A lot gets said about AI and little of it is concrete. When we talk to business owners, the same question comes up almost every time: which part of this can we actually use? The honest answer is shorter than most conference talks. Language models can reliably read, categorize and write. They cannot reliably act on their own. Everything that works in everyday office use builds on that distinction.

Where Austrian companies stand right now

According to Statistik Austria (ICT usage in enterprises, 2025), 30 percent of Austrian companies with ten or more employees use artificial intelligence; the EU average is 20 percent. In 2021 the Austrian figure was 9 percent. The other side of the number is just as telling: 77 percent of companies not using AI have not looked into it at all. And where AI was considered but dropped, the most common reason is missing in-house expertise; lack of benefit comes last. Starting now does not make you late. It puts you in the first half.

Companies using AI in Austria

20219%

202311%

202420%

202530%

Source: Statistik Austria / Eurostat, 2025

What an AI workflow is, and what it is not

An AI workflow is not a chatbot you shout tasks at. It is a fixed, defined process: an email arrives, gets read and categorized, the data lands in the right system, a draft reply is waiting. The AI takes over exactly those steps that used to require a person, because software could not understand free text. The rest remains ordinary automation with fixed rules.

This embedding is where many AI projects fail. A model that understands invoices is worthless if the extracted data still gets typed into the ERP by hand. In our projects, connecting the mailbox, the ERP and the document storage regularly takes more work than the AI part itself.

Example 1: The shared mailbox sorts itself

In many companies, the office@ mailbox is a mix of orders, inquiries, invoices, complaints and spam that somebody goes through every morning and forwards. A language model can take over this triage: it reads each email, recognizes the request, assigns it to the right person or category, and attaches a draft reply for standard inquiries.

The draft is not sent automatically. That is not a technical limitation but a deliberate choice: the time savings come from pre-sorting and pre-writing, not from cutting out the final human glance. Automating that away too saves seconds per email and risks an annoyed customer per mistake.

Example 2: Documents enter the system without retyping

Invoices, delivery notes and order confirmations arrive as PDFs, scans or phone photos, each laid out differently. Classic software needed a template per supplier. A language model reads the document the way a person would: it finds the supplier, amounts and line items even in a layout it has never seen.

The extracted data is checked against fixed rules, for example whether the line items add up to the total, or whether the purchase order number exists. Only then does it enter the system. Cases where the model is uncertain go to a person. With clean PDFs the recognition rate is very high in practice; with crooked phone photos of bad receipts it is noticeably lower. That is why the uncertainty branch belongs in every such workflow.

Example 3: A free-form inquiry becomes a draft quote

A customer writes three paragraphs describing roughly what they need. Until now, somebody read the email, wrote out the requirements, looked up prices and typed a quote. A workflow can extract the requirements, match them against the price list and produce a draft that builds on earlier, comparable quotes. If information is missing, the system suggests a follow-up question instead of guessing.

Here too, the draft waits for approval. A quote is a legally relevant commitment, and a language model has no sense of whether this customer has earned a discount or whether this job fits your current capacity.

What does not work reliably today

For completeness, the other side. Autonomous agents that carry out multi-step tasks in external systems on their own are impressive in demos and error-prone in operation: a mistake in step two carries through every following step, and nobody notices in time. Equally shaky are tasks where exact numbers matter and the model is supposed to calculate them itself. Let software that can calculate do the calculating; the AI reads and assigns.

And there are processes for which AI is simply the wrong tool. If a process follows clear rules and has structured data, ordinary automation is cheaper and, on top of that, deterministic: same input, same result. We regularly tell prospects that their case is not an AI project. That is part of honest consulting.

How to recognize a good first use case

In our experience, four questions help pick the first workflow:

Does the task occur frequently? A process that comes up ten times a day pays off faster than one per week.
Is the core of it reading, categorizing or writing? Those are the strengths of language models.
Can a mistake be found and corrected before it causes damage? Then approvals can do their job.
Does the necessary data exist digitally? A workflow built on paper slips in a shoebox starts with a different project.