Enterprise AI assistant
Business users needed a way to ask questions about reporting data in plain language. I built the assistant end to end and worked with users to improve how it handled their questions.
- My role
- End-to-end application development
- What I focused on
- Report ingestion and question handling
Handle inconsistent reports
Report identifiers sometimes changed, and some reports arrived without fields the application expected. I updated the ingestion path to handle those conditions explicitly so that changes in the source data wouldn’t quietly carry through to the assistant.
Test the language users actually use
During acceptance testing, business users phrased familiar requests differently from my initial test questions. In some cases, the assistant reported that information was unavailable even though the underlying reports contained it. I worked with users to understand their terminology and adjusted how the assistant handled those questions.
- Test wording
- “Which records are still open?”
- Alternate wording
- “What hasn’t been closed yet?”
The expected result is the same set of records.
What I took from it. When an assistant says it can’t find an answer, check whether the data is present and whether the system understood the request. Each calls for a different fix.