USD 0.0006 per AI scan: GenAI cost

The question “how much does generative AI cost” usually gets a vague answer. We have our own production number: USD 0.0006 per AI scan in mojApteczka. But that figure is only the tip — the real cost of a GenAI implementation sits elsewhere. Let’s break it down, on production data.

The number: USD 0.0006 per scan

That is the cost of one AI scan in production — what the system pays when the AI extracts structured data from a drug package. At that unit cost, ongoing inference is usually not the main cost barrier — the bigger risk sits in data, quality and operations.

Where the cost really is: before the model, not in it

The most honest lesson from building mojApteczka: the biggest work was before the model, not inside it — ETL, cleaning and versioning the knowledge. 302,516 records of drug interactions in the production knowledge base did not come from a prompt. This is the dominant, usually hidden cost of GenAI: data engineering, not tokens.

The three layers of GenAI cost

Inference (ongoing model calls) — the smallest line item; it sits inside the per-scan cost (USD 0.0006).
Data and knowledge (one-off and ongoing) — ETL, cleaning, embeddings, versioning. This is the bulk of the budget.
Operation — storage, retrieval (RAG), monitoring, quality evaluation. A fixed cost that grows with scale.

A budget built on line item 1 alone is incomplete and often too low — which hurts most after go-live.

TCO: the biggest cost comes after go-live

The third layer — operation — is not a one-off line item but the total cost of ownership (TCO) that builds up after launch. A system that answers correctly today drifts tomorrow: the vendor’s model changes, documents accumulate, questions evolve, rules get sharper. Keeping quality up costs money — and it’s the cost you don’t see in an implementation quote:

Evaluations — a golden set and regression on every change, so fixing one question doesn’t break ten (how to measure RAG quality).
Production monitoring — refusal rate, drift, user reports.
Governance and compliance — oversight, an audit trail and continuous AI Act compliance as systems and rules change (governance for regulated industries).

Leaving this layer out of the budget doesn’t lower the cost — it shifts it in time and usually raises it. That’s why we compute TCO over a horizon of months, not the first invoice.

Structure your own TCO model

Open the free GenAI TCO worksheet to put inference, retrieval, storage, data, evaluations, monitoring, operations and error cost in one place. You supply every value, while static formulas preserve the source of each assumption. A separate section captures audit scope but does not calculate a price or publish Semitora rates.

A cheap scan that gets it wrong is expensive

A low unit cost only matters with controlled quality. USD 0.0006 is therefore not standalone evidence of effectiveness: read it alongside a versioned validation set, a scoring definition and per-case results. We withdrew the earlier extraction metric because the public material did not retain a reproducible artifact.

Optimizing AI costs: how to lower GenAI cost without losing quality

Match the model to the task — a smaller model where it fully suffices.
Cache and batch for repeated calls.
RAG instead of long prompts — shorter context means fewer tokens per call.
Managed AWS services (e.g. Amazon Bedrock) instead of running models and infrastructure yourself.
Measure the unit cost from day one, not after the fact.

What an AI implementation costs for you

Your unit cost will be different — it depends on the task, the model and the state of your data. The structure is usually similar: tokens are the minority, data is the majority. The worksheet gives you an initial assumptions model. If you need an audit, we first confirm the process, systems, data, risk and expected deliverables; only then do we prepare the scope, schedule and individual quote. If you are weighing a cheap price list against a production system, see what you buy at each level: cheap chatbot vs production system.

What’s next

mojApteczka is our production proof: a June 2026 operating snapshot contained a USD 0.0006 inference-layer cost per scan and 302,516 records in the knowledge base. We disclose the scope and the withdrawn metric in the case methodology. See how it works or read the case study. To learn the cost for your own process, get in touch.