The real cost of AI isn't on the invoice.
Model calls are only part of the cost. Retries, orchestration, evaluation, human review, and infrastructure keep running long after the demo. Olensys measures the full cost of an AI system in production, identifies the engineering that drives it, and tells you whether it deserves to scale.
Measure the cost. Prove the value. Scale with control.
Cost per successful task · Model usage · Retries and orchestration · Evaluation and oversight · Human intervention · Production infrastructure
The invoice is the part above the waterline.
Choose a workload, set its volume, and select the model. Change the assumptions and see how each cost component affects the total. Your inputs stay in your browser.
AI reads each document and puts its fields into your system.
Document processing, size M: a two-page invoice. Move any of these and the figures follow; a new task or size brings its own.
Each call re-reads the task: about 5.5k tokens a call
10 extra calls for every 100 planned
1 in 5 documents, about 2 minutes each
Output accepted after the system's own retries
10 minutes a document, at $60 an hour: the same rate as review. 160 hours a month is one person full time.
- Each task
- 2 pages at about 2,300 tokens each (text plus an image of the page); the extract and check calls both read every page.
- Tokens a task
- 17k in and 750 out, at $4 and $20 per million. A token is about ¾ of a word (OpenAI, 2026).
- Cloud infrastructure
- $0.004 a model call: compute, retrieval, storage, gateway.
- Review
- 2 minutes a review, at $60 an hour.
- By hand
- 10 minutes a document: reading at 238 words a minute (Brysbaert, 2019) plus the time to key the fields or write the reply. One person full time is 160 hours a month: 2,087 paid hours a year (OPM, Oct 2026) less holidays and leave. Not counted: hiring, training, managing, and equipping people.
- Arithmetic
- Cost per successful task = (model + infrastructure + retries + review) ÷ successful tasks. Not counted: caching discounts, evaluation runs and monitoring, building the system, and the engineers who keep it running.
Sources: model list prices from Anthropic, OpenAI, Google, Together AI and Mistral, checked 27 Oct 2026. A PDF page is read as text plus an image of the page (Anthropic, 2026). Everything else on this panel is an illustrative assumption.
Blank uses the figure shown. Prices are per million tokens; a token is about ¾ of a word (OpenAI, 2026).
- Cost per successful task
- $0.56
- Monthly operating cost
- $5,023
The planned calls cost $0.08 a task. Each task that actually succeeds costs $0.56. Fully loaded, that is $5,023 a month.
- Model usage. Above the waterline, what you budgeted:$810
- Cloud infrastructure. Below the waterline, not in the budget:$120
- Retries and loops. Below the waterline, not in the budget:$93
- Human review. Below the waterline, not in the budget:$4,000
Human review: a person checks 1 in 5 documents, about 2 minutes each. Review falls as the success rate holds up; see the levers below.
Your task looks different? Tell us about it. An engineer works out what yours really costs.
Talk to us about your taskIllustrative · assumptions shown
Moving a pilot to production "nearly quadruples" AI spend, and a poorly designed agent can use up to 30× the tokens of a well-designed one on the same task (McKinsey, 2026). Gartner expects at least half of GenAI projects to overrun budget through 2028 (Gartner via Campus Technology, Jun 2026).
AI economics starts with the number that matters.
What AI really costs, the number that matters, and how engineering changes it.
01 · The full cost
Model spend is only the starting point. Inference pricing is one part of the equation. Retries, infrastructure, orchestration, evaluation, and human oversight all contribute to operating cost.
02 · The right metric
Measure successful outcomes. Cost per successful task reveals what it takes to deliver an accepted result, including failures and human intervention.
03 · The engineering lever
Change the system, not just the model. Identify unnecessary calls, inefficient workflows, excessive retries, and expensive infrastructure. Engineering changes should address the cost drivers that matter.
The system you'll operate.
From your cloud telemetry and model traces. Recorded in terms you can inspect.
- Unit economics
- Cost per successful task
- Model usage
- Model, tokens, and inference cost
- Success rate
- Accepted outputs against attempted tasks
- Human intervention
- Tasks requiring review or correction
- Retry overhead
- Failed attempts and additional model calls
- Latency
- End-to-end response time
- Infrastructure
- Cloud, storage, and orchestration cost
- Fully loaded monthly operating cost
- Total measured operating expense
Baseline: measured against an agreed workload, volume, and acceptance criteria. Measurement period: recorded across the agreed production measurement window. Data sources: cloud billing, model usage, application traces, evaluation results, and workflow records. Delivery is measured separately, against your own git and tracker history — see the Sprint.
Measurement finds the cost. Engineering moves it.
Each layer of the iceberg has a lever. The Review identifies which ones apply to your system. The Sprint builds them.
| Cost driver | What engineering changes |
|---|---|
| Expensive model calls | Route suitable tasks to smaller models. Reserve stronger models for tasks that need them. |
| Agent loops and retries | Bound execution, remove unnecessary steps, and control retry behavior. |
| Oversized context | Reduce redundant context and retrieve only relevant information. |
| Human review overhead | Improve evaluations and confidence-based escalation. |
| Excessive tool calls | Consolidate operations and remove unnecessary orchestration. |
| Evaluation and monitoring | Focus checks on the behaviors and failures that matter. |
| Cloud infrastructure | Right-size compute, storage, and supporting services. |
| Caching and reuse | Avoid repeating work where results can safely be reused. |
A dashboard shows you the first column. Engineering is how you fix the second.
How the numbers stay honest
Production workloads.
Only work executed in production counts. Prototypes and test workloads are labeled separately.
Agreed baselines.
Baselines come from your systems. You review and sign off before optimization begins.
Source data.
Numbers come from telemetry, not recollection. We use billing records, traces, evaluation results, and workflow data.
Visible assumptions.
Ask us to walk through any number. We will show the calculation and underlying assumptions.
Your data.
Your figures remain yours. We publish engagement results only with written client permission. Internal builds are labeled separately.