“AI agent” is a term you now hear everywhere, rarely with an explanation. In plain business language, an AI agent is software that is given a task rather than a question: it reads what arrives, works out the next step within rules your company sets, carries it out in your tools and passes anything important or unusual to a person. This guide explains how that differs from a chatbot and from a fixed automation, and how to tell whether your company needs one.
What an AI agent does in a company
An agent combines three things: AI that understands text and documents roughly as a person does, access to the tools your company already uses, and rules that say what it may do alone. That combination is what people mean by agentic AI: AI that does not stop at an answer but carries a task through several steps. Typically, an agent does four things:
- It reads what arrives. Emails, invoices, enquiries and forms, in German, French, Italian or English, however each sender writes them.
- It decides the next step. Within your rules: which customer or supplier this is, who should handle it and whether anything is missing.
- It works across your tools. It looks up an order, prepares a booking, files a document or drafts a reply, in the system where the work belongs.
- It hands over what matters. Anything important or unusual goes to a person with a short summary, instead of being guessed.
For a task with several steps, the agent plans, acts in your systems and checks its result before anything is final.
How an agent differs from a chatbot
Most people know AI as a chat window: you type a question and get an answer. For individual questions, a chat tool is often all a person needs. But it waits until someone asks, helps one person at a time, knows nothing about your company unless you paste it in, and someone still has to copy the result into your other tools.
An agent works the other way round. It starts by itself when an email, invoice or enquiry arrives, works with your documents, prices and rules, and enters the result directly into your tools. A chatbot answers questions; an agent does a piece of work.
How it differs from a fixed automation
Many companies already automate with fixed rules: when a form arrives, copy these fields into that system. Such rules are fast, cheap and predictable, as long as every input looks exactly as expected. When a supplier changes the invoice layout or a customer writes in French, the rule stops or does the wrong thing.
An agent understands content that varies, so it can handle cases nobody wrote a rule for. In exchange, it can make mistakes, as people do, so it needs checks. A well-built system often combines both: fixed rules where the input never changes, the agent where understanding is needed.
Tasks that suit an agent, and tasks that don’t
An agent earns its place where a task has several steps, comes up every week and follows rules you could explain to a new team member, but arrives in forms too varied for a fixed rule. Examples from a Swiss SME:
- Supplier invoices. Read the PDF or QR-bill, match it to the order, prepare the booking and set aside anything that does not match.
- Customer enquiries. Recognise what the customer wants, look up the order and draft a reply in the customer’s language; complaints go to the right person.
- Payment matching. Match the e-banking statement with open invoices and list what is still unpaid after the due date.
An agent is a poor fit when the task happens only a few times a year, every case needs a new, personal judgement, nobody can say what a correct result looks like, or a mistake would cause harm before anyone could catch it.
Human approval and checks
An agent that acts in your systems needs limits, and you set them before anything is built:
- Approval before anything leaves. Replies, bookings, payments and anything else that leaves your company or costs money wait for a yes from a person.
- A check before every important step. The agent checks its own result, for example that an amount matches the order, before it moves on.
- Flags instead of guesses. Missing or unusual information goes to a person rather than being filled in.
- A visible record. You can see what the agent did, so a mistake can be found and corrected.
The agent should reach only the data and systems the task needs. The written offer states which data the AI sees, which provider processes it and where, so you can meet your duties under the revDSG. If data must stay in-house, models can run on your own infrastructure.
How to start with one task
The best first agent is narrow: one task that costs time every week. Write down three things:
- The task as you would explain it to a new team member, including the exceptions.
- A handful of real examples, as they actually arrive.
- What a correct result looks like, and where a person must approve it.
Bring these notes to a first conversation with MomentumQ: 30 minutes, free and without obligation, online or in Zürich by appointment. For a structured view of several workflows first, the review “Find the right starting point” takes one week, costs from CHF 2'500 and ends with prioritised suggestions with estimated value and a written plan.
A defined project, such as an agent for one agreed workflow, starts from CHF 8'000 and typically takes three to eight weeks. The written offer states the scope, timeline, price and expected running costs before any work starts.