Comparing AI 智能体 repetition vs 独立系统 with 智能路由 for 100 SME bank statement sets.
This is a concrete demonstration of the 智能路由 + 分离式系统 layer in the AINNA 效率飞轮. 87% is an internal benchmark on this workload.
100 SMEs upload 100 different sets of bank statements. The goal is to generate accurate financial statements for all SMEs.
小 & 中等 企业版s
Different 银行 格式s
已生成 报告s
The AI agent reads, understands, classifies, calculates, validates and generates each report individually repeated 100 times.
AI is used once to build a reusable detached 系统. 银行 statements are segmented and intelligently routed with minimal context.
Fewer 令牌. Same outcome. Much smarter 系统 设计.
This study demonstrates the 智能路由 + 分离式系统 layer of the 效率飞轮. 87% token reduction is an internal benchmark on the tested bank-statement workload. 完整 loop: 分段 → 智能路由 → Distillation → 分离式系统 → 私密 基础设施. See LLM 战略es and 模型蒸馏.
Transparent model inputs for investor due diligence. 验证s the RM 288K/SME/year unit economics claim in the pitch deck.
| Parameter | 基础 Case | Best Case | 最差 Case |
|---|---|---|---|
| SMEs processed | 100 | 100 | 100 |
| 报表 per SME | 12 / year | 12 / year | 12 / year |
| Avg 令牌 / statement (方法 1) | ~8,500 | ~6,000 | ~12,000 |
| Avg 令牌 / statement (方法 2) | ~1,100 | ~800 | ~1,600 |
| Token price (external API) | $0.002 / 1K | $0.0015 / 1K | $0.003 / 1K |
| Annual savings (100 SMEs) | ~$5,200 | ~$7,800 | ~$3,100 |
研究 date: July 2026 · 本地 inference API fee assumed RM 0/token (GPU infrastructure amortization 已跟踪 separately)
This study is supporting evidence for the RM 2M investor round. See the full financial model, cap table and growth roadmap.
* 内部 validation based on 实时 platform data. 独立 audit available on investor request.
更低 GPU compute and data center load
Significant during heatwaves when cooling demand spikes
结构d 系统s make automation viable for smaller businesses
规则 + selective AI + human review layer
