As an accountant (Kai — "Isn't there an easier way?") I want the system to handle documents differently based on the overall confidence — straight through, quick confirm, or targeted clarification — so that I dispatch routine cases with no effort and only touch the exceptions. TODO (2026-05-29): Story not yet round. Open strategic questions: How do we build customer trust to allow STP in the first place? Are 1-click postings the precursor to STP (ramp-up mechanism per instruction/vendor)? Will the pilot put us in a sustained state where many 1-click postings happen before STP kicks in? The story cut becomes final only after this is clarified — it touches the pilot activation strategy, pricing telemetry, and customer success accompaniment.
With a high overall confidence the document runs all the way to posting automatically (STP, straight-through processing) — no user intervention needed With medium overall confidence the document receives status "AI suggestion" per affected value or line item; the accountant can confirm the values with a single click (see #340367 ) With low overall confidence or individual low-confidence values the document receives status "clarification required" per affected value or line item, with a structured handling instruction in the detail view The counter metric "correction rate after posting" is calculated continuously per client; if it exceeds the threshold the system automatically holds back the activation of new client-specific instructions The system action "confidence evaluation and routing" is recorded in the log with who, what, and when Out of scope for MVP Customer-side configuration of the confidence thresholds — phase 3 (linked to #340359 ) Cross-client threshold recommendations — phase 3 Learning the thresholds from correction patterns — phase 3
No child work items.
No linked pull requests.
Plan review. The agent proposed a plan. Review it below, then approve so it can start coding — or request changes.
Implement the approved work packet "Add structured logging to the preprocessing document-processing pipeline" with a scoped branch and human-reviewed PR.
medium riskThe implementation agent can change source code and open a PR, so dashboard approval is required before coding and before PR creation.
None listed.