# Token Economics for CFOs: How AI Usage Becomes a Software Bill > Translate AI tokens, credits, tool calls, retries, and platform fees into cost per workflow, accepted output, and business outcome. - Source: https://www.zarifautomates.com/blog/token-economics-for-cfos - Published: 2026-08-09 - Updated: 2026-08-12 - Pillar: Enterprise AI - Tags: token economics, AI costs, AI FinOps, CFO AI, LLM costs - Author: Zarif --- An AI invoice can be numerically precise and economically useless. “We used 800 million tokens” tells Finance that a meter moved. It does not explain which business unit caused the spend, which workflows succeeded, how many results employees accepted, or whether the company should buy more. The CFO's job is not to become a tokenizer expert. It is to build a bridge from technical consumption to business value. AI token economics is the discipline of measuring, allocating, and optimizing the full cost of model consumption relative to a defined unit of useful business work. - Separate resource meters such as tokens from business units such as cases, documents, or qualified leads - Model input, cached input, output, tools, retries, and provider-specific features separately - Convert SaaS credits back into operations and outcomes before comparing vendors - Measure cost per good outcome, not just cost per request - Include orchestration, data, review labor, observability, and unused commitments in total cost - Use low, base, and high workload cases because averages hide agent loops and power users ## Start With the Cost Stack AI operating cost has at least five layers.
LayerExamplesTypical owner
ModelInput, cached input, output, reasoning, fine-tuningEngineering
Model toolsWeb search, code execution, file processing, embeddingsEngineering
ApplicationOrchestration, vector store, database, APIs, loggingPlatform team
OperationsHuman review, support, evaluation, incident responseBusiness and operations
CommercialPlatform fees, minimums, unused credits, overagesFinance and procurement
Optimizing only the first line can raise total cost. A smaller model may require more retries or human review. A shorter prompt may remove context and reduce acceptance. A self-hosted model may reduce API spend while increasing infrastructure and operations. ## How Model Token Charges Work Text models split usage into input and output. ### Input tokens These include system instructions, user prompts, retrieved documents, conversation history, tool definitions, and tool results sent back to the model. ### Cached input tokens Providers may charge a lower rate when a repeated prompt prefix or context is served from cache. Cache rules, write charges, duration, and eligibility vary. ### Output tokens These include generated text and structured tool calls. Output is often priced higher than input. The exact ratio varies by model. ### Additional metered features Search, code execution, images, audio, data residency, fast processing, and provider-hosted tools may carry separate rates or multipliers. Anthropic's documentation, for example, explains that tool definitions and tool-use system prompts add input tokens, while some server-side tools create additional usage charges. ## The Core Model-Cost Formula For one model: **Model cost = input millions × input rate + cached-input millions × cached rate + output millions × output rate + tool charges** For a workflow: **Workflow model cost = sum of every model attempt, including retries and evaluator calls** For a business outcome: **Model cost per good outcome = total model cost ÷ accepted outcomes** The denominator matters. If 1,000 requests produce 700 usable results, dividing by 1,000 understates the cost of useful work. ## A Worked Token Example Assume an extraction workflow uses a model with these illustrative rates: - Input: $2 per million tokens - Cached input: $0.20 per million tokens - Output: $12 per million tokens One request uses: - 8,000 uncached input tokens - 2,000 cached input tokens - 1,000 output tokens The request cost is: - Input: 0.008 × $2 = $0.016 - Cached input: 0.002 × $0.20 = $0.0004 - Output: 0.001 × $12 = $0.012 - Total: $0.0284 At one million requests, the direct model bill is $28,400 before retries, tools, data, or review. Now assume 8 percent of requests retry once and 7 percent of completed outputs fail quality review. A simple effective-cost calculation is: **Model spend after retries = $28,400 × 1.08 = $30,672** **Accepted outcomes = 1,000,000 × 93 percent = 930,000** **Direct model cost per accepted outcome = $30,672 ÷ 930,000 = about $0.033** The listed request cost was 2.84 cents. The accepted-output cost is 3.3 cents. Add reviewer labor and the gap can become much larger. These numbers are illustrative, not a quote. Use the current