Financial Planning
AI Inference Costs: Model the Margin Per Completed Job
Build a customer-delivery cost tree for an AI product, including retries, tools, review and infrastructure, before explaining product margin.
What you will learn
- reconcile provider consumption to completed paid jobs
- define the customer-delivery cost boundary
- test how retries and human review change margin
Model AI inference costs around the completed customer job, not one successful model call. Include failed attempts, retries, tool charges, delivery infrastructure and required human work. Then compare the defined delivery cost with the corresponding customer revenue. A low token bill can coexist with an expensive product.
Use a customer job as the denominator
Define a job that the customer receives and pays for: a completed analysis, processed document or accepted delivery. Count every attempt needed to produce it. Keep nonbillable trials, failed deliveries and refunds visible instead of removing their costs.
AWS’s pricing structure illustrates why model and service dimensions must be checked separately. Match provider consumption to the same job cohort; do not divide a whole-month bill by only the final week’s successful jobs.
Draw the delivery cost tree
| Cost branch | Evidence to collect | Common omission |
|---|---|---|
| Model consumption | Input/output usage, cache policy, model, rate and date | Retry and failed-call tokens |
| Tools and data | Search, enrichment, OCR or other supplier usage | A second vendor invoice |
| Infrastructure | Storage, compute, orchestration and capacity allocation | Idle committed capacity or data transfer |
| Human delivery work | Required review, exception handling and support boundary | Labor needed to make the promised output usable |
| Transaction costs | Applicable payment and refund costs | Costs retained when the customer is refunded |
a16z’s AI-economics discussion highlights cloud and human-support costs. Classify actual expenses with the finance owner. A delivery contribution view is not automatically the reported gross margin.
Calculate a fictional customer-delivery cohort
Fictional assumptions, not vendor prices: 10,000 completed paid jobs sell for $0.20 each. Total consumption, including retries, is 100 million input tokens at $1 per million and 20 million output tokens at $5 per million. Other delivery costs are stated below. For this teaching view only, all listed costs are included and no additional delivery costs are assumed.
| Item | Cost |
|---|---|
| Input tokens | $100 |
| Output tokens | $100 |
| Tools and data | $150 |
| Allocated delivery infrastructure | $90 |
| Required human review, 10 hours at $30 | $300 |
| Transaction costs | $60 |
| Total defined delivery cost | $800 |
| Customer revenue | $2,000 |
| Delivery contribution / margin | $1,200 / 60% |
Cost per completed job is $0.08 under this boundary. Revenue minus these costs is not company net profit; product development, sales and other excluded expenses still exist.
Stress-test the expensive exception
| Scenario | Defined delivery cost | Margin on $2,000 revenue |
|---|---|---|
| Base | $800 | 60% |
| Token consumption doubles; other inputs unchanged | $1,000 | 50% |
| Required review rises to 40 hours; other inputs unchanged | $1,700 | 15% |
These scenarios isolate different drivers. More review can overwhelm token savings. Do not cut necessary review merely to improve a spreadsheet; quality and safety remain part of the product promise. Use the existing agent evaluation lesson for the separate task of testing internal output quality.
Copy the customer-delivery margin sheet
Customer job, completion and billing definitions:
Cohort period and job count:
Paid, failed, refunded and trial jobs:
Model/provider/version, rate date and consumption:
Retries and tool calls included once:
Infrastructure and committed-capacity allocation:
Required human review and exception labor:
Other delivery costs and exclusions:
Customer revenue on the same basis:
Cost per completed job and defined margin:
Sensitivity, quality guardrail and evidence owner:Exercise: do not count retries twice
A provider export already includes $240 of successful and retry calls. A separate log says retries account for $40 of that total. There are $160 other delivery costs and 2,000 completed jobs. What is the cost per completed job?
Answer: ($240 + $160) / 2,000 = $0.20. Adding the $40 again would double-count retries. Keep the retry share as a diagnostic inside the total, not another invoice.
Make the margin explanation reviewable
Tie each cost branch to an actual source, show exclusions and retain the quality requirement. Reconcile billed revenue, credits and completed jobs before comparing customer segments. Refresh rate cards and contract terms when they change.
Keep the approved delivery-economics brief with your fundraising materials in Finta. Ask Aurora to help explain which cost branch drives the sensitivity and what evidence is missing. Validate the figures in the source workbook before sharing the explanation with an investor or changing the delivery plan.
Finta Field Guide. By Finta Editorial Team. Sources checked November 14, 2026. Examples are original, fictional teaching exercises, not customer results, benchmarks or individualized financial, accounting, legal or tax advice.