We are automating a KYC onboarding process and evaluating how Agentic AI handles unstructured documents compared to traditional Document Understanding generative ML models. We have complex, variable table data that requires precise labeling.
If we utilize a Large Action Model (LAM) within an agentic workflow to parse the document instead of standard DU extractors, how do we enforce a strict confidence threshold? I need the agent to autonomously route the document to Action Center for human validation if the extraction confidence drops below 85%, but LAMs don’t output standard DU confidence scores.
@anon2539286 I wouldn’t rely on the LAM itself to provide a DU-style confidence score. Instead, add a validation step after the agent extracts the table data.
You can validate the extracted values using required fields, data types, formats, and business rules, and calculate a validation/risk score based on those checks. If the result is below your 85% threshold, route the document to Action Center for human validation. Otherwise, continue the process automatically.
A hybrid approach could work well here: use the Agent/LAM for flexible extraction of variable tables, then use UiPath Document Understanding validation and business rules to verify the extracted data.
This keeps the agent flexible while the 85% human-review threshold is controlled by the workflow, rather than depending on a native LAM confidence score.
Flow: Document → Agent/LAM Extraction → Validation → Score/Rules → <85% → Action Center | ≥85% → Continue
Hi @anon2539286
before going the LAM route, worth asking if you actually need to leave DU extractors. The generative extractors in IXP already return per field confidence and plug straight into confidence based routing and Validation Station, so you keep the 85% rule and the audit trail without building any of it yourself. Going agentic here means giving up exactly the thing you need most
and about asking the model for a self assessed score like suggested above, id be careful with that in KYC specifically. Those numbers are usually poorly calibrated, the model is often confident when its wrong, so you end up with a threshold that looks rigorous but isnt. In a regulated process thats hard to defend later when someone asks how the number was produced
what i would do is keep the extraction on IXP for everything it handles, use the agent only for the parts that really break it, and for those apply the rule based checks people mentioned, row counts, totals, formats. Rules are deterministic and you can explain them to an auditor
Hey @anon2539286 Since LAM doesn’t provide the same standard confidence score as DU, I think the better approach would be to add our own validation layer after extraction.
We can calculate a confidence score based on things like field validation, table structure, format checks, and consistency between extracted values. Then keep 85% as the threshold if the score is below 85%, the document is automatically sent to Action Center for human validation. If it’s above 85%, the process can continue automatically.
This way, we still get the flexibility of LAM while keeping a clear human-in-the-loop control for KYC.
Hey @anon2539286,
For LAM-based extraction, you can’t directly enforce the 85% DU confidence threshold because LAMs don’t provide standardized confidence scores.
A practical approach is to define a custom confidence/validation score based on field accuracy, schema checks, and extraction consistency.
If the score falls below 85%, have the agent automatically route the document to Action Center for human validation.