agentic-ai-engineering · Opportunity Framing

Invoice Processing Agent — Opportunity Assessment

Skill: agentic-opportunity-framing · Command: /opportunity-framing · Generated 2026-06-22
Verdict
BUILD
Strong fit across 6 of 7 process traits. Compounding value is moderate — consider starting with exception routing before automating approval decisions end-to-end.
Process Fit Score
6.2 / 7.0

7-Trait Process Fit Assessment

Trait Score Evidence Rating
Multi-step workflow High 8-step approval chain: extract → validate → classify → route → review → approve → post → archive
0.9
Judgment required High Policy-based approval thresholds, exception routing, vendor anomaly detection
0.85
Structured outputs High JSON extraction → ERP API integration, deterministic posting format
0.95
Compounding value Moderate Improves with volume but each invoice is independent — no state accumulates across runs
0.55
HITL feasible High Existing AP team review gate maps cleanly to HITL approval checkpoint
0.9
Cost-benefit positive High ~$40k/yr labor cost, 3,200 invoices/month, estimated 70% straight-through rate
0.88
Data available High 2yr invoice history, vendor master data, GL codes — sufficient for fine-tuning extraction
0.92

Risks and Mitigations

Financial Risk — HITL gate required
Invoice approval affects cash flow. Do not automate final posting without a human-in-the-loop checkpoint for exceptions above $5k threshold. Shadow mode recommended for 30 days before live routing.
Extraction accuracy — vendor format variance
Current vendor base has 47 distinct invoice formats. Extraction F1 must exceed 0.97 before enabling straight-through. Slot confidence thresholds need calibration per vendor tier.
Scope creep — start with exception routing
High-value win is routing unusual invoices to the right reviewer, not full automation. Constrain Phase 1 to extraction + classification + routing. Automate posting in Phase 2 after 60-day accuracy review.

Recommended Next Steps