AI process automation in SMEs: start with one Vorgang
Contents
AI process automation in an SME means an AI employee takes over the daily reading and sorting of documents that someone types into the ERP by hand today. It classifies each document, extracts the details, creates a task in the existing system and hands unsure cases to a person. Start with one Vorgang. Five steps get you there:
- Name the bottleneck: where do delivery notes, incoming invoices, orders or complaints pile up because someone reads and retypes each one?
- Cut one Vorgang: one intake, one task, one destination, for example every delivery note that arrives by email and ends up in the ERP. Vorgang is our term for exactly this unit, and we keep it in German.
- Answer questions: your Fachbereich, the department that owns the work, names the classes in one conversation and shows examples. Building and running the AI employee is our job.
- Agree a hit rate: before the start, as the bar for acceptance in week three.
- Hand over the work: we build the AI employee and operate it: we monitor, correct and take responsibility.
Full order books are a capacity problem
Full order books help little when there are not enough hands to work through them. According to the DIHK Skilled Labour Report 2025/2026, more than a third of German companies still cannot fill vacancies, at least in part; among firms with more than 20 employees it is over 40 percent. Of the consequences firms expect, a heavier workload for existing staff ranks second, after rising labour costs.
That workload reaches the office too: in the sales back office that keys orders into the ERP, at goods-in where delivery notes are checked, in accounts where invoices are sorted. The work is necessary, but it rarely calls for judgement. It calls for time.
AI process automation starts exactly there. It does not create demand and it does not replace a skilled worker. It takes over the time a person spends reading and retyping, so that person can deal with the cases that need experience.
AI takes over the reading; your people keep the decisions
Classic automation follows fixed rules: if a field sits in a certain place, it is copied. That breaks as soon as every supplier sends a different form. An AI employee instead recognises from the content what a document is, who sent it and which details it contains, even when the layout changes.
What it does with each document is narrowly bounded:
- It assigns the document to a class, such as delivery note, invoice or complaint.
- It extracts the details the target system needs, such as supplier, order number, date and line items.
- It routes the document into exactly one target system, usually the ERP or the inventory system.
- It creates a task there for the responsible person or team.
- It records every assignment with timestamp and reasoning.
What it does not do: post, approve, pay, delete or trigger deadlines. Those steps stay with your people, and that is what makes the automation hold up in daily work.
One Vorgang is the smallest unit you can measure
A Vorgang is everything that arrives daily, answers the same question and lands in the same system. Every delivery note that arrives by email and reaches goods-in as a task in the ERP is one Vorgang. Delivery notes and invoices together are two.
The limit has a practical reason: only one Vorgang can be measured with a single number, the hit rate. It states the share of documents the AI employee assigns correctly. Automating three intakes at once gives you three sources of error and no clear answer to whether it works.
A second intake channel or a second decision rule therefore becomes a Vorgang of its own, with its own hit rate and its own Pilotbetrieb, our fixed-price entry engagement. After each step you know what it achieved.
A good first Vorgang arrives daily and ends in the ERP
Five questions show whether a Vorgang is a good place to start. The more often you answer yes, the better.
| Question | Why it matters |
|---|---|
| Does the document arrive every day, in noticeable volume? | Only volume makes the relief measurable |
| Does every document answer the same question? | One question means one hit rate |
| Does it always land in the same system? | Exactly one target system keeps the Vorgang simple |
| Is there a manageable list of classes with examples? | We use them to build and train the AI employee in week two |
| Does someone read every document by hand today? | Then the Vorgang costs time that can be freed up |
Multi-stage processes with branching are not a good start. Complaints are a good example: reading a complaint, assigning it and passing it to quality assurance as a task is a Vorgang. Handling it afterwards, with queries, credit notes and replacement deliveries, is not.
