UK-based independent AI efficiency projectEvidence before claims · 2026

Token Optimisation

AI efficiency · measured
Discuss a diagnostic

LLM or Deterministic Tool? A Decision Framework for Recurring AI Work

Known exact rules route to deterministic code; interpretive tasks route to an LLM with verifier and escalation.

Written by

Token Optimisation

in

,

Classification: DECISION FRAMEWORK, not a measured contest between code and a model.

LLM or Deterministic Tool? Route by Uncertainty and AcceptanceDECISION FRAMEWORK · NO MEASURED WINNER. RULE IS EXACT? → DETERMINISTIC TOOL → MODEL FOR AMBIGUITY → VERIFY OR ESCALATE. A design framework. No paired production speed, quality, or cost comparison is available.DECISION FRAMEWORK · NO MEASURED WINNER1RULE IS EXACT?2DETERMINISTIC TOOL3MODEL FOR AMBIGUITY4VERIFY OR ESCALATEUse code for exact repeatable rules · Use interpretation only where needed
DECISION FRAMEWORK · NO MEASURED WINNER. A design framework. No paired production speed, quality, or cost comparison is available.

View the full infographic → · View this article

Use an LLM when the task genuinely needs language interpretation, synthesis or judgement under a stated rubric. Use deterministic code or a direct query when exact rules and structured inputs can answer it reliably. Then verify whichever path you choose.

Start from acceptance criteria

  • Exact lookup, arithmetic, schema validation or deduplication: prefer a deterministic operation when the source and rule are known.
  • Ambiguous language, open-ended synthesis or nuanced classification: a model may be appropriate, with examples and a review rubric.
  • High-impact or uncertain output: add an independent verifier or human gate regardless of the generator.

This is not a claim that deterministic tools always win. Bad source data, incorrect rules or brittle parsing can fail as clearly as an LLM. Compare accepted task quality, maintenance effort, retries, runtime and human review.

A staged routing pattern

  1. Try the simplest deterministic method that can satisfy the contract.
  2. Check output against explicit acceptance criteria.
  3. Escalate to a suitable model only when the deterministic route lacks needed interpretation or fails its defined check.
  4. Carry a concise verified summary into the stronger route rather than replaying unrelated context.
  5. Record model, provider, attempts, usage, acceptance and human handling for comparisons.

Token routing is one lever, but a cheaper model is adequate only if quality still passes. The token rules tool includes deterministic-first and quality-gated escalation rules. For cost scenarios, use