One task, one tool.
OCR, extraction, classification, reasoning and drafting are not the same. Each has a tool with a better cost, latency and quality ratio. Picking the same frontier model for everything means paying a lot for results that a dedicated component delivers better.
Which tool for each case, and why
- Task
OCR of invoices and delivery notes
Tool: Dedicated component (Datalab, Textract, Azure Document Intelligence).
Extracting tables and positional fields from scanned PDF is solved by specific tools, with better precision and a fraction of the cost of a frontier model doing the same.
- Task
Semantic document classification
Tool: Lightweight API model with short prompt and examples.
Detecting whether a document is invoice, contract, memo or ruling needs language understanding but not deep reasoning. A lightweight model costs cents per thousands of calls.
- Task
Free-field extraction
Tool: Lightweight API model with structured schema in the prompt.
Extracting entities with no fixed position (amounts, deadlines, references, clauses) works well with lightweight models when the schema is clear. Frontier only if there is real ambiguity.
- Task
Legal or technical reasoning
Tool: Frontier model with mandatory human review.
Interpreting a tender, cross-checking clauses against case law, or proposing a legal strategy needs frontier models. And human sign-off on the output is not optional where responsibility lies.
- Task
Assisted drafting
Tool: Frontier model with own templates and cited sources.
Generating reports, executive summaries or commercial proposals works well on frontier models when built from templates and internal data, not from thin air. Full traceability via citations.
In SVP each level of the Semantic Scale uses the tool that fits that task
Architecture separates tasks and assigns the tool with the best cost-quality ratio. Lower cost per call, higher precision, and the ability to swap any piece without rewriting the system.