Four Vorgänge share the same shape in SMEs
Whether at goods-in, in accounts or in the sales back office, many paper flows in an SME follow the same pattern. A document arrives, someone reads it, copies details and passes it on to whoever is responsible.
| Vorgang | What arrives | What the AI employee does | Target system |
|---|---|---|---|
| Delivery note | Paper or PDF from the supplier | reads the details, checks them and forwards the note to the warehouse, quality assurance or accounts | ERP or inventory system |
| Incoming invoice | PDF by email | captures the header data, proposes the assignment and hands the invoice to accounts | ERP or accounting system |
| Incoming order | Order by email or fax | recognises customer, items and requested date and creates a task for the sales back office | ERP |
| Complaint | Letter or email from a customer | assigns the complaint, extracts order and item numbers and creates a task for quality assurance | ERP or ticket system |
How the delivery-note Vorgang runs step by step is covered in our piece on delivery note processing. Invoices have two pieces: invoice capture explains how header data is captured and assignments are proposed. Invoice processing automation draws the line between what the AI employee takes over and what accounts keeps.
More examples from public administration and SMEs are collected in AI agent examples. How documents are classified and read is covered in our piece on how to automate document processing.
An order arrives as a task in the ERP, not as a PDF in an inbox
Incoming orders show how a Vorgang runs day to day. Today someone in the sales back office opens every order, looks up the customer in the ERP and types in items, quantities and dates. With an AI employee, the same flow looks like this:
- The order arrives as a PDF by email in the mailbox the Vorgang covers.
- The AI employee recognises that it is an order, not an enquiry or a complaint.
- It extracts customer, customer number, item numbers, quantities and requested date.
- It creates a task in the ERP for the sales back office, with those details as metadata and the original attached.
- It records the assignment with timestamp and reasoning.
If the order carries an item number it cannot match, its certainty falls below the threshold. It then hands the order to a named person in the sales back office. Confirming the order, checking prices and committing to delivery dates stay with the sales back office in every case.
The boundary is fixed before the Pilotbetrieb starts
What an AI employee may do is set before the Pilotbetrieb begins, not negotiated during operation. The boundary applies to every Vorgang:
| Area | We do this | We do not do this |
|---|---|---|
| Scope | One intake, one task, one destination | A second intake channel or decision rule |
| Term | Fixed, defined end | Open-ended or auto-renewing |
| Target system | Exactly one | Multiple targets, or writing back into a second |
| Action | Classify, extract metadata, route, create a task | Post, approve, pay, delete, trigger deadlines |
| Integration | Existing interface, watch folder, sFTP, email, standard API | Custom development against a legacy system |
| Workflow | Single-stage | Multi-stage with branching |
For you, the boundary means two things. The AI employee never posts anything and never changes stock levels. And the work does not turn into an IT project: we use the access your ERP already offers.
Your business answers questions; we build and operate
The work is clearly divided. The only change is who does the reading and sorting.
| Who | Role in the Vorgang |
|---|---|
| Management | chooses the Vorgang and agrees the hit rate |
| Fachbereich, such as goods-in, accounts or the sales back office | answers the process questions in week one, shows examples and resolves the cases the AI employee hands over |
| AI employee | classifies, extracts, routes, creates tasks and hands unsure cases to a person |
| IT service provider or ERP partner | provides access that already exists |
| Agentivo | builds the AI employee in the first two weeks and then operates it: monitors, corrects, takes responsibility |
You do not need an IT department of your own. Access runs through an interface that already exists; custom development against a legacy system is not part of it. After go-live we keep operating the AI employee: we monitor it, correct it and answer for the result. What that involves is described in our piece on AI agent governance.
Unsure cases go to a person
An AI employee is not equally certain about every document. A smudged delivery note, a new supplier, an invoice without an order number: in such cases its certainty falls below the agreed confidence threshold. It then stops and hands the case to a named person in your Fachbereich instead of guessing.
