OpenAI and Anthropic invoices are real but incomplete. Year-one cost includes data pipelines, human review, compliance review, and the model-update tax — re-embedding, re-evaluating, and redeploying when providers or policies change. Founders who budget only OpenAI line items discover in month four that human review, re-embedding jobs, and legal review hours exceeded model spend — plan for the full stack.
Inference is often minority spend. Data, eval, compliance, and ops dominate year one.
Build-phase economics
Focused AI MVPs typically run eight to sixteen weeks depending on data messiness and compliance. Data cleaning and access negotiation often consume thirty percent of early sprint capacity. Founders who budget only engineering hours underestimate calendar time. Discovery sprints typically cost ten to twenty percent of total MVP budget but prevent thirty to fifty percent rework when data reality surfaces. Include founder and domain expert time in cost models — their interview and review hours are not free.
Discovery and scoping should precede fixed bids — ambiguous data estates inflate cost unpredictably. Thin-slice architecture reduces rework when the first golden-set eval fails. Pay for clarity before paying for scale. Compliance-heavy verticals add security questionnaire responses, pen test remediation, and BAA legal fees that flat-rate dev quotes exclude. Factor two weeks of buffer for data access delays even when founders insist data is ready.
Run-phase variable costs
Model API spend scales with users, context length, and tool-call depth. Vector databases, embedding jobs, and human-in-the-loop review add margin in high-risk domains. Chat-heavy B2C products face different unit economics than batch B2B summarization. Unit economics should model p95 token usage, not average — power users destroy margins calculated on mean consumption. Embedding refresh jobs after document updates are recurring compute most spreadsheets omit.
Cache stable retrieval results and compress prompts where quality allows. Cost dashboards per successful task beat monthly invoice surprises. Plan tiered models — small for classification, large for generation — from architecture day one. Human-in-the-loop review for healthcare or finance may cost more per task than inference — price products accordingly or automate review queues aggressively. Multi-model routing reduces spend when small models handle classification and large models handle generation only.
Hidden line items
Legal DPAs, BAAs for healthcare, security reviews, and pen tests appear on enterprise deals. Prompt regression engineering and eval CI are ongoing — not one-time setup. Prompt regression CI, observability SaaS, and vector storage grow with MAU — tag these as COGS in financial models. Model deprecation forces re-eval and sometimes re-tuning — budget quarterly maintenance sprints.
Budget fifteen to twenty-five percent of initial build cost annually for AI ops and model churn. See our pricing guide for how we structure engagements against these realities. Compare build versus buy for generic RAG infra — sometimes managed retrieval saves one FTE worth more than the vendor fee. Our pricing guide maps engagement phases to these cost categories explicitly.
What to do next
Start with one measurable outcome — latency, accuracy, adoption, or revenue — and design your stack around that constraint. GenOrbit ships production AI systems across healthcare, astrology, fintech, and SaaS; if you want a second pair of eyes on architecture or delivery, we are happy to talk.
Related: /services/product-strategy-mvp · /blog/pricing-ai-product-development · /blog/ai-observability-from-day-one