The Transformations page on the Organization specific tab, showing the Rules list sorted by Title with two rule groups expanded to reveal their individual rules and switches: an enabled group, and a fully disabled group shown in grey text

Transformations are programmatic rules that reshape, normalise, and standardise extracted data after AI processing but before export to downstream systems. They sit in the middle of Docupath's three-part business rules and instructions layer - after Instruction Builds (which guide extraction) and before Rejection Rules (which enforce compliance guardrails). Transformations handle field-level adjustments such as setting default values, converting units, applying percentage adjustments, and performing CSV-based lookups, ensuring that data leaving Docupath matches the exact format expected by receiving systems.

Transformation Rules execute after AI extraction is complete and after Instruction Builds have guided the extraction process. They run before Rejection Rules evaluate the document, and before the data is exported. This sequencing means transformations operate on fully extracted data and can clean or reshape values before compliance checks are applied.

To create a transformation rule:

Once deleted, the rule no longer applies to documents processed under that scope.

A transformation rule group can be disabled without being deleted, either as a whole or one sub-rule at a time. Only enabled rules are evaluated when a document is processed; a disabled rule stays configured but is skipped entirely.

On the Transformations list:

The Title column is the first column in the list and can be sorted, so groups can be found by name before toggling them.

Each rule also carries its own switch on the create and edit screens, so a rule can be saved in a disabled state and switched on later once it has been reviewed.

Switch changes take effect in the interface immediately and are saved after you stop toggling. If a save fails, the group reverts to its last saved state and an error message appears.

Rules created before this capability existed are enabled by default, so existing behavior is unchanged until a rule is explicitly switched off.

When the AI generates a transformation rule, the output is composed of three components:

Users do not typically need to edit the generated rules manually, but they can be modified directly if needed.

Transformation rules can be configured at multiple scopes. When multiple rules could apply, the most specific scope takes precedence:

Transformation rules can reference supporting files (e.g., master data, vendor information, price lists etc.) for value mapping scenarios - for example, mapping country codes to country names. Supporting files are uploaded to the tenant instance separately and can be linked to a rule during creation by clicking the "Link Supporting File" link. Use the CONTAINS and NOT_CONTAINS ("Does not contain") operators to act on whether a field matches an entry in a linked file - for example, filling a field when a lookup value is found (CONTAINS) and clearing it only when none is found (NOT_CONTAINS). Because NOT_CONTAINS is a true substring check, it affects only genuine non-matches, so values that did match keep their result.

Actions within a generated rule execute top-to-bottom. When sequence matters (e.g., convert units first, then add tax), actions should be ordered accordingly within the same rule, or split across separate rules with explicit sequencing.

Each generated rule shows a short, automatically generated title in its header, next to its position number - for example, Rule 1 : Set Currency to "USD". The title is derived from the rule's own content, so you can identify a rule at a glance instead of reading it by position alone. It is separate from the Title you enter for the rule group as a whole.

A transformation rule's title is built from its first action:

Field paths are shown with arrow notation (for example, Buyer -> Name). When a rule has more than one action, the title describes the first action and appends +N more to stand for the rest. A normal rule always has a describable first action, so a title is shown; the header falls back to just the position number (with no title) only in the rare case of an incomplete rule with no usable action.