What is data validation?
Data validation is the set of automated checks that decide whether a record is fit to accept. The checks work at several levels. Field validation looks at one value at a time, confirming for example that a date is real and a quantity is positive. Record-level checks compare fields within a record, such as a ship date that must not precede the order date. Cross-system checks compare the record with reference data, such as whether the customer exists in the ERP or the cost center is still open. Business rules sit on top, such as a discount limit that depends on the customer’s pricing tier.
Validation decides whether to accept new data, while data cleansing repairs data that is already stored. Both are part of managing data quality, and a data contract often lists the rules a producing team commits to.
Why data validation matters
An invalid record is cheapest to correct before it is accepted. Once it reaches the ledger, a forecast, or a customer’s inbox, fixing it means tracing every place it went. AI raises the stakes in two ways. Models trained on unvalidated history learn its errors. AI systems that extract or generate data produce plausible values that can still be wrong, and a language model asked for an invoice total returns a number every time. Validation is how the system tells whether that number matches the line items.
Data validation in practice
An industrial equipment manufacturer uses an AI agent to read supplier invoices and prepare them for payment. Each extracted invoice passes a validation layer before it reaches the accounts payable system. The purchase order must exist and be open, the supplier must match the one on the order, line amounts must add up to the stated total, and invoiced quantities must not exceed what the receiving dock recorded. Invoices that pass go to a clerk for approval, and invoices that fail go to an exception queue labeled with the rule that failed. Kirk Tech Solutions typically places a validation layer and a human approval step in front of any system of record that an agent can change.
What to watch for
- Write the rules with the people who run the process, because they know which exceptions are legitimate.
- Decide for each rule whether a failure blocks the record, holds it for review, or only raises a warning.
- Track how often each rule fails, since constant failures point to a broken upstream process or a wrong rule.
- Check AI output with the same rules as human input, starting with a schema check on structured outputs.