Abstract
Teams lose time moving digital assets between systems. To place a product image intoa slide, a user has to leave PowerPoint, search a separate Digital Asset Management
(DAM) web interface, download the file and paste it back. This thesis investigates how
an agent inside Microsoft 365 Copilot can remove that detour and let a user find and
insert an asset in a single conversational turn. The concrete case is the integration of
CELUM Content Hub into PowerPoint Copilot.
The Copilot platform offers no single prescribed way to build such an agent. This
thesis compares four candidate architectures: a Declarative Agent with an API Plugin
(DA+API), a Declarative Agent with an MCP Server (DA+MCP), a Custom Engine
Agent (CEA) and a Copilot Studio Agent with an MCP Server (CS+MCP). Each can
didate is scored against seven evaluation dimensions: development complexity, OAuth
integration effort, Adaptive Card support, image insertion capability, maintainability,
platform dependency and flexibility.
DA+MCP is selected as the most suitable architecture. The decisive factor is image
insertion. Only the two manifest-driven candidates can use the Action.InsertImage
Adaptive Card action, which places an image straight onto the slide. Between those two,
DA+MCP wins because its typed MCP tool surface makes agent-built requests more
reliable, keeps request guidance out of the tight instruction budget and lets one backend
serve more than one Copilot agent.
The selected system is then designed, built and evaluated. It consists of a Declarative
Agent manifest package on the Copilot side and a Spring Boot MCP server on the
backend side, deployed on Azure Container Apps. The user-facing OAuth 2.0 flow runs
against Azure AD B2C and every call to Content Hub is made with the user’s own
access token, so a user only sees assets they are already allowed to see.
The technical evaluation shows that the implemented system meets all nine func
tional and non-functional requirements. Two of them carry a caveat because they depend
on Action.InsertImage, which is rendered by the Copilot host and can change without
notice. Latency measurements show that Content Hub accounts for about 96 percent of
the user-perceived wait and the MCP server itself adds only milliseconds of overhead.
The work has clear limits. The evaluation is technical and includes no user study.
The cross-architecture comparison rests on an uneven base, because only DA+MCP
was built end to end while the other three remained PoCs. All platform observations
are point-in-time readings taken during the implementation period, because the Copilot
extensibility surface evolves on a near-weekly cadence.
| Date of Award | 2026 |
|---|---|
| Original language | English (American) |
| Awarding Institution |
|
| Supervisor | Andreas Stöckl (Supervisor) & Rene Loitzenbauer (Supervisor) |
Studyprogram
- Software Engineering
Cite this
- Standard