Extraction Errors: What to Do When the AI Reads a Field Incorrectly

Extraction Errors: What to Do When the AI Reads a Field Incorrectly

In this section, you will learn how to act when Dost’s AI incorrectly extracts a specific field: supplier VAT ID, invoice date, amounts, VAT types, invoice number, units, discounts, and more.

Understanding which tool to use in each case helps you resolve issues in a structured way instead of repeatedly fixing the same errors manually.

Why Can the AI Misread a Field?

Dost’s AI performs its best interpretation of the document, but errors may occur in situations such as:

• The supplier uses an unusual invoice format (complex tables, multi-page layouts, special characters).
• The field appears in an unexpected position or non-standard format (for example, dates in Catalan or uncommon formats).
• The document is a low-quality scan or image.
• The supplier has recently changed their invoice layout.
• It is a new supplier with insufficient training history.

Decision Tree: Which Tool Should You Use?

When you detect a misread field, follow this decision flow:

Step 1 – Is it a one-time error or recurring for this supplier?

• One-time error (first occurrence or atypical document):
  Correct the field manually and mark the document as Reviewed. No further action is needed.
• Recurring error for the same supplier:
  Proceed to step 2.

Step 2 – Is it a reading error or a business rule issue?

• Reading error (AI misinterprets the value):
  Use self-training (see Self-Training section). The goal is to teach the AI to correctly read the field for that supplier.
• Business rule transformation (value is correct but needs adjustment):
  Use a mapping (see Mappings section) to transform or complete the value according to your rules.

Step 3 – Is it a format-related issue (date, currency, decimals)?

• If dates are in a non-standard format (for example, Anglo-Saxon format or Catalan text), check Format Settings in document configuration.
• If decimal or thousand separators are incorrect, also adjust them in Format Settings.

Common Cases and How to Fix Them

VAT ID (CIF/NIF) Not Detected or Incorrect

Common cause: VAT ID appears in an unusual position or includes variations (e.g. dashes, “ES” prefix).

Recommended solution:

• Correct the VAT ID manually.
• Use self-training and tag the document with the supplier’s VAT ID (see How to Train Dost’s AI for a Specific Supplier).
• Repeat for at least 5 invoices from the same supplier.

💡 If the supplier includes the “ES” prefix and this causes issues, you can create a mapping to remove it or transform it into the format required by your ERP.

Invoice Date Read in Incorrect Format

Common cause: Supplier uses a different date format (MM/DD/YYYY instead of DD/MM/YYYY) or writes dates in text form (e.g. “15 March 2026” or Catalan format).

Recommended solution:

• For a specific supplier: adjust Format Settings in document configuration.
• For textual or language-based dates: use self-training.
• If it is a global issue: update default format in Format Settings.

Amounts, VAT, or Net Values Incorrect

Common cause: Complex table structure, multiple VAT rates, or non-standard invoice layouts.

Recommended solution:

• Correct values manually.
• Use self-training if the issue is recurring.
• If calculations are required (e.g. deriving net amount from total and VAT), use a mapping (see mapping use cases).

Invoice Number Misread

Common cause: Invoice number appears in a non-standard location or is mixed with other identifiers.

Recommended solution:

• Correct manually.
• Apply self-training.
• If the issue persists, contact Dost support with supplier VAT ID and a sample invoice.

What If Training Does Not Improve Results?

If self-training does not resolve the issue after several documents:

• Verify that training examples are representative and not edge cases or corrupted formats.
• Check in the Self-Training section that training status is Completed (see Training Status article).
• Consider whether a mapping is more appropriate than training.
• If the issue persists, contact support with:
  • Supplier VAT ID
  • Failing field
  • Expected vs actual behavior
  • At least one representative invoice example

Best Practices

• Do not correct the same error manually more than 2–3 times without applying a corrective action (self-training or mapping). Manual effort will exceed setup effort over time.
• When you detect a new error pattern, document it internally: supplier, field, and applied solution. This improves long-term maintainability and onboarding of new team members.
• Self-training and mappings are complementary: you can train the AI to read data correctly and also use mappings to transform it according to business rules.