Every assignment, and every handover to a person, is recorded with timestamp and reasoning in the decision log. Anyone who later wants to know why a delivery note went to quality assurance can look it up there. For a business going through a handover, that is more than control: it is a written record of how the Vorgang actually runs.
Every operation also 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: your team can stop the AI employee, correct it or take over the Vorgang.
From first conversation to written evaluation in seven weeks
The Pilotbetrieb covers one Vorgang at a fixed price, agreed before we start. It takes seven weeks, four of them productive operation on your real documents.
| Week | What happens |
|---|---|
| 1 | Process conversation with the Fachbereich, define classes, review examples, clarify access |
| 2 | We build and train the AI employee, set up quality checks and test on real documents |
| 3 | Acceptance against the agreed hit rate, go-live, handover |
| 4 to 7 | Productive operation on real volume; we monitor and correct |
| End of week 7 | Written evaluation: hit rate achieved, cases processed, time saved |
The hit rate is agreed before the start. If it 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.
The first use case is the hurdle, not the technology
In German SMEs, AI is still the exception. According to KfW Research, 2026 (in German), one in five uses at least one AI technology. Text recognition, used for example to pre-sort incoming email or to process invoices, reaches around one in ten. Technologies that automate workflows are used by only about 5 percent; the study attributes this to their having to be tailored to the individual firm and integrated into its IT. The study sees finding first, practical applications as a central hurdle.
KfW Research (in German) names the main obstacles: a lack of skilled staff and know-how, limited time, missing data and doubts about the reliability of AI. A Pilotbetrieb addresses each of them. Your Fachbereich contributes one process conversation and examples, not a project team. The data are the documents that arrive every day anyway. The know-how for building and operating comes from us. And reliability is not promised; it is tested in week three against an agreed hit rate.
After the first Vorgang comes the next, not a platform
Once the first Vorgang runs, the second is often close at hand: first the delivery notes, then incoming invoices, then incoming orders. Each gets its own AI employee, its own hit rate and its own acceptance. Automation grows one Vorgang at a time, and every step has a number it is measured against.
In firms facing a handover, that sequence is an advantage. The next generation can show with one Vorgang what changes, without rebuilding the whole operation at once. Which Vorgänge suit that first step is described in our piece on AI in SMEs.
Frequently asked questions
How is AI process automation different from classic automation?
Classic automation follows fixed rules and stops when a document looks different from what it expects. An AI employee recognises from the content what a document is and which details it contains, even when suppliers use different layouts. It classifies, extracts, routes into one target system and creates a task. When it is unsure, it hands the case to a person instead of following a rule blindly.
Do we need our own IT department?
No. Your Fachbereich answers questions about the process in week one and shows examples. We do the building and the operating. The AI employee uses access your ERP or inventory system already offers: a standard API, a watch folder, sFTP or email. Custom development against a legacy system does not belong in a Pilotbetrieb. If you work with an IT service provider, they only set up that access.
Which Vorgang is a good place to start?
One that arrives every day in noticeable volume, answers the same question and lands in the same system. Typical candidates are delivery notes, incoming invoices, orders arriving by email or incoming complaints. What matters is a manageable list of classes with examples. Multi-stage processes with branching, such as full complaint handling, are not a good first step.
What may the AI employee do with an invoice?
It captures the header data, proposes an assignment and hands the invoice to accounts. Posting, approving, paying and deleting stay with your people, as does triggering deadlines. That is part of the boundary fixed before the Pilotbetrieb. Every assignment is recorded with timestamp and reasoning in the decision log, so accounts can trace every proposal.
What happens after the seven weeks?
If the agreed hit rate is met, operation begins under a monthly operating agreement that can be cancelled monthly. We keep monitoring the AI employee, correct it when suppliers or processes change and take responsibility for the result. If the rate is missed, the Pilotbetrieb ends with no further obligation. The price is settled beforehand either way: fixed price, agreed before we start.
The next step
If delivery notes, invoices or orders are typed into your ERP by hand 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.