Most companies think of a chat when they think of AI. Bitkom 2025 shows how early
the field still is: only 2 % of companies use AI in IT-supported production,
and only 26 % give their employees access to AI at all. Anyone who just hands
out a chat window leaves the largest part on the table. In our AI adoption work,
we see four use-case classes that keep appearing.
In sheer usage, chat dominates — at one mid-sized company we measured, it was
over 70 % of all interactions. The economic value per use case, however, sits in
workflows and AI apps. A single workflow often delivers more in a month than
hundreds of chat requests.
Building in only one class means losing a large part of the potential.
Teams confuse the two middle classes most often. The dividing line is simple: an
assistant is flexible — the model decides depending on the request, calls
tools itself and returns Markdown or artifacts. A workflow has fixed steps,
typed variables and its own model per step. File output in PowerPoint, Excel or
Word goes exclusively through a workflow.
Same flow, structured inputs → workflow.
Open question, the model uses tools on its own → assistant.
Badly built workflows deliver a good result on the first test and fail silently on
the tenth run, when a different person enters different inputs. Two rules prevent
that.
Manually first, then model it. Run the process through the chat three times —
with three different inputs — before you save it as a workflow. You notice in the
process which variables really vary and whether the model delivers what you need.
Three components per step. Every good step names at least one variable, one
reference and one output schema. Three to four steps at most — beyond that the
model loses the original task.
Anti-pattern: output without a schema
The most common mistake in the workflow builder: the prompt says what should be
produced, but not in what form. Models then vary from run to run in length,
order and bullet style — and a colleague gets a different format a week later.
Write the output schema into the prompt explicitly, as a list of sections. Where
possible, enable document output — a .docx or .xlsx enforces a fixed
structure.
For AI to take on real work, it has to reach your systems. Two routes,
deliberately separated by risk:
Connector: a live API with actions — the assistant reads current data and
writes back (send emails, create tickets). Permissions match those of the
individual user in the source system.
Data pool (RAG): read-only knowledge search across large document sets,
maintained centrally by admins.
Careful with data-pool permissions
A data pool does not automatically inherit the permissions of the source.
Unlike a connector, SharePoint permissions for instance do not apply by
themselves — the admin configures access manually. Settle that before sensitive
knowledge is indexed.
The numbers from real introductions show the difference between a toy and a
production system — all anonymised.
Task
Before
With AI
Job ad (workflow)
20 min
2 min
HVAC quote (assistant)
45 min
5 min
Sales dashboard (AI app)
weeks
30 min
Requirement spec, first version
3–5 days
60 % in one afternoon
Saving on the HVAC quote
80–90 %
The sales dashboard was built by an employee with no programming skills — the
classic route would have needed a business analyst and several weeks. This is
exactly where AI tips from a helper into a process.
Start read-only: search and read first, then write and act. How to govern write
permissions and data access cleanly is in governance & security.
Which tasks in your team offer the greatest leverage is shown by the
use-case map — and what breadth across all four classes is
really worth is calculated by measuring impact.
ISO Certified
GDPR Compliant
EU Hosting
Start with AI in your company
Together we find the right use cases, connect your systems, and bring AI into daily work in line with your business.