Document classification: assigning post at department level
Contents
Document classification here means assigning every incoming item to a document type or to the responsible department, so that it arrives there as a task. It does not mean grading information by confidentiality, as information security under the BSI's IT-Grundschutz or ISO/IEC 27001 and the protection of classified material do. The two meanings side by side:
| Question | Classification by document type | Grading by confidentiality |
|---|---|---|
| What is decided? | what kind of item this is and which department is responsible | how much protection the information needs |
| Typical classes | planning application, objection, invoice, citizen enquiry, or the departments themselves | normal, high, very high; open, internal, confidential, secret; or grades such as VS-NUR FÜR DEN DIENSTGEBRAUCH |
| What for? | so that the item arrives as a task with the responsible department | so that access and safeguards match the protection need |
| Who sets the classes? | the department, following the organisation's responsibilities | information security management, or the classifying authority for classified material |
Information security asks what damage follows if confidentiality, integrity or availability is breached, and groups protection needs into three categories, "normal", "high" and "very high", according to BSI-Standard 200-2 of the Federal Office for Information Security, which notes that some organisations grade confidentiality as open, internal, confidential and secret. For classified material of the German federal government, section 4 of the Security Clearance Act (SÜG, German text) sets four grades, from VS-NUR FÜR DEN DIENSTGEBRAUCH (restricted) to STRENG GEHEIM (top secret). This article is about the other question: where an item belongs. That is AI document classification at department level.
The classes are your organisation's responsibilities
For the incoming post of a public administration, the deciding question is: which Fachbereich (the department that owns the work) is responsible? That is why an AI employee assigns post at department level, not against a generic catalogue of document types.
The classes come from your organisation, not from a model. In the first week, we sit down with the department and the Poststelle (the mailroom) and write down which departments or case types exist, how post for each can be recognised and which examples are typical. The starting point is what the Poststelle already knows today, for example along the allocation of responsibilities or the filing plan.
A good list of classes is short enough for every class to have real examples, and precise enough for every item to fall into exactly one class. Post that fits no class goes to a named person.
Classification starts once the post is digital
The AI employee reads files, not envelopes. Opening, scanning and checking completeness stay in the Poststelle. For the scanning itself, the BSI's Technical Guideline TR-03138 (TR-RESISCAN) sets out the technical and organisational measures for replacement scanning. For federal authorities, section 7 of the E-Government Act (EGovG, German text) requires the electronic document to match the paper in image and content according to the state of the art; state and municipal authorities follow the law of their federal state.
What has to exist before classification is covered in the article on the digital mailroom: a digital intake, a list of classes, a target system and a rule for cases of doubt.
Every item goes through the same five steps
For every item, the AI employee follows the same sequence:
- It reads the scanned item and recognises sender, date, subject and, where present, the file reference.
- It weighs the agreed classes and rates how certain it is for each.
- If the highest certainty is above the confidence threshold, it assigns the item to the department and creates the task in the DMS, with the recognised details as metadata.
- If it is below, it creates no task and hands the item to the named person.
- It records every assignment and every handover with timestamp and reasoning in the decision log.
It does not decide on the matter, set deadlines or delete anything. The item arrives at the department with its metadata; what follows from it is handled by the people there.
Below the confidence threshold, a person assigns the item
The confidence threshold is the certainty below which the AI employee does not assign but hands the case to a person. It is agreed with the department before the Pilotbetrieb, our fixed-price entry engagement, starts.
The named person can sit in the Poststelle, or each department can have its own. They receive the item, the classes weighed with their certainty, the class the AI employee would have chosen and the reason for its doubt. Their assignment is recorded in the decision log and becomes an example for the next correction.
The other way round, a named person can stop the AI employee at any time through the human intervention path, correct an assignment or take over a case. This path is documented before go-live.
Automated document classification is checked against one hit rate
Whether the assignment works is shown by one number: the hit rate, the share of items the AI employee assigns to the right department. It is agreed before the start and tested at acceptance in week three, on real documents.
For that number to mean something, a Vorgang (one intake, one task, one destination) is single-stage. Assignment into two systems, or with branches, would leave no single number on which acceptance and operation can rest.
Invoices and delivery notes become Vorgänge of their own
Not every item belongs in the same Vorgang. An invoice for the finance office needs different details from a letter to the building authority: invoice number and header data instead of a file reference. A delivery note for the municipal depot is read, checked and forwarded.
In the incoming post, the AI employee recognises such items as a class of their own and assigns them to the responsible department. Their further processing later becomes a separate Vorgang with its own hit rate. The articles on delivery note processing and invoice capture show what that looks like.
How to automate a Vorgang step by step is covered in the article on how to automate document processing. The overview of every step from intake to task is in the article on intelligent document processing.
Traceable to leadership, the finance office and auditors
For a public administration, it matters not only whether the assignment is right but whether it can be explained later. That is why every operation includes:
- Operation entirely within the EU, responsibility from Düsseldorf.
- GDPR-compliant processing with an AVV (data processing agreement) included.
- A complete decision log available as evidence for EU AI Act audits.
- A defined human intervention path, documented and triggerable at any time.
Which post belongs in the Vorgang and which stays out, such as personnel matters or letters graded as classified, is for your organisation to set, for example in an internal instruction. We provide the AVV and the log for it.
Frequently asked questions
Is document classification the same as grading by protection need?
No. Grading by protection need or classification level asks how much protection a piece of information needs. Document classification in this article asks what kind of item it is and which department works on it. The two exist side by side: an item has a responsibility and a protection need. The AI employee assigns only the responsibility; it does not change the grading.
Who sets the classes?
Your department together with the Poststelle, in the first week of the Pilotbetrieb. We ask what arrives, where it belongs, who assigns it today and what happens in doubt, and we review real examples for each class with you. If responsibilities change later, for example after a reorganisation of departments, we adjust the classes during operation.
How many classes make sense?
As many as there are responsibilities that post is distributed to today, and few enough for every class to have real examples. A very fine-grained list makes the distinction hard and lowers certainty. One that is too coarse sends items to the wrong department. The right list comes out of the process conversation and is tested on real documents in week two.
What happens to post that fits no class?
It goes to a named person. If an item fits no class, the AI employee's certainty falls below the confidence threshold. It then creates no task but hands over the case with its reasoning and logs the handover. If such cases pile up, that points to a missing class, which we add together with the department.
Is this document classification software?
Not in the sense of a product your organisation installs and trains. We build an AI employee for one Vorgang and then operate it: we monitor it, correct it and take responsibility for the result. It writes into exactly one target system your organisation already runs, through existing access such as a standard API, a watch folder, sFTP or email, agreed with your IT or data centre in the first week.
The next step
If your Poststelle distributes incoming post to departments by hand today, that assignment 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.