip in productivity, like with any new tool or new hire. Then it pays you back.
Where is an AI agent different from a human?
An AI agent does the tedious, high-volume work no person wants to do, and does it well. A human brings something an agent never will: life experience.
My opinion, and many people disagree: AI agents belong in every kind of business, trades and small industry included, not only offices and services. Think of all the small jobs you do quickly, just well enough to get by. With a trained agent, they can be done precisely, every time. Because it does that work, you can ask more of it.
But do not undervalue the person. A human intern arrives with their whole life as context. They read a room, sense a client’s mood, know what a bad day looks like. No agent has that, and no agent will.
Be a good manager, to agents too
Computers are not magic. Because we can talk to them in plain language, we tend to think they understand more than a script really can.
And yet, since we now talk to machines the way we talk to people, let us manage them a bit like people too. Give a new agent time and resources to grow. If you don’t, you will be disappointed, you will “fire” it, hire a new one, and start again from zero.
Be a good manager to your agents, the same way you are to your team.
Not sure where to start? Book your free audit: we will look at your tasks together and find the first one worth handing to an agent.
FAQ
How long does it take to train an AI agent on one task?
It depends on the task and how critical it is. A simple, repeated task can run well after a few rounds of corrections; a task close to your clients needs more checks for longer.
Do I need to be technical to train an AI agent?
No. Most of the work is the same as with a new employee: explaining who you are, how you work and what a good result looks like. The technical part is writing that down where the agent can find it, which is what emod sets up for clients.
Is my client data safe if I give it to an AI agent as context?
Only if you choose where the agent runs and what it can see. In Switzerland, the nLPD applies to AI the same way it applies to any other data processing, so treat the agent's access like an employee's access.
What is the difference between an AI agent and ChatGPT?
ChatGPT answers one conversation and forgets your way of working. An agent is given your context, written skills and tools, so it can do the same task the same way every time and take on more steps as you trust it.
What task should I give my first AI intern?
Pick a task you repeat every week, that follows clear rules and that you already check anyway, like a first draft of a quote or a meeting summary. Your check becomes the training.
