Examples of AI agents: four back-office tasks in administration and SMEs
Contents
Examples of AI agents in everyday office work have one thing in common: each agent handles exactly one Vorgang, which means one intake, one task and one destination. It classifies, extracts details, routes and creates a task. When it is unsure, it hands the case to a person. Four examples from public administration and SMEs:
| Vorgang | Intake | What the agent does | Target system | Unsure cases go to |
|---|---|---|---|---|
| Posteingang (incoming post) | Scanned letter post as PDF | Captures the item, assigns it to the responsible department, creates the task in the DMS | DMS | a named person in the Poststelle (mailroom) |
| Bürgeranfrage (citizen enquiry) | Enquiries arriving in one shared mailbox | Recognises the request, routes it to the responsible office, confirms receipt | ticket system or DMS | a named person in the citizens' office |
| Lieferschein (delivery note) | Delivery notes as scan or PDF | Reads the details, checks them against fixed rules, forwards them to the warehouse, quality assurance or accounting | ERP or inventory system | a named person in goods-in |
| Rechnung (invoice) | Invoices as PDF, scan or e-invoice | Captures the header data, proposes the assignment, hands it to accounting | the accounting team's existing system | a named person in accounting |
An AI agent works on its own, but only inside one Vorgang
The EU AI Act describes an AI system as a machine-based system designed to operate with varying levels of autonomy, which infers from its input how to generate outputs such as predictions, content or recommendations (EU AI Act, Regulation (EU) 2024/1689, Article 3). An AI agent is software that carries out a task on its own in this way.
How much autonomy makes sense is set by the Vorgang. In all four examples the agent assigns cases on its own, but within fixed limits: one intake, one task, one target system, and a confidence threshold below which it stops and hands the case to a person.
Incoming post: the agent assigns each item to the responsible department
In many organisations someone reads the incoming post every day and distributes it to departments. The agent takes over that assignment. It captures the incoming PDF, recognises sender, date, subject and, where present, the file reference, assigns the item to the responsible department and creates the task in the DMS.
The Poststelle keeps opening and scanning as before. The agent does not decide on the matter, set deadlines or delete anything. How this works in a public administration is described in our article on mailroom automation.
Citizen enquiries: the agent recognises the request and confirms receipt
Citizens write to their local authority, often without knowing which office is responsible. The agent recognises the request, routes it to the responsible office and confirms receipt. The office gives the answer on the matter.
The same boundary applies here. The agent does not answer an enquiry on its substance, issues no decisions and sets no deadlines. When it is unsure what the request is about, the enquiry goes to a named person who assigns it.
Delivery notes: the agent reads, checks and forwards
In an SME, every delivery arrives with a delivery note, on paper or as a PDF. The agent reads the details, such as supplier, delivery note number and line items, checks them against fixed rules and forwards the note to the warehouse, quality assurance or accounting, in exactly one target system such as the ERP or inventory system.
Which rules apply is agreed with the department in week one: which details every delivery note should carry and which remarks point to a discrepancy. The agent does not post goods receipts and releases nothing. More in our article on delivery note processing.
Invoices: the agent captures the header data and proposes the assignment
Incoming invoices arrive as PDFs, scans or e-invoices. The agent captures the header data, such as supplier, invoice number, date and amounts, proposes the assignment to a department or cost centre and hands the invoice to accounting as a task in its existing system.
Checking, approving, posting and paying stay with accounting. The agent makes a proposal, not a booking. More in our article on invoice capture.
Four AI agent examples, one boundary: the matter stays with people
The four examples differ in their intake and target system, not in their boundary. The agent assigns, extracts, routes and creates a task. Anything with legal effect or that moves money stays with the people in the department.
| What the agent does | What the agent never does |
|---|---|
| Classify and assign to the responsible department | Approve or release |
| Extract metadata such as sender, date or reference numbers | Post or pay |
| Create the task in exactly one existing target system | Write into a second system |
| Hand unsure cases to a named person, with its reasoning | Delete, set deadlines or decide on the matter |
AI agents for organisations: a Vorgang fits when it meets three conditions
Not every task suits an AI agent. A Vorgang fits when it meets three conditions:
- It arrives every day, not just now and then.
- Every item raises the same question, such as: which department is responsible?
- The result lands in a system that already exists.
A second intake channel, a second decision rule or a second target system turns it into a second Vorgang with its own hit rate. Multi-stage processes with branches are out of scope. That narrowness is what makes the result measurable.
We build the agent in conversation and operate it afterwards
The same sequence applies to each of the four examples. In week one, your Fachbereich (the department that owns the Vorgang) answers our questions about it, names classes and examples and clarifies access with us. In week two, we build and train the agent and test it on real documents. In week three, it is accepted against the agreed hit rate and goes live. After that, we operate it: we monitor, correct and take responsibility.
Every assignment is recorded with timestamp and reasoning in the decision log. When the agent stops is governed by the confidence threshold. What operation involves in detail is covered in our guide to AI agent governance. An agent built and operated this way is what we call one of our AI employees.
Frequently asked questions
What is an example of an AI agent in an office?
A typical example is incoming post: the agent reads every incoming item, assigns it to the responsible department and creates the task in the DMS. When it is unsure, it hands the case to a named person. Other examples are citizen enquiries, delivery notes and incoming invoices. What they share is one intake, one task and one target system.
Does the AI agent make decisions on the matter?
No. It assigns, extracts details, routes and creates a task. It approves nothing, posts nothing, pays nothing, deletes nothing and sets no deadlines. Below the agreed confidence threshold, it stops and hands the case to a named person with its reasoning. Every assignment is recorded with timestamp and reasoning in the decision log.
Which systems can the agent connect to?
Exactly one existing target system per Vorgang, such as a DMS, an ERP or a ticket system. Access runs through an interface that already exists: a standard API, a watch folder, sFTP or email. Custom development against a legacy system is not part of it. We agree the right interface with your IT in the first week.
Can we start with several examples at once?
A Pilotbetrieb, our entry engagement, covers exactly one Vorgang, because only then can the result be measured with one number: the hit rate agreed before the start. Every further Vorgang is a separate engagement with its own hit rate. That keeps each agent measurable. Whether and how to continue after the written evaluation is your call.
Is an AI agent the same as a chatbot?
No. A chatbot is a conversational interface that answers questions. The agents in these examples hold no conversations. They work in the background on one Vorgang: they read incoming documents, assign them and create tasks in the target system. When an agent is unsure, it hands the case to a person instead of guessing.
The next step
If one of these four examples is done by hand in your organisation every day, that is the Vorgang for a Pilotbetrieb: one intake, one task, one destination. Fixed price, agreed before we start. Seven weeks, four of them productive operation on your real documents.