Agent Reimbursement Pre-Audit & Acceleration System
AI Pre-Audit + Human Adjudication · Clear 300 Pending Reimbursements in 5 Days
Built for finance departments in precision-manufacturing and other large enterprises, the system uses the paradigm of "agent pre-audit + risk tiering + human adjudication": AI takes over the four repeated steps — invoice verification, duplicate detection, standard matching, and image check — while the finance manager keeps the final adjudication right. We ran a 6-day PoC for a 480-person manufacturing client and cleared all 300 pending reimbursements within 5 business days: manual review workload dropped 76%, average handling time fell from 15 to 4 minutes, and every abnormal judgment came with evidence — 100% explainable.
Core Modules
Before: 300 Pending Reimbursements on the Original ERP
Expense Review Queue Workbench
Multi-Dimensional Risk Filtering
Policy & Evidence Library
Core Philosophy
AI does not replace finance accountability — it takes over the standardized, repeatable labor. Audit shifts from "investigate four steps from scratch" to "adjudicate on Agent-supplied evidence."
Scope Discipline First
Stay on the single line of "reimbursement audit" — no ERP rebuild, no post-audit flow, no non-reimbursement work, no auto-payment. Not creeping the scope is the prerequisite for 6-day delivery.
Tiering Is the Leverage
Low-risk auto-pass (manager spot-check), medium/high routed to the human review workbench. Reviewers shift from "look up four steps" to "adjudicate on Agent-supplied evidence".
100% Explainable
Every abnormal judgment must attach evidence fragments and rule basis. Managers can send the rejection reason to employees directly without further explanation — explainability is the prerequisite for finance to trust the system.
On-Premises Deployment
The whole stack runs on the client's virtualized in-house resources: data never leaves the network, rules are configurable, ERP is read-only via API — no intrusion into the core ledger, simple ops.
Industry Status & Pain Points
FDE cross-validated the pain points through system data, policy materials, and 7 stakeholder interviews during the PoC. The pain points were not directly given by the client.
Fully Manual Per-Claim Review
Reviewers hand-run "verify → check amount → check standard → check duplicate", ~15 min per claim. 300 claims ≈ 9–10 person-days, chronic backlog.
Invoice Authenticity by "Feel"
No verification API, reviewers judge by experience. Sampling found 3 cases of abnormal invoice codes — risk of fake/cloned invoices slipping through.
Duplicates by "Memory"
Whether the same invoice/amount has been reimbursed depends on memory and Excel cross-check. Cross-validation found 11 suspected duplicates.
Inconsistent Expense Standards
Travel/entertainment standards conflict in interviews: B says "lodging ≤400/night", D says "tier-1 ≤500, others ≤350". Policy files are missing; same case, different decisions.
Missing/Blurry Receipt Images
Of 4,222 images, 217 lack claim linkage or are unclear. Manual flipping is easy to miss, with no system prompt.
No Priority, No Tiering
300 claims mix high-risk (large amount, over-standard) and low-risk (small amount, complete docs) in one queue — resources misallocated.
Conclusions Not Traceable
Rejection reasons are mostly verbal/abbreviated, lacking structured evidence. Employees dispute, managers cannot audit.
Core Capabilities · 10 Functions
Designed around the full chain "ingest → pre-audit → tier → adjudicate → feedback". Every capability has clear priority and boundaries.
Multi-Source Data Ingest
Pull claims, invoice ledger, and receipt images via open-platform API; build claim–invoice–image linkage (F1).
Receipt Image OCR
Recognize seller, amount, tax ID, invoice number, and date from images; output structured fields (F2).
Invoice Verification / Consistency
Cross-check invoice number, code, and amount against ledger and OCR; if tax-authority API available do real-time verify, otherwise rule-approximate with confidence label (F3).
Duplicate Reimbursement Detection
Flag second occurrence of same invoice number / same amount + same seller. Cross-checked 11 suspected duplicates in the ledger (F4).
Expense-Standard Rule Engine
Validate lodging, meals, transport upper limits by type. Rules visual + configurable, finance team self-service (F5).
Budget / Department Attribution Check
Verify claim department matches budget attribution and does not exceed department budget (F6).
Risk Tiering
Output low/medium/high risk; low risk auto-passes, medium/high goes to human review (F7).
Human Review Workbench
List claims to review with Agent-flagged items, evidence fragments, and suggested conclusions; single-screen review (F8).
Structured Rejection & Audit Trail
Rejections must include structured reasons; all judgments leave a trace and are exportable (F9).
Acceptance Metrics Dashboard
Real-time display of throughput, auto-pass rate, abnormal recall, manual effort, and more (F10).
TO-BE Flow: From Ingest to Adjudication
Employee submits → system ingest layer (API + OCR) → agent pre-audit engine → risk tiering → low-risk auto-post / medium-high human adjudication → conclusion feedback + metrics dashboard.
Key Results · 7 Core Metrics
Based on 6-day PoC measurements: baseline = pre-PoC fully-manual; target = post-PoC AI pre-audit + human adjudication.
Key Breakthroughs
The difficulty is not single-point rules but connecting discover–tier–adjudicate–feedback into an auditable, quantifiable, reusable pipeline.
Scope Discipline for Delivery
Stuck to the single line of "reimbursement audit" for 6 days. No ERP rebuild, no post-audit flow, no non-reimbursement work, no auto-payment. Not creeping the scope is what made 6-day delivery possible.
Tiering as Leverage
Not "AI replaces human review" but "AI frees humans from repeatable labor": low-risk auto-passes (manager spot-check), medium/high enters the workbench, reviewers shift from "look up four steps" to "adjudicate on Agent-supplied evidence".
100% Explainable
Every abnormal judgment must attach evidence and rule basis. Managers can send the rejection reason to employees directly without further explanation — explainability is the prerequisite for finance to trust the system.
On-Premises Runnable
The whole stack runs on the client's in-house virtualized resources. Data never leaves the network, ERP is read-only via API, rules are visual + configurable. 300-claim batch pre-audit ≤ 4 hours.
Business Value
Convert technical capability into quantifiable, reusable business value.
Content Under Active Update
Product content has been published based on internal materials. The following areas are planned for further development:
Other Products & Solutions
Explore Xianma AI solutions in other domains