provider rate card and measured token distributions. ## Why “Cost Per Million Tokens” Is Not Comparable by Itself Two models with the same rate can produce different bills because: - Tokenizers split the same content differently - One model needs a longer prompt or more examples - Output length differs - Tool schemas and context differ - Quality may require retries or an evaluator - One model completes the task while another escalates Anthropic's live pricing documentation notes that newer tokenizers can produce a different token count for the same text. That is why model selection should use a representative evaluation set and **cost per accepted task**, not rate-card arithmetic alone. ## Credits Add a Translation Layer AI SaaS vendors often hide raw model economics behind credits. Translate the invoice through four steps. **Dollars → credits → operations → accepted outcomes** For each feature, calculate: 1. Dollars per purchased credit 2. Credits per billable operation 3. Operations per completed workflow 4. Accepted workflows per period Suppose a $100,000 annual pool contains one million credits. One operation costs 20 credits, and a completed workflow averages three operations. - Dollar per credit: $0.10 - Dollar per operation: $2 - Nominal workflow cost: $6 If 15 percent of workflows fail or are abandoned but still consume actions: **Cost per accepted workflow = $6 ÷ 85 percent = about $7.06** Now include the base platform fee and unused credits. A credit plan can be more predictable and still have a worse effective unit cost than pay-as-you-go. ## The Fully Loaded AI Workflow Formula Use this management formula: **Fully loaded workflow cost = model + tools + orchestration + data + review labor + allocated platform + expected failure cost** Where: - **Model** includes every call, retry, router, and evaluator - **Tools** include search, extraction, third-party data, messaging, and other APIs - **Orchestration** includes n8n or another workflow/agent runtime - **Data** includes storage, retrieval, vector search, and transfer - **Review labor** includes average human minutes multiplied by loaded labor rate - **Allocated platform** includes seats, support, minimum commitments, and unused allowance - **Expected failure cost** captures remediation, service credits, or business loss where material This is not a financial-accounting rule. It is a decision model. Finance should define consistent allocation and distinguish cash cost, accounting treatment, and expected economic cost. ## Separate Cost per Request From Cost per Outcome The FinOps Foundation recommends moving from technical unit measures such as cost per token toward business units such as cost per assist, action, or case deflected. Use a metric ladder: 1. Cost per token 2. Cost per model call 3. Cost per completed workflow 4. Cost per accepted output 5. Cost per business unit 6. Value per business unit 7. Net value after AI and operating cost Different stakeholders own different rungs. Engineering can optimize the first four. Product and Finance have to define the business unit and value. ## Goodput: Pay Attention to Useful Work Borrow a systems concept: **goodput** is useful output, not raw throughput. For AI, failed requests are obvious waste. Harder categories include: - Outputs nobody reads - Suggestions users repeatedly reject - Summaries generated for inactive records - Duplicate analysis of unchanged data - Agent actions reversed by humans - Answers that trigger a repeat contact - Drafts that take longer to fix than to write Tag each workflow result as accepted, edited, rejected, escalated, failed, or unused where practical. The right level of instrumentation depends on risk and volume; not every low-cost internal draft needs forensic accounting. ## How Deterministic Logic Changes Unit Economics The first optimization is not always a cheaper model. It may be no model. Assume a customer-email workflow receives 100,000 messages: - 40 percent are automated receipts or system notifications - 30 percent match known templates and routing rules - 30 percent contain ambiguous language that needs classification Calling an LLM on every message creates 100,000 model requests. Filtering known senders and applying deterministic rules first reduces the ambiguous set to 30,000. The saving comes from removing calls, not trimming a few tokens. n8n can implement the filter, lookup, branching, API calls, and logging, then invoke a model only on the ambiguous branch. It can also route low-risk requests to a smaller model and escalate selected cases. The platform itself has a cost, but deterministic steps do not add external LLM tokens. Use the [deterministic workflow decision framework](/blog/deterministic-workflows-vs-llm-calls) before optimizing