What an AI “agent”
actually is.
People use the word agent to sound smart. Here is what it actually looks like when it is sitting on a desk, handling real paperwork, for a real business.
If you follow technology news, you have heard the word “agent” several thousand times in the last eighteen months. Most uses of it are promotional. An agent, in those contexts, means something that sounds impressive and is vague enough to mean anything. A chatbot with a few extra features. A piece of software that “takes action.” A productivity tool rebranded for a new cycle.
This article is about what an agent actually looks like when it is built for a specific business, is running on a real desk, and is handling real work. The concept is less mysterious than the marketing makes it seem.
The difference between a chatbot and an agent.
A chatbot is reactive. You open it, you type, it replies. Every conversation starts fresh. You bring it context each time you use it. It does not remember last week's client or last month's pricing update. It waits for you to speak.
An agent is proactive. It has a job to do, access to the things it needs to do that job, and the ability to act without you starting the conversation. You do not go to it. It works in the background and surfaces the results.
ChatGPT is a chatbot. It is a very capable one. If you need a writing partner or a research tool, it is the right thing. If you need something that watches your inbox at 11 PM and handles an urgent client query while you are asleep, a chatbot is not the tool.
What “having a job” means in practice.
When I build an agent for a client, the first thing we do is write a job description. Not for a person. For the software. It sounds strange the first time you do it. It produces useful results.
A job description for an agent might say: “You are the first reviewer for incoming client evidence files at a small law firm. When a new folder arrives in the shared drive, you read every document in it. You translate anything that is not in Swedish. You extract every name, date, and case reference. You flag any contradiction between what the client has written and what the documents show. You produce a summary brief for the lead lawyer, in the format they use, ready to review.”
That agent has a job. It knows what to do without being asked. It does not wait for a prompt.
The “set of keys” problem.
An agent without access to anything cannot do anything. This is the part the promotional material usually skips. An agent is only as useful as the tools you give it.
In practice, most business agents need three types of access. Read access to the information they work with — documents, emails, spreadsheet rows. Write access to record what they have done — filing a summary, updating a status column, adding a note. And the ability to alert you when something needs your attention — a message, an email, a flag in a shared system.
The access is scoped carefully. The agent can see the client folder; it cannot see your banking. It can update the status column; it cannot delete rows. This is deliberate. A well-built agent has the minimum access it needs to do its job, and nothing else.
Why “custom” is the important part.
A generic agent built for nobody in particular knows nothing about your business. It does not know that your senior partner likes summaries in bullet points, not paragraphs. It does not know that cases with a missing tax identification number are always flagged as urgent. It does not know your pricing, your preferred suppliers, or the specific way your forms are structured.
A custom agent knows all of that, because we taught it. The knowledge comes from your actual documents, your existing notes, the way you describe your own work. The agent does not guess. It applies the logic you already use.
Aiko · Solo accountant — she spent an afternoon writing down her own rules for reviewing a client's monthly books: what she checks first, what flags she looks for, what a clean set of accounts looks like versus one that needs a conversation. Her agent now applies those same checks to every client file before she opens it. By the time she sits down with the numbers, the boring first pass is already done.
What to be sceptical about.
-
“Our AI agent does everything automatically.”
No agent works without a human in the loop for anything that matters. The agent does the first pass, the human does the final check. Fully autonomous agents handling sensitive decisions are a liability, not an efficiency. Ask where the human review step is.
-
“Just connect it to your data and it learns.”
An agent learns what it is taught. If you hand it a pile of documents with no guidance on what to look for, it will produce something. Whether that something is useful depends entirely on whether you have told it what useful looks like. The training matters as much as the technology.
-
“It works for any business out of the box.”
Generic agents produce generic outputs. For administrative tasks that vary by industry, client type, and individual preference, generic is not good enough. A custom agent for your specific workflow will outperform a generic one every time, because it knows what good looks like in your context.
The short version.
An agent is a piece of software with a job description, a set of keys to the systems it needs, and the ability to work without you starting the conversation. The custom part is what makes it useful: the rules, the formats, the judgments your business has developed over years.
If you want to see what a specific agent would look like for your particular mess, the 60-second audit is the right starting point. I will write back with a plain-English description of what it would do and whether it is worth building.