Semantic Scale
Determinism where certainty is required. Semantics where understanding is needed. Power only where it pays off.
An AI model is not the base of an enterprise system. It is one tool inside the architecture. The right question is not whether to use AI. It is at which point in the flow AI delivers the highest value against cost. The Semantic Scale answers that with seven processing levels, and each task rises only to the level it needs.
Each task rises only to the level it needs
No dogma. A dictionary beats a model when the criterion is fixed. A lightweight API model beats a frontier one when just interpreting is enough. And some decisions do not go to any model. Architecture picks the level per task, not per trend.
Determinism
Dates, calculations, permissions, formats, identifiers, amounts. Code with boolean logic. Zero ambiguity, zero probability. Exactness.
When · Whenever possible.
Dictionaries and structured knowledge
Equivalences, catalogs, regulatory codes, sectoral terms. Not AI, just table lookups. Scales better than rewriting the criterion.
When · When the criteria already exist and do not change.
Dedicated components
OCR, embeddings, reranking, anonymization, specialized classification. Point tools with a clear goal. Not a general-purpose LLM.
When · When there are proven open components for that task.
Lightweight API models
Semantic classification, flexible extraction, normalization, short summaries. Low-cost, manageable-latency models from the major providers.
When · When the task requires understanding language but is not complex.
Frontier models
Cross-analysis, ambiguity, complex drafting, technical or legal reasoning. Significant cost. Medium latency.
When · When the problem is genuinely complex and the responsibility justifies it.
Frontier model with human review
The same as level 5, with an expert validation step. The model proposes, the person resolves.
When · When legal or reputational responsibility is at stake.
Pure human review
No model. Expert interprets, decides, approves. The system provides context, references, data. The decision is human.
When · Professional interpretation, non-probabilistic decisions, full responsibility.
Frequently asked questions
What is the Semantic Scale?
A seven-level technical architecture that assigns each task to the lowest-total-cost mechanism able to solve it: determinism, dictionaries, dedicated components, lightweight models, frontier models, frontier model with human review and pure human review.
How many levels does it have?
Seven. From deterministic code (level 1) to pure human review (level 7). Each task rises only to the level it needs, no more.
Why use levels instead of a single model?
One model for everything blows up cost and drops precision. A dictionary beats a model when the criterion is fixed. A lightweight API model beats a frontier one when just interpretation is enough. Architecture picks the level per task.
Can the level assigned to a task be changed?
Yes. Configuration lives in a system file, not in every call. Changing a task's default level is one config line, not a code refactor.
Is the Semantic Scale applicable to any AI workflow?
Yes, it is agnostic to sector and model vendor. It is used as a design criterion in every SVP implementation and also works as an audit framework for existing AI systems.
The model is not the system. The system is the architecture that decides what to use at each step.
The Semantic Scale is the internal architecture used to deliver Own Vertical Software. If you want to see how it applies to a specific case, let us talk 30 minutes with no commitment.