Abstract
Employees increasingly turn to generative AI, including through private accounts onfreely available services, placing their inputs beyond the company’s control. Organiza
tions that want to respond with a managed AI chatbot platform face a fundamental
choice between buying an external solution, running a hybrid architecture and hosting
everything themselves. In practice, this decision is often made ad hoc and without an
established procedure, all while the EU AI Act and data protection law continue to
apply.
This thesis develops a transparent procedure for that decision, following a Design
Science Research approach. It begins with an analysis of the current state of knowl
edge and technology: why organizations turn to managed platforms (data sovereignty,
integration of internal knowledge, auditability, cost control), which deployment models
exist from cloud services to on-premises operation, and how turnkey SaaS products differ
from the component market of self-operated architectures. This foundation is completed
by the obligations of the AI Act, above all the division of roles between providers and
deployers, and the security risks of generative AI.
From this foundation and an empirical requirements elicitation, the thesis derives
its decision framework. At its core sits a decision tree that treats the admissibility
of external data processing as a hard entry criterion and only then weighs economic,
strategic and functional considerations. Security and compliance are deliberately not
selection criteria. They are downstream design layers of whichever strategy is chosen.
The framework was put to the test in a single case study at SWIETELSKY AG,
an international construction group with around 13,000 employees. There it points to
the hybrid strategy, reflecting the decision the company had actually made. The imple
mented platform keeps control, knowledge and identities inside the corporate network,
obtains model inference from external providers through a central AI gateway and serves
around 4,200 registered users.
The evaluation drew on six semi-structured expert interviews with independent prac
titioners, vendors of the core components and internal key people, validated through
member checking. All six endorsed giving data admissibility precedence over every other
driver and rated the framework’s effectiveness unanimously with the best possible score.
The most important gap, raised by five of them, concerns operability, that is, whether
an organization has the staff and expertise to run even a partially self-hosted platform
at all.
These findings have been incorporated into a revised version of the decision tree,
which now checks the project’s success criteria and its operability before entering the
decision logic itself. This revised version has yet to be validated. The thesis thus delivers a practice-tested decision instrument together with a documented reference implemen
tation and clearly marks the limits of its validity.
| Date of Award | 2026 |
|---|---|
| Original language | German (Austria) |
| Awarding Institution |
|
| Supervisor | Harald Lampesberger (Supervisor) |
Studyprogram
- Information Security Management
Cite this
- Standard