Non-Commercial Evaluation License (v1.0): Free for personal operator review, CPA/CFO inspection, and non-commercial academic research only. Commercial AI model training, RL environment building, and commercial benchmarking require a written license from sam@levelcfo.com. View License & IP Notice · Canary: LEVEL-IP-CANARY-9F4A-7C2E-PERSONAL-USE-ONLY-CONTACT-SAM-AT-LEVELCFO-COM
The Level Index · AI Accounting, Fractional CFO & Operations Benchmark (v4.1.0)

Can AI Close Real SMB Books? 38 AI Systems Evaluated on 100 QuickBooks Takeover Checks, 15 SMB Case Studies & 30 AI Mistakes

Distilled from 2,200+ SMB client cleanups by Level. Unassisted raw foundation models score 31.0% to 71.0% and post bad entries on every run, while AI guided by Level Playbooks and controller review achieves 100.0% accuracy and 0.0% bad book entries.

100.0%
AI + Level Playbooks
0.0% bad-entry rate across 100 takeover checks & 15 SMB case studies
71.0%
Best Raw AI Model Alone
Gemini 4 Argon (Medium) · 81.0% with SOP prompt · Opus 5.5 at 69.0%
84.0%
Best Unassisted CLI Tool
Gemini CLI (Argon 4) · 100.0% on 26-check suite with Level checklist
15 Worlds
Anonymized SMB Close Cases
31 close tasks · 214 binary rubric checks · 27.83ms pre-flight
30 Mistakes
AI Accounting Fault Catalog
95 close guardrails (18 audit gates, 61 controller rules, 16 auditors)
Rank System / Model Tier Lab Raw Zero-Shot SOP-Guided 3-Run Pass^3 Bad-Entry Rate Exact Math Accrual Cutoffs Hold & Ask Subledger Time (s) Cost ($/close)