prompts. Then apply the [LLM cost-reduction sequence](/blog/reduce-llm-token-costs). ## A CFO Dashboard That Supports Decisions For each major use case, show: ### Demand - Workflows started and completed - Volume by business unit, product, and customer - Growth, seasonality, and peak rate ### Resource cost - Tokens and dollars by model - Tool and external API cost - Orchestration and data cost - Retry and failure spend ### Quality and adoption - Accepted, edited, rejected, escalated, and unused outputs - Evaluation pass rate - Human review minutes - Active users or downstream systems ### Business unit economics - Fully loaded cost per accepted outcome - Baseline process cost - Capacity saved or throughput added - Revenue, loss avoided, or service improvement where defensible - Net value and attribution confidence ### Commercial position - Commitment used and remaining - Forecast exhaustion or underuse - Effective unit rate versus list rate - Renewal and price-protection dates The dashboard should lead to actions: remove calls, change routing, renegotiate commitment, improve quality, expand a positive use case, or stop a weak one. ## Common Finance Mistakes ### Treating tokens as the value metric Tokens measure activity. Tie them to a workflow and outcome. ### Allocating by seat count Machine workloads may serve several departments through one service account. Allocate from request metadata, workflow ownership, and customer or cost-center tags. ### Ignoring unused commitment Prepaid credits reduce nominal rates only when used. Amortize committed cost across actual accepted outcomes. ### Comparing list rates without evaluation A cheaper model can be more expensive per good result. Test real workload distributions. ### Celebrating spend reduction alone A lower bill caused by lower adoption is not optimization. Review value, quality, and volume together. ## The First 30 Days 1. Inventory direct model APIs and AI SaaS credit pools 2. Select the top five use cases by spend or risk 3. Define one business unit for each 4. Add business-unit and workflow tags to requests 5. Separate model, tool, platform, and labor costs 6. Measure retry, failure, and acceptance rates 7. Calculate fully loaded cost per accepted outcome 8. Create low, base, and high forecasts 9. Assign Product, Engineering, Finance, and Commercial owners 10. Start a monthly decision review Continue with the [AI cost forecasting guide](/blog/forecast-budget-token-based-ai-costs) and the full [AI FinOps operating model](/blog/ai-finops-token-spend-unit-economics). ## Frequently Asked Questions **What is token economics in enterprise AI?** Token economics connects the variable units models process to the cost and value of a business workflow. It includes metering, allocation, forecasting, optimization, and measurement of cost per useful outcome. **How do you calculate the cost of an LLM request?** Multiply uncached input, cached input, and output token quantities by their respective current model rates, then add tool or feature charges. Include every retry, router, and evaluator call when calculating workflow cost. **What is the best AI unit-economics metric?** Use the closest stable business unit to value, such as cost per accepted document, completed case, qualified lead, or successful transaction. Keep token and call measures underneath it so Engineering can diagnose and optimize the cost. **Should human review cost be included in AI ROI?** Yes. Include the loaded cost of review, correction, escalation, and exception handling. AI that reduces model cost while increasing human cleanup may worsen total economics. **Can workflow automation reduce token cost?** Yes when it removes unnecessary model calls, reduces context, routes requests, or prevents retries. A platform such as n8n adds its own cost, so compare the fully loaded before-and-after cost per accepted outcome. --- ## Sources and Further Reading - [OpenAI API pricing](https://openai.com/api/pricing/) - [Anthropic Claude Platform pricing](https://platform.claude.com/docs/en/about-claude/pricing) - [Token Economics: The Atomic Unit of AI Value — FinOps Foundation](https://www.finops.org/insights/token-economics-the-atomic-unit-of-ai-value/) - [Capability: Unit Economics — FinOps Foundation](https://www.finops.org/framework/capabilities/unit-economics/) - [Token Economics: Managing AI Value in SaaS Model Token Costs — FinOps Foundation](https://www.finops.org/wg/token-economics-saas/) - [Agentforce pricing — Salesforce](https://www.salesforce.com/agentforce/pricing/) - [Intercom pricing FAQs](https://www.intercom.com/help/en/articles/8344190-pricing-faqs) - [n8n pricing](https://n8n.io/pricing/) - [From Seat-Based to Token-Based Pricing](/blog/seat-based-vs-token-based-pricing)