OpenText's May 2026 release blog describes Core Capture for SAP Solutions 26.2 in one packed sentence: the intelligent voter "combines the three capture disciplines (knowledgebase, self-adaptive learning, and LLM-based extraction) to deliver the most accurate results ever." The same note says you can boost out-of-the-box extraction with GenAI, cover header and item fields across invoice and related document types, and activate the new functionality with a simple switch, without a mandatory VIM upgrade from 7.6 onward.
We already wrote about turning that GenAI switch on without a VIM upgrade: baseline first, keep the switch apart from the add-on upgrade, and name who flips it. This piece is about the voter itself. When three disciplines disagree, who decides which one wins, and when GenAI is allowed to overrule the other two?
Three voices, one field. A knowledgebase answer reflects rules and templates you already trust. Self-adaptive learning reflects what your validators have corrected over time. LLM-based extraction reflects what the model reads on this document. Straight-through processing improves when those voices agree. Governance starts when they do not.
The public blog does not publish a tie-break table. That is fine for marketing. It is not enough for a production switch. Someone in your shop has to write the policy the voter is allowed to follow.
Policy before the switch. In our view the decisions sit with named owners. These labels are ours, not OpenText's:
- Capture owner. When GenAI may propose a field the knowledgebase left blank, and when a GenAI value may replace a learned value.
- AP lead. Which fields are never auto-accepted from GenAI alone (bank details, tax IDs, and high-risk coding are the usual shortlist).
- VIM process owner. How a disagreed field lands in validation: soft highlight, hard stop, or silent accept into a confidence band.
- Internal audit / SoD. Whether the extraction log shows which discipline supplied the winning value, so a later sample can explain a posting.
Without that, "most accurate" becomes whoever spoke last in the stack.
What we'd write down. Four lines are enough to start:
- Agree path. When all three (or the two that fired) agree within tolerance, accept and measure.
- GenAI-only path. When only the LLM returns a value, which fields may post, which must stop in validation, and who reviews the queue.
- Conflict path. When GenAI and the knowledgebase disagree, which one the voter may prefer by default, and which disagreements always need a person.
- Learn path. Whether a human correction updates self-adaptive learning, the knowledgebase, both, or neither for that supplier.
Keep measurement honest. The earlier GenAI post argued for a fixed supplier set and field-level correction counts before and after the switch. Add one cut: disagreements between disciplines. If GenAI "wins" often against a stable knowledgebase on bank or tax fields, that is a control finding, not a productivity win. Report it on its own line.
Version and scope notes. OpenText's lifecycle page, when we last read it on 10 October 2026, listed Core Capture for SAP Solutions only up to 24.2 (Current Maintenance to 30 April 2029) and had no rows yet for the 25.x and 26.x releases OpenText has announced since. Ask OpenText which release your tenant runs, and confirm in the release notes how the voter exposes winning discipline and confidence. The blog's "all languages and countries" claim is OpenText's; prove it on the languages you actually receive.
The intelligent voter is a design choice dressed as a feature. Name who sets the policy when GenAI wins, then turn the switch on against a baseline you can defend.
Where does your landscape stand?
Take the free Assessment. It's about 10 questions, and you get a one-pager with your estimated level of effort and suggested next steps.
