What an agent actually is

You have probably used a chatbot. You typed a question, it produced an answer, and you copied the useful part into an email. That experience was real, and it is exactly why so many executives now carry the wrong mental model into what is, in practice, a staffing decision.

Because the thing arriving in companies now is not a better chatbot. It is something your organization already knows how to think about, once you see it clearly: a new kind of worker.

A chatbot answers, an agent acts

The distinction fits in one line. A chatbot answers questions. An agent does work.

Ask a chatbot "how should I respond to a supplier who missed a delivery date?" and you get a well-written suggestion. Useful, the way a knowledgeable friend is useful. But you still open the inbox, find the email, write the reply, update the order, and inform the warehouse. The work remained yours.

An agent is given the inbox. Overnight, 80 supplier emails arrived. The agent reads all of them, drafts replies to the routine ones, updates delivery dates in the order system, and puts three emails on your desk with a note: these need a decision from you. One supplier is asking for a price increase, one is signaling a delay that affects a customer commitment, one is a complaint that smells like a relationship problem. The other 77 are handled and logged.

That is the whole shift. Not better answers. Finished work.

The word agent itself deserves a plain definition, because the technology industry rarely offers one. An agent is a digital worker. It is closer to a new hire you give an email account and instructions to than to software in the usual sense: it works through tasks on its own, deciding step by step what to do next. Traditional software follows a fixed script. An agent reads the situation, like a person, and chooses its next step, like a person. Just much faster, and without ever getting bored.

The anatomy of an agent

If you were to onboard a junior employee next Monday, you would prepare four things. An agent needs exactly the same four. This is not a loose metaphor, it is genuinely how these systems are set up.

The job description. An agent runs on written instructions, in plain language: what the role is, what good work looks like, what the priorities are, how to handle the common cases. Whoever writes these instructions is doing management work, not technical work. A vague brief produces vague work, from people and from agents alike.

The brain. Underneath every agent sits a model, the brain the agent runs on. You rent it from a provider the way you rent capacity from a payroll bureau, a supplier you choose and manage like any supplier. Anthropic, for example, is a provider, and its model is called Claude. A newer model is simply a sharper brain, and you can switch. You do not build the brain, you do not maintain it, and you should never let anyone tell you that you need to.

The keys. An agent can only work in systems it has been given access to. Each of these is called a tool: the email account, the calendar, the bookkeeping software, a folder of contracts. Think of the logins and permissions you give a new employee on day one. No access, no work. Too much access, too much risk. The list of tools is a decision you make, not something the agent grants itself.

The limit. Finally, every agent needs a mandate: the limits you set on what it may do alone, what it must ask about, and what it may never touch. It is the same idea as an attestation limit for a new employee. A controller fresh out of school can book invoices under 10,000 SEK alone but escalates anything above. An agent gets a written version of exactly that rule.

Notice what is on that list, and what is not. There is no programming on it. Setting up an agent well is a delegation exercise, and the people in your company who are good at delegation, the ones who write clear briefs and set sensible limits, are the people who will be good at this.

One invoice, minute by minute

Abstractions hide more than they reveal, so here is one agent doing one task, end to end, at a 120-person logistics firm in Jönköping. The agent's job: handle inbound supplier invoices.

09:12. An invoice arrives as a PDF attachment in the invoices inbox. The agent opens it and reads it the way a clerk would: supplier name, invoice number, amount of 48,750 SEK, 30-day payment terms, a reference to purchase order 4471.

09:13. The agent looks up order 4471 in the bookkeeping system, one of the tools it has been given. The quantities match the delivery. The amount does not. The invoice is 1,850 SEK higher than the order, and a line at the bottom mentions a freight surcharge that was never agreed.

09:14. Here the mandate decides what happens next. The rule says: deviations above 500 SEK are escalated, never booked. So the agent does not book the invoice. Instead it writes a short note to the controller: here is the invoice, here is the original order, here is the 1,850 SEK difference, here is the supplier's stated reason, and here is the relevant clause from the framework agreement, which does not mention freight surcharges. The whole package lands in her review queue.

09:15. The next invoice arrives. This one matches its order exactly. The agent books it as a draft awaiting her sign-off and moves on.

By the time the controller finishes her first meeting of the morning, the day's invoices are processed: the clean ones booked as drafts, the exceptions sitting in her queue, each with a one-paragraph explanation. Her job has quietly changed from typing in invoices to reviewing decisions, which is what she was actually hired for.

Nothing in that sequence is magic. Every single step is one a junior clerk performs today. The difference is that the agent reads 200 invoices in roughly 20 minutes, works nights and month-end peaks without overtime, and records every action it takes, which makes it easier to audit than a busy human. The difference is speed and consistency, not intelligence beyond yours.

Why this exists now

A fair question: if this is so sensible, why was nobody offering it in 2022? Because the technology could not do it. The models of that era could answer questions impressively but could not be trusted to act: to use systems, hold a multi-step task together, and know when to stop and ask. Around 2024 that line was crossed. Models became reliable enough at acting, not just answering, for serious companies to hand them real work under real limits. That is the entire history lesson. The capability is young, which is precisely why the advantage of understanding it early is real.

What an agent is not

A definition is only trustworthy if it includes the limits, so here are five, stated plainly.

It is not a robot. There is no machine in the hallway. An agent lives in your systems, with logins and permissions, like any office worker who happens to have no body.

It is not a colleague with feelings. It has no ambition, no loyalty, no judgment of its own beyond its instructions, and it does not care about the company. The warmth in its writing is craft, not sentiment. Treat it as capacity, not as company.

It is not general intelligence. It is a capable worker within the role you define. Move it outside that role and quality drops fast, the way a brilliant accountant does not automatically make a brilliant negotiator.

It is not always right. Sometimes an agent produces a hallucination: it states something false with full confidence, like a new hire who guesses rather than says "I don't know." You manage this the way you manage any unproven employee, with spot checks and sign-off rules, not with blind trust. This is also why the mandate exists.

It is not a replacement for judgment. In the invoice story, the agent surfaced the 1,850 SEK question. A person answered it. Deciding what the company should accept, risk, and stand behind remains your work, and nothing in this guide will suggest otherwise.

One more boundary worth drawing: an agent is not the same as automation, the fixed routines your company may already run, which do the same steps every time, like a dishwasher. An agent is more. It handles variation and decides what to do when the steps are not obvious. Automation breaks on the exception. An agent's job begins there.

The sentence to keep

The rest of this guide builds on this one mental model: what agents can and cannot do, what they cost and return, where the risks really sit, and how to start small. But if you take a single sentence into your next leadership meeting, take this one.