Hi everyone,
I’m facing an issue while using Appian AI Skill (GenAI) with an Excel file.
Has anyone experienced this issue before? Any insights on why the classification behavior changes like this?
Discussion posts and replies are publicly visible
This is expected behavior for GenAI-based classification.
Why This Happens: GenAI Classification is Context-Dependent
When the file is empty, the model classifies based on:Column headers onlyFile structure and formatting
When the file has data, the model classifies based on:Column headersActual data valuesPatterns and content within cells
These are two different inputs — different outputs are expected
Classification changes because the model reads the document's actual content, not just the structure. An empty Excel shows only headers, while a populated one adds text and data that shift the classification features; so the same template gets treated as two different documents.Add both empty and populated files under the same class label in your training set, or convert Excel to PDF before classification (AI Skills work most reliably with PDFs).
hi Shubham Aware , thanks for reply,
We have a requirement where we receive Excel files from users through email. Once the email reaches Appian, we need to classify those Excel documents.
Each time, the data inside the Excel file is different, so our GenAI model has to classify based on different structures and content.
Initially, we trained our GenAI model using an empty Excel file and some sample data. However, when real data comes from users, it becomes difficult for us to handle, and many files are getting classified as Unclassified. This is unexpected behavior.
We want to know if there is any way to achieve this effectively?
Also, we are using PDF files as well, and there are no issues with PDFs — classification is working fine.
Document Classification needs a stable layout (that's why PDFs work). Excel with constantly changing data has no consistent signature, so it lands as Unclassified; more training won't fix it.Convert each Excel to PDF first, then classify. Or skip the AI model for Excel and route by filename, sender, sheet name, or header row (via Excel Tools plugin) using a rule.