Best practices to train AI for smarter automation

To maximize the benefits of AI-powered predictions, follow these best practices to train the AI effectively. Regular use, consistent feedback, and clear corrections will help AI refine its predictions over time.

1. The more you use it, the smarter it gets

AI/ML improves with continuous use—the more transactions you process, the better predictions will become.

Every correction refines the model, enhancing accuracy for future transactions.

2. Post transactions to train the AI

AI only learns from bills that are posted or submitted for approval. Unposted draft transactions do not contribute to training.

To improve predictions, always finalize transactions instead of leaving them unposted.

3. Set up default GL accounts

Assign a default GL account for your supplier records. To ensure that the default expense account is being considered and up-to-date, edit each supplier and select Save.

If the AI has low confidence on its GL account prediction, it will fall back on the default GL account for accuracy.

4. Provide accurate feedback based on transaction structure

AI predictions work best when feedback aligns with the requested transaction structure. In this context, feedback refers to any action completed and recorded by the AI to learn your preferences.

Single-line versus multi-line processing

• If you request a single-line prediction, only submit single-line feedback. Multi-line edits will not be captured by the AI in this instance.

• If you need multi-line details, make sure that your AI request is set for multi-line before submitting feedback.

• For multi-line corrections: Always update the existing lines instead of deleting and adding new ones. This keeps AI learning in the correct order.

5. Avoid deleting incorrect predictions—correct them instead

Avoid deleting transactions with incorrect AI predictions. Instead of deleting and re-uploading the same transaction, correct the data and post the transaction so AI can learn from the adjustment.

6. Train AI by creating transactions on different days

AI retrains every 24 hours, meaning diversity in transaction submission leads to better learning. Instead of entering all transactions on the same day, spread them across different days to give AI varied training data.

This helps AI recognize different transaction formats, suppliers, and patterns more effectively.

7. Understand AI learning patterns and accuracy expectations

AI extracts transaction data with high accuracy, but early results might require more corrections for GL account classifications. GL classification accuracy improves over time as AI learns from your company's specific patterns.

8. Enhance AI prediction on memos with clear descriptions

When correcting the memo field, write text that best matches the item description as provided in the e document. This helps the AI understand context and improves future predictions in the memo section.

9. AI populates most fields, but manual review is still needed

AI automatically fills in most fields on transactions, but some manual validations might still be required. Review key fields like GL accounts, tax codes, and department assignments before posting.

The more corrections that you provide, the faster AI adapts to your company's accounting practices.