Imagine a bicycle repair shop. A customer changes a pickup note to Friday afternoon without mentioning a phone number. An AI-generated update must distinguish an omission from an explicit deletion. This is a hypothetical example.

For a JSON Merge Patch endpoint, a patch containing only the pickup note leaves the phone field unchanged. Represent business intent before building the update. Do not add null values merely to fill out a record.

In that same protocol, a null phone value means remove the phone member, not that its value is unknown. An extraction result must not silently become a deletion instruction. Deletion needs explicit intent and permission.

Arrays have a separate boundary. Merge Patch replaces an array; it does not append automatically. Submitting just one new tag can remove the existing tags. Review the full before-and-after difference.

Identify the format first. JSON Patch uses an ordered list of operations and is a different format. Sharing the PATCH method does not mean sharing deletion semantics. Check the contract when the business needs to store a literal null.

Test omission, explicit deletion, an empty string and a list with synthetic records. Write down the expected difference, inspect the actual request and read back the result. The goal is no unintended deletion or list loss. This is a proposed test, not one performed on a customer system.

Leave the team with one rule: reading can be uncertain, but writing must say retain, set or remove. Review the data difference, not just the AI-generated sentence.