Context
Whoever works with SAP Analytics Cloud usually has the question ready: revenue by region, as bars, for the current quarter. Until that answer is on the screen, several steps are needed, each one somewhere else in the tool, some of them in Excel. Whoever does not do this every day asks someone who can.
If the result is then not only to be looked at but used further, back into planning and from there into the SAP system, more steps follow, again somewhere else and again with knowledge of their own.
A chat inside Analytics Cloud
The chat sits in Analytics Cloud itself, in the same interface as the open story. Users write in their own words what they need, and the chat carries out the step: an analysis, a chart, an import, a retraction. There is no switch into another tool, and technical knowledge is not required.
The backend behind it runs as an SAP CAP application on SAP BTP. That is where the orchestration and the connections into the systems live. The AI models, Claude Sonnet and Opus, come through AWS Bedrock from EU regions.
What the chat does
Asking. The chat reads the data of the open story, with the filters and variables that are set right now, and if needed the raw data of the model behind it. On that basis it spots trends, finds outliers, compares periods and gives recommendations.
Showing. Results come as a chart, as a table or as code. Charts are available as bars, lines, pie, waterfall, heatmap, flowchart and mind map, with zoom and full screen. Tables can be searched, sorted and filtered per column. SQL, JavaScript or JSON appear formatted, with a copy function.
Bringing in. An Excel file is uploaded, analysed and imported into the story. A screenshot is taken, cropped, marked with pen or highlighter and analysed. Web search and further SAP and non-SAP systems are added per project by configuration.
Writing back. Data goes from the chat back into the Analytics Cloud model, from story data or from the Excel upload. Raw, story and analysis data can be exported as Excel or CSV.
Retracting. Retractions on the SAP system are started by the user directly from the chat, optionally as a test run first.
How it is built
The chat is a framework of four parts: chat interface, agent orchestration, tool routing and the connection to Analytics Cloud. A project configures it and extends it with its own sources. Because the backend runs on BTP, the connections come from the platform: destinations and Cloud Connector lead into the SAP on-premise systems, reading and writing.
| Component | Technology | Decision |
|---|---|---|
| Chat | SAP Analytics Cloud | Inside the story’s interface, no second tool |
| Backend | SAP CAP on SAP BTP | Orchestration, tool routing and all connections in one place |
| AI | Claude Sonnet and Opus via AWS Bedrock | Models from EU regions, chosen per project |
| Analytics Cloud | Story data, SAC interface | Reading from the story with filters, reading and writing on the model |
| SAP on-premise | Destinations, Cloud Connector | Reading, writing and retractions without open ports |
Learnings
- The chat sits with the story. The question comes up while looking at the numbers, and that is where it is asked, with the filters and variables that currently apply. A chat in a second window would not have known that view.
- BTP brings the connections along. Destinations and Cloud Connector were configuration, not a build of their own. Only through them does the chat read from and write into SAP on-premise systems and trigger retractions there.
- The framework is configured per project. Which sources the chat knows, which systems it reaches and which models it uses is decided by the project. Here: AWS Bedrock, EU regions.