Process Automation for SMEs: Three Processes to Start With
By Dennis Weidner · 11. November 2026
Process automation means that a recurring workflow runs without manual work: data is captured, checked, passed on and filed, and people only step in for exceptions. For mid-sized companies this is not a new topic. What is new is that generative AI now makes it possible to automate workflows that used to fail on unstructured text: emails, PDF invoices, delivery notes, tender documents. This guide shows which method fits which process, which three processes are the best starting point, and how to set up a first pilot that ends with a clear decision. As of November 2026.
What is process automation?
Process automation is software that carries out a business process fully or partly on its own. Three families of tools belong to it: interfaces and workflow software for rule-based processes, robotic process automation (RPA) for operating applications without an interface, and AI agents for tasks where content has to be read, understood and assessed.
The difference to digitisation: digitisation turns paper into a file. Automation makes sure something happens with that file without anyone opening it. Our page on AI automation for business covers services, process and cost.
Where German mid-sized companies stand
Official statistics show how fast things are moving. According to the Federal Statistical Office (Destatis), 12 percent of German companies with ten or more employees used artificial intelligence in 2023, and 20 percent in 2024. The 2025 ICT survey shows 26 percent; from that year Destatis counts by the EU definition of an enterprise, so the figures are not fully comparable, but the trend is clear.
Among companies using AI in 2025, text mining and the generation of images, video and audio lead equally (52 percent each), followed by speech recognition (42 percent) and natural language generation including code (35 percent); only 27 percent use AI to automate workflows or decisions. By business function, marketing and sales led in 2024 (33 percent), followed by production and service processes (25 percent), administration (24 percent) and accounting, controlling and finance (24 percent). Much of it is about reading, sorting and writing text, while only about a quarter actually automates workflows. That gap is where the new automation potential lies.
Types of process automation
| Type | What it does | Limit |
|---|---|---|
| Workflow automation | connects applications through interfaces with fixed rules | needs structured data |
| Business process automation (BPA) | automates end-to-end processes across systems and departments | needs well-defined processes |
| Robotic process automation (RPA) | software bots operate screens like employees | breaks when screens change |
| Intelligent process automation (IPA) | combines workflows and RPA with AI that reads, classifies and prepares decisions | needs approvals and checkpoints |
Rules, RPA or AI: which kind of automation fits which process
| Type of process | Example | Suitable method |
|---|---|---|
| Fixed rules, structured data | Transfer an online order into the ERP system | Interface or workflow software |
| Fixed rules, operated through screens | Re-type data from a legacy system without an interface | RPA |
| Changing content, judgement needed | Read an enquiry, classify it and answer it with data from your own systems | AI agent with human approval |
| Mixed | Read an invoice (AI), check it (rules), book it (interface) | Combination, the agent orchestrates the flow |
In practice the methods are almost always combined. Where exactly RPA ends and agentic AI begins is explained in AI agents vs. RPA.
Three processes to start with
1. Capture and pre-code incoming documents
Invoices, delivery notes and order confirmations arrive in many formats and are often still typed in by hand. An agent reads them, matches them against purchase orders and supplier master data and prepares the booking. Approval stays with a person. See AI invoice processing.
2. Sort and answer enquiries
Many enquiries by email or contact form repeat themselves: delivery status, documents, appointments. An agent assigns each one to a topic, answers standard cases with data from your systems and hands everything else to the right person with a summary. See automating customer service with AI.
3. Monitor sources and report
Tender portals, funding databases, supplier price lists: many teams check the same sources by hand. An agent screens them daily, scores hits against fixed criteria and delivers a short overview with deadlines. See tender and funding monitoring.
Worked example: invoice processing
Assumptions: 600 incoming invoices a month, six minutes of manual work each, 45 euros fully loaded cost per hour. After automation four out of five invoices pass with one minute of review, every fifth remains an exception with the previous effort.
| Item | Before | After |
|---|---|---|
| Standard cases | 600 × 6 minutes | 480 × 1 minute |
| Exceptions | included above | 120 × 6 minutes |
| Hours per month | 60 | 20 |
| Hours per year | 720 | 240 |
| Cost per year at 45 euros per hour | 32,400 euros | 10,800 euros |
In this example 480 hours a year are freed up, roughly 21,600 euros, before the cost of building and running the agent. How those costs break down is covered in how much does an AI agent cost.
How to set up the first pilot
- Pick one process, not three. The first pilot should answer one question.
- Measure the baseline: volume per month, time per case, number of queries.
- Define approval: which results does a person check before they take effect?
- Agree a stop criterion: what must be achieved for go-live, and what ends the test?
- Involve the team early: the people doing the work today know the exceptions.
In our projects first measurable effects usually show after four to six weeks, and regular operation is in place after about twelve weeks. Why some projects still stall is covered in moving AI pilots into production.
Why process automation fails, and how to avoid it
Technology is rarely the bottleneck. More often it is undocumented exceptions, master data maintained differently in three systems, a missing baseline, and a team that experiences the new flow as control rather than relief. Knowledge helps most against the last point; see our AI training for teams.
Checklist: is your process ready?
- The workflow occurs at least a few hundred times a year.
- Someone is responsible for the result today and can check it.
- Input arrives digitally, as email, PDF or data record.
- The main exceptions can be written down in a list.
- You know the current effort per case, at least as an estimate.
- Errors are caught before they take effect, for example through approval.
If four or more points apply, the workflow is a good candidate. We are happy to go through it with you in a short intro call.
Transparency note: this article was produced with the help of artificial intelligence.