The Economics of LLM Deployment in Capital Markets Operations
This research note examines the cost structure of LLM-based services in capital markets operations. Using the AAL benchmark corpus as a case study, we show that $150–200 in API expenditure generated a 750-case evaluation framework that underpins four commercial product lines at 98–99% gross margins. The margin profile is not accidental — it follows directly from an architecture that restricts LLM inference to classification tasks and delegates all arithmetic to deterministic code. We present unit economics by product tier, scenario planning for 2026–2028, and implications for capital markets AI deployment strategy.
The cost of building the evidence base
D-001 (~$45), D-002 (~$50), and D-003 (~$60–80) total approximately $150–200 in API expenditure across 750 cases and 5 frontier models. In traditional capital markets technology procurement, building a comparable evidence base would cost tens or hundreds of thousands of dollars. The compression by two to three orders of magnitude is the first indication that something structurally different is occurring.
The architectural driver of margin
The 98%+ gross margin follows directly from the AAL architecture: LLMs handle classification only (Claude Haiku at ~$0.003–0.004 per call), and deterministic Python handles all arithmetic. If arithmetic were routed through frontier models (Sonnet or GPT-4o at $0.02–0.05 per call), the API cost per audit would rise from ~$33 to $100–200 and gross margins would collapse from 99.6% to approximately 60–70%. The deterministic architecture is superior on both cost and reliability dimensions.
Implications for capital markets AI strategy
Four implications: (1) margin structure should inform architecture — map task types to model tiers; (2) the evidence base is a moat — 750 validated cases are a proprietary asset; (3) incumbent vendors face a structural challenge — 98%+ margins enable aggressive pricing; (4) buyers should scrutinize vendor cost structures — a vendor who cannot explain their inference cost is either overpaying for compute or underinvesting in evidence.