Skip to main content

I have created several learning instances, and I am noticing that there are cases where the extractions results/field values differ between what I get when running my automation vs. testing for the same document under the AI tab > Document Automation > Process.

I removed all other possible differences, by running a task bot with a single command to use Document Extraction: Extract Data, and saving the output to a local folder. Then on the same machine, (having already enabled test mode) manually selecting the same document for processing under the Document Automation tab for the same learning instance.

In the end, I get different results for the same field.

Ex.
Task bot Document Extraction returns: 505.20, “#111000”
Document Automation returns : 505.20, “123 Some Address, State, 99999”

Is this some known issue? This makes it hard for me to have confidence on the document extraction as a whole, if I can’t even rely on the testing/training I do for the learning instances.

Differences between results you see when running a Task Bot with Document Extraction → Extract data, and
Processing the same file under AI → Document Automation → Process does happen. In almost every case I’ve seen, it comes down to version/config drift (and sometimes provider differences) between what the UI is using vs. what your bot is invoking—not a random bug.

 


1. You may be running different versions of the Learning Instance

 

In test mode, the Process UI lets you explicitly pick a version before you process/validate. If your Task Bot isn’t pinned to that same version, it will use the Learning Instance’s current/default config behind the extraction bot—and that may not match what you tested. 

 

2. Different providers/settings per version (OCR/GenAI/rules)


Test mode was designed for A/B testing—changing OCR provider (e.g., “Digital PDF Extractor”), turning GenAI on/off, tweaking prompts, adding field/document rules, adjusting aliases, etc. Each tweak creates a new version with potentially different extraction behavior. If your bot points at a different version, you’ll see different values. 

 

3. GenAI prompts vs. non‑GenAI extraction


When GenAI is enabled and a “Search query for generative AI model” is set on a field (e.g., an Address), the engine may extract a full postal address instead of a short token like “#111000.” If your test version has GenAI enabled but the bot uses a non‑GenAI version (or vice versa), outputs will diverge. (Automation Anywhere also cautions that GenAI can produce varying outputs; validating and locking versions is key.) 

 

4. Validation & learning feedback affect results in Test mode


In test mode, documents always route to the validation queue so you can review results; when you submit & reprocess, your corrections and rules can alter extraction for that version. If your bot run isn’t using that updated version, you won’t see the same values. 


5. Package/version mismatches


The underlying Document Extraction package version can differ between environments/runs and impact extraction behavior. Newer releases even show which package version was used in the test-mode validator to aid debugging. 


6. Rules & data types can transform values


Field/data types (text, number, date, address) and validation rules (field or document level) can normalize, replace, or combine values—again, these are versioned. If rules differ between the version used by your bot and the one used in test mode, values will differ.


Hi Padmukar,

I hadn’t specified earlier, but the LI versions were the same. Before 9/29 there was only the v1 available (v2 was created when test mode was enabled). I had also tested specifying an LI version on the task bot to be sure.

Thankfully that ruled out all the points except 5, and since the original task bot has been in production for some months now, it was running one package version behind  the current “default”.

 

I’ll have to monitor and use Bulk update to default package option more often.