In brief
- Model cost depends on input and output tokens, model class, caching, runtime and number of steps, among other factors.
- Search, storage, document processing, tool calls and observability create additional variable costs.
- Development, testing, governance, support and human review belong in the total cost calculation.
- Cost per request is less meaningful than cost per correctly completed case or productive hour released.
Which cost components make up an AI solution?
A production application consists of more than a model call. It needs identity, data access, potentially search and document preparation, business logic, an interface, logs and monitoring. People and organisational introduction add further implementation and operating costs.
| Cost type | Typical components | How to plan it |
|---|---|---|
| One-off | Discovery, data preparation, integration, development, migration and adoption | Capture as project cost and amortise over the expected useful life |
| Recurring fixed | Base licences, platform, monitoring, support readiness and minimum capacity | Monthly regardless of actual case volume |
| Variable | Model calls, OCR, search, storage, data transfer, tools and human review | Measure by case type and multiply by volume, growth and peak demand |
| Hidden or irregular | Rework, outages, security assurance, model change, training and exit | Use observed pilot and operating effort plus a risk reserve |
What drives variable usage cost?
Text models are commonly billed for input and output tokens. Long system instructions, large document contexts, conversation histories and agent loops increase usage. Multimodal processing, real-time modes and premium service tiers have their own pricing structures.
A cheap model call can become expensive if it is repeated often or sends many cases to human rework. A more expensive model may be cheaper overall if it completes the process with fewer errors and steps.
- Volume, peak demand and expected growth
- Context length, output size and number of model steps
- Model routing, caching, batch processing and reserved capacity
- Failed attempts, retries and human correction
How do you build a defensible cost estimate?
Work with real case types rather than one average request. A short standard case, a document-heavy case and a complex exception can have very different costs. A load and usage scenario connects these cases to expected frequency and growth.
- Step 1
Measure case types
Capture tokens, tool calls, runtime and rework on representative cases.
- Step 2
Model volume
Calculate normal, high and extreme usage scenarios including peak load.
- Step 3
Add fixed cost
Allocate development, platform, people, security and support across the lifecycle.
- Step 4
Show uncertainty
Document price, exchange-rate, usage and quality assumptions as a range.
What does a realistic worked example look like?
This example is not a price quote; it demonstrates the method. A document process handles 24,000 documents per month. One-off implementation is spread over 24 months, while variable usage and human review use measured pilot assumptions.
| Cost block | Assumption | Monthly amount |
|---|---|---|
| Amortised implementation | CHF 48,000 for integration, testing and rollout / 24 months | CHF 2,000 |
| Fixed operation | Platform CHF 1,400, monitoring CHF 600, support and operation CHF 1,400 | CHF 3,400 |
| Variable technology | 24,000 documents × CHF 0.045 for model, OCR, search and storage | CHF 1,080 |
| Human review | 8% of documents × 2 minutes × CHF 75 per hour | CHF 4,800 |
| Reserve | Regression tests, exceptions, minor incidents and price variance | CHF 800 |
| Total | 12,080 / 24,000 documents | CHF 12,080 or CHF 0.50 per document |
Which metric connects cost with value?
A token or API call is a technical consumption unit, not a business outcome. Useful unit cost relates the fully loaded cost to a domain result: a correctly prepared case, a qualified assist or a manual handling step avoided.
- Cost per successful, quality-checked case
- Cost per accepted draft or avoided manual handling
- Cost split by case type, team, model and environment
- Budgets and alerts tied directly to usage and accountable owners
Example: classifying incoming documents
The team measures more than the price of each model call. It includes OCR, storage, classification, clarifications, human correction and operations. A smaller model is cheaper per call but creates more rework on complex documents. Routing uses it for simple cases and a stronger model for difficult ones. The decisive metric is total cost per correctly classified document.
What to remember
Plan and measure fully loaded cost per usable outcome. This makes architecture, model and process alternatives comparable and exposes scaling surprises early.
Sources and further reading
These primary sources provide further detail on definitions, technical foundations or responsible use.
Content reviewed