Agentic large language model (LLM) workflows are entering finance back offices, yet
causal evidence on their operational impact remains scarce. We evaluate two agentic
workflows—(i) an AP agent that detects reverse-factoring eligibility, prioritizes
invoices, and orchestrates outreach; and (ii) an audit-export agent that shards and
validates large database extracts under file-size limits. Using a stepped-wedge rollout
across AP pods and audit teams, we estimate effects with a difference-in-differences
design that compares treated and not-yet-treated units over time while controlling
for supplier mix, period-end effects, weekday, and seasonality. The primary outcome
for AP is realized early-payment discount per € of eligible spend; secondary outcomes
include cycle time from eligibility to action, manual-touch rate, exception/rework
incidents, compliance deviations, and team Net Promoter Score (NPS). For audit support,
the primary outcome is export completeness and time-to-delivery; secondary outcomes
include re-run rate, checksum mismatches, and analyst time saved. We pre-specify heterogeneity
by invoice complexity (lines, attachments), supplier risk, demand peaks (month/quarter-end),
and team tenure. To test mechanisms, we analyze agent trace logs (plan length, tool
choices, retries), human-in-the-loop (HITL) rates, and policy-rule violations to distinguish
gains from earlier eligibility detection vs. faster routing and fewer errors. The
study provides practice-oriented effect sizes that translate into incremental discounts
captured, labor minutes saved, and avoided exceptions—offering managers a transferable
design, metric framework, and analysis plan for agentic automation in finance operations.
