Economics v2.0

Charles Dana · Monce AI · May 2026

Numbers below are measured, not modelled — from the live sweep of all 7 reference POs on 2026-05-29.

The Equation

Cost(PO) = CVLM + CSnake + Cinfra

Measured per-PO cost (live sweep, 2026-05-29)

POLinesTokens in / out$ BedrockWallPath
AM40753072~14k / ~1k$0.0740 sVLM
30026308002~25k / ~1.4k$0.1073 sVLM
45002398284~33k / ~2k$0.1489 sVLM
50024992071~38k / ~1.5k$0.14110 sVLM
PO00470915021~10k / ~0.7k$0.0531 sVLM
Satair 45005959861~22k / ~1.2k$0.0955 sVLM
Saft Ferak P250225326493k / 0.9k$0.1013 stext-mode + Haiku stages 0/1/5
Total275$0.69412 s

Average: $0.10 per PO, 0.25¢ per line item on the live sweep. The 264-line Saft Ferak PO costs the same as a 2-line Verizon PO because Stage 2 takes the deterministic text-mode path when totals reconcile to the cent.

Per-stage breakdown (typical balanced-mode run)

ComponentCallTokens (typ.)$/PO
Stage 0 (Client + plant ID)Haiku 4.5 × 11k in / 0.1k out$0.0013
Stage 1 (Doc Analyzer)Haiku 4.5 × 12k in / 0.3k out$0.0031
Stage 2 (text-mode — when totals reconcile)local$0
Stage 2 (VLM — otherwise)Sonnet 4.6 × 1–34k in / 2k out$0.04
Stage 3 (Rules)local$0
Stage 4 (plant-aware Snake)local$0
Stage 5 (Validation)Haiku 4.5 × 13k in / 0.5k out$0.0044
Stage 6 (Router)local$0

A "small" balanced-mode PO (1–6 lines, VLM stage 2) lands at ~$0.05–$0.15. A text-mode PO of any line count (so long as totals reconcile) lands at ~$0.01–$0.03 regardless of size.

Matching leverage

Stage 4 (Snake) is free (< 10 ms/line, in-process) but it resolves the heart of the problem: mapping heterogeneous manufacturer / customer part IDs to the right plant's master data. Measured tier-1 exact match rate on the sweep: 274 / 275 (99.6%). The single rejection is correct — a 3.4% Textron customs surcharge with no Saft SKU counterpart.

Manual todayMonce pipeline
Avg time per PO8–15 min (entry + SKU lookup)15–90 s automated + 30 s spot check
Avg time per 264-line PO4–6 hours of pure data entry13 s end-to-end
Tier-1 SKU match rate (measured)≈ 100% (operator does it)99.6% tier-1 exact, audited
Hallucination guardn/a0.85 floor on Snake; tier ⊥ rather than wrong SKU
Share routed to auto-approve0%~70% target once T ≥ 0.85

Break-even vs manual entry

Assume a fully-loaded cost of $45/h for a planning clerk. Manual PO ingestion averages ~12 min → $9.00 per PO. The pipeline runs at $0.10 measured average.

Saving per auto-approved PO ≈ $8.90

At 30 POs/day across both plants, that's $267/day in recovered operator time, or ~$67k/year gross before fixed infra. The 264-line Saft Ferak PO alone replaces ~5 hours of clerk time at $225 of opportunity cost — for $0.10 in Bedrock fees.

Fixed infrastructure (May 2026)

ResourceMonthly
EC2 r6i.large (2 vCPU / 16 GB / eu-west-3a)$96
EBS 30 GB gp3$3
EIP 51.44.2.200 (in use)$0
S3 archive (~5 GB versioned)$1
Route53 + CloudWatch + data out$4
Total$104/month

The bump from t3.medium ($30/mo) to r6i.large ($96/mo) was forced by Bordeaux's 800 MB Snake article model: deserialised to Python objects it consumes ~4 GB heap, which the 4 GB t3.medium OOM-killed. r6i.large gives ~11 GB free with both plants warm, comfortable headroom.

Volume scaling at measured rates

 10 POs/day  → ~$30/mo Bedrock  + $104 infra = ~$134/mo
 30 POs/day  → ~$90/mo Bedrock  + $104 infra = ~$194/mo
100 POs/day  → ~$300/mo Bedrock + $104 infra = ~$404/mo
500 POs/day  → ~$1,500/mo Bedrock + r6i.xlarge ~$200 = ~$1.7k/mo

r6i.large handles ~3 concurrent workers and ~30 POs/min sustained. The bottleneck is Bedrock throughput, not CPU or memory. Above ~250 POs/day a second worker queue pays for itself in latency; above ~1k POs/day SnakeBatch v9 (the 100-shard Lambda mesh) takes over the matcher.

Depth knob

The model_mode parameter on /extract trades cost for accuracy:

modestage 0/1/5stage 2 fallback$/PO (typical)
cheapHaikuHaiku$0.01–$0.03
balanced (default)HaikuSonnet$0.05–$0.15
accurateSonnetSonnet$0.10–$0.20

Text-mode short-circuits all of this when the document totals reconcile to the cent — that's how the 264-line Saft Ferak PO cost $0.10 instead of $1.50+.

VLM-dollars = accuracy-dollars

Cheapest first; escalate only when Snake's top-1 confidence is below θauto.