Data preparation and analysis
An ERP assistant with prepared data behind the answer
PitchAI connected an assistant to an organization's ERP data, then developed the preparation and refresh processes behind its analysis. The system includes defined datasets, full-table analysis and records of data refreshes.
- Context
- Business operations
- Stage
- Deployed system
- Published
Start with the connection
The early work covered API setup and connecting a language-model assistant to the ERP system. That established access to the source data.
Later work gave the assistant a more structured basis for analysis. Business terminology and prepared datasets sit between the underlying sources and the questions asked of them.
Define the tables before using them
The data layer builds derived tables from declared inputs. It checks whether dependencies exist and whether a transformation actually returns a table. A missing input or an invalid result causes preparation to fail explicitly.
The production assistant requires this preparation to finish successfully before it starts. Its available datasets are therefore prepared before it handles questions.
Information about each dataset remains attached to the prepared view, alongside its values.
A preview helps find data; the full table supports analysis
Search can return a short preview of a dataset. That preview helps the assistant identify relevant material without printing every row into the conversation.
Analysis works with the full loaded table. The short preview helps find data; it does not set the calculation's row limit.
Keep refreshes tied to the running assistant
The assistant was deployed in February 2026. A September production update addressed how refreshed data reaches the running assistant: it publishes prepared datasets as a generation, records freshness information and reloads the assistant.
This extended the original API connection into a maintained data workflow. The source connection, dataset preparation and refresh record each answer a different question: where the data comes from, how it is prepared for analysis and which prepared generation the assistant has loaded.