Your AI Business Case Is Missing Its Biggest Cost
A practical framework for modeling AI inference costs, measuring cost per outcome, and building more accurate AI business cases
Most AI business cases I review have a revenue model and no cost of goods sold.
Nobody does this on purpose. Token pricing looks trivial, cents per million, so teams assume it rounds to zero. Then the invoice arrives.
Here is the P&L I build instead. One deployment, three versions, and the ticket count never changes.
BASELINE: An internal support agent resolving 40,000 tickets per day. Each ticket carries 8,000 input tokens of retrieved context and returns 700 output tokens.
Version 1: One model call per ticket on an efficient tier ($0.10 / $0.40 per million): 320M input + 28M output per day = about $1,300 per month, or $0.0011 per ticket. Rounding error. This is the number that gets into the deck.
Version 2: The same 40,000 tickets, now agentic. Six model calls each. Context accumulates across turns, roughly 8K on the first call rising to 20K on the last, averaging 14,000. Intermediate turns are terser than a final answer, so output runs about 500 tokens per call rather than 700: 3.36B input + 120M output per day = about $11,500 per month, or $0.0096 per ticket.
Version 3: The identical agent on a frontier tier ($3 / $15 per million): about $356,000 per month, or $0.30 per ticket. Same tickets. A 275x spread.
What moved was not on anyone's dashboard: how many model calls a task takes, and how fast context grows inside the loop. Seat count barely matters. Token path is everything.
And notice that $0.30 per ticket is not obviously wrong. Against a loaded human cost of several dollars per ticket, the frontier tier still clears. That is not a bad decision. It is an unmodeled one, which is a different problem, and a fixable one.
So stop tracking cost per user. Track these three:
Tokens per resolved task, not per request. A cheap model that needs four attempts is not a cheap model.
Cost per successful outcome, meaning a closed ticket, an approved document, a completed transaction. The only number that sits next to revenue on the same page.
Prefill cache hit rate. In a RAG or agentic workload, 90 percent or more of your tokens are input. Reusing that prefix instead of reprocessing it is usually the largest margin lever available, and most teams have never measured it.
Token prices fell roughly 1,000x in three years, and aggregate spend still went up. That is what happens when unit cost collapses faster than teams instrument consumption.
If you cannot state your cost per resolved task today, you do not have an AI business case. You have an AI hypothesis.
What is yours? I am collecting benchmarks across verticals.
Source: Scenario ACG Research modeling math, priced against published July 2026 API rates. Historical price decline per Introl, "Inference Unit Economics: The True Cost Per Million Tokens."
#Tokenomics #AIEconomics #LLMInference #FinOps #AIInfrastructure #ACGResearch
