Document processing7 min read

How to automate document processing with AI: one Vorgang in six steps

Contents

To automate document processing with AI means an agent takes over reading, assigning and forwarding the documents of one intake that arrives every day. It creates the task in the system you already use and hands every unsure case to a person. Six steps get you there:

  1. Choose one Vorgang: one intake, one task, one destination, such as the delivery notes someone passes by hand to the warehouse, quality assurance or accounting today.
  2. Name the intake: the email inbox, scan folder or upload through which the documents arrive.
  3. Write down classes and metadata: where a document can go and which details the task should carry, with real examples for each class.
  4. Fix the target system: exactly one system, such as the ERP, DMS or ticket system, through an interface that already exists.
  5. Agree the hit rate and the confidence threshold: how many cases the agent should assign correctly, and below which certainty a person takes over.
  6. Hand the operation to us: we build the agent, test it against the hit rate at acceptance and then operate it.

Steps one to five are questions your Fachbereich (the department that owns the work) settles with us: the hit rate before the start, everything else in the first week. Building and operating is our job. What intelligent document processing covers as a whole is in the overview; this article is about automating a single Vorgang.

What gets automated is the assignment, not the decision on the matter

Automated document processing replaces the manual work between intake and casework: reading, recognising, retyping, forwarding. What comes after stays with the department.

AutomatedStays with a person
Collecting the document from the intakeAssigning cases below the confidence threshold
Recognising the document type or who is responsibleJudging the matter itself
Extracting details as metadataPosting, approving, paying
Creating the task in the target systemSetting deadlines, deleting documents
Logging the assignment with its reasoningStopping or correcting the agent at any time

This split makes the result measurable and the responsibility clear. We answer for the assignment, your department for the matter. How the classes used for assignment come about is covered in the article on document classification.

Structured data needs no retyping; paper and PDF still do

Some documents already arrive as data. In Germany, since 1 January 2025, an e-invoice is only an invoice in a structured electronic format; a simple PDF does not count. Between domestic businesses it becomes the rule, with transition periods until the end of 2026, and until the end of 2027 for issuers with a smaller prior-year turnover. Until then, invoices may still arrive on paper or, if the recipient agrees, as a PDF, according to the Federal Ministry of Finance (in German).

For automation, that means nobody needs to retype the details of an e-invoice; they are already in fields. Letters, delivery notes, complaints, applications and invoices on paper or as PDFs have no fields. That is where the agent comes in: it reads what arrives as an image or free text and turns it into a task with metadata.

The agent connects to the system that is already running

Among German enterprises that have considered AI but not adopted it, almost half cite incompatibility with existing equipment, software and systems as a reason, according to the Federal Statistical Office (Destatis), 2025.

That is why the agent writes into exactly one existing target system, through access that already exists: a standard API, a watch folder, sFTP or email. Custom development against a legacy system does not belong in a Pilotbetrieb, our fixed-price entry engagement. A second target system becomes a second Vorgang, not an extension of the first.

One delivery note, one pass: what an automated case looks like

An example from an SME. Delivery notes arrive as PDFs by email or as scans from goods receipt. Today someone reads each one, retypes numbers and passes it to the warehouse, quality assurance or accounting. Automated, the same case runs like this:

  1. The delivery note arrives in the agreed inbox, and the agent collects it.
  2. It extracts supplier, delivery note number, order number and date, and checks the details.
  3. It weighs the classes warehouse, quality assurance and accounting and rates how certain it is for each.
  4. If the highest certainty is above the confidence threshold, it creates the task for the chosen team in the inventory system, with the extracted details as metadata.
  5. If it is below, it creates no task and hands the delivery note to the named person in goods receipt.
  6. Either way, it writes the assignment with timestamp and reasoning to the decision log.

The article on delivery note processing describes reading, checking and forwarding delivery notes in detail.

The log shows for every case why it landed where it did

Every assignment leaves an entry. For the delivery note in the example, it looks like this:

FieldWhat the entry records
Intakethe delivery note as a PDF from the goods receipt inbox, with time of arrival
Classes weighedwarehouse, quality assurance, accounting, each with the agent's certainty
Assignmentwarehouse
Confidence and thresholdthe certainty of this assignment and the threshold in force that day
Basisin the department's words: which sender, which details and which class examples tipped the balance
Human interventionnone, or who corrected what and when
Follow-up actionthe task in the inventory system with its identifier

An entry like this answers the question asked after a mistake: what arrived, what was weighed, how certain the assignment was, and whether a person intervened. Which fields a complete entry needs is set out in the article on the AI audit trail.

Below the confidence threshold, a person takes over

Automation without a fallback is unsupervised automation. That is why every agent has a confidence threshold: below that certainty it creates no task but hands the case to a named person, with the classes it weighed and the reason for its doubt.

The threshold is agreed with the department. Set too high, almost everything lands on a desk and little time is saved. Set too low, mistakes surface later. The right value is the one the department can defend to management. A named person can also stop the agent at any time through the human intervention path.

Seven weeks, four of them productive operation

The entry point is the Pilotbetrieb: one Vorgang, fixed price, agreed before we start.

WeekWhat happens
1Process conversation with the department, define classes, review examples, clarify access
2We build and train the agent, set up quality checks and test on real documents
3Acceptance against the agreed hit rate, go-live, handover
4 to 7Productive operation on real volume; we monitor and correct
End of week 7Written evaluation: hit rate achieved, cases processed, time saved

If the agreed hit rate is met, operation begins under a monthly operating agreement that can be cancelled monthly. If it is missed, the Pilotbetrieb ends with no further obligation. What we build and run is one of our AI employees, for exactly this one Vorgang.

Frequently asked questions

Which documents are a good place to start?

Documents that arrive every day, raise the same question and land in the same system. Typical starting points are delivery notes, incoming post, complaints, or invoices on paper and as PDFs. Incoming post may first pass through a central Poststelle (the mailroom). What matters is a manageable list of classes with real examples and one target system for the task. Rare special cases do not belong in the first Vorgang.

Do we have to set anything up or train anything ourselves?

No. In the first week your department answers our questions: what arrives, where it belongs, who decides that today and what happens in doubt. Your IT provides access to the target system. We build, train, test and operate the agent. After go-live we monitor and correct it and take responsibility for the result.

What happens when suppliers or forms change?

The agent's certainty on those documents drops. We see that in operation, because more cases fall below the confidence threshold or assignments get corrected. We then correct the agent with the new examples. Until then, the unsure cases go to the named person, and every handover is recorded in the decision log.

How accurate does the assignment have to be?

As accurate as the agreed hit rate: the share of cases the agent assigns correctly. It is agreed before the start and tested at acceptance in week three on real documents. Cases where the agent is unsure do not count as errors; they go to a person. The confidence threshold controls how many of those there are.

Can we automate more documents later?

Yes, as a separate Vorgang. A second intake channel, a second target system or a second decision rule is not bolted onto the first Vorgang but set up as a new one with its own hit rate. That keeps every Vorgang measurable, and every acceptance depends on a single number agreed before the start.

The next step

If the same kind of document arrives every day and someone reads it, retypes it and passes it on, 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.

Request a Pilotbetrieb

← Back to insights