Saltar al contenido
APFerrer
Technical architecture

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.

Seven levels

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.

  1. Determinism

    Dates, calculations, permissions, formats, identifiers, amounts. Code with boolean logic. Zero ambiguity, zero probability. Exactness.

    When · Whenever possible.

  2. 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.

  3. 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.

  4. 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.

  5. 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.

  6. 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.

  7. 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.

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.