Economics
Forensic methodologies to audit, categorize, and optimize R&D capital allocation and GenAI margins.
Observed Evidence
The Direct Experience
"While conducting R&D audits for PE-backed B2B SaaS companies, I kept seeing the same margin leak: teams scaled AI features assuming flat SaaS software costs, only to experience margin collapse. The pattern was clear—power users consumed model queries far faster than subscriptions recovered costs. The underlying mechanism was Synthetic COGS—variable GPU and API runtime execution capitalized incorrectly as fixed hosting. This reveals a general principle: AI features have variable unit economics that scale non-linearly with user activity. The broader implication is that SaaS companies must pivot to consumption-capped or credit-based pricing models to remain structurally solvent."
Core Analytical Axioms
Forensically proven concepts in this operational boundary.
AI Unit Economics
The marginal cost structures of running generative inference models per user activity.
Organizations scale AI features assuming standard SaaS gross margins (80%+), only to experience margin collapse as AI costs scale dynamically with usage.
Calculating cost-per-interaction allows organizations to adjust pricing tiers or model routing before running at a loss.
Synthetic COGS
Attribute-based variable costs (GPU cycles, embeddings, vector search) that replace traditional static server opex.
Variable AI inferencing is incorrectly capitalized as fixed server hosting, masking structural gross margin erosion.
Correctly identifying Synthetic COGS ensures accurate gross profit reporting and models true product contribution margins.
Innovation Tax
The hidden cost of maintenance and bug-fixing disguised as new feature velocity.
VP of Engineering reports that 70% of R&D goes to new capabilities, when forensic audit reveals only 20% produces growth assets; the other 50% is legacy maintenance.
CFOs over-capitalize R&D spend, leading to surprise write-downs and stalled roadmap timelines.
Cost Per Outcome
The cumulative inference cost required to achieve a successful user result, accounting for failures, retries, and formatting errors.
A single user transaction requires multiple LLM round-trips due to prompt drift or formatting failures, multiplying the real variable cost.
If cost-per-outcome is high, user scaling causes rapid profitability decline rather than scaling efficiency.
Want to apply this to your organization?
Run a free diagnostic first. If the numbers concern you, book a session to build a remediation plan.
Richard Ewing — AI Economist & Capital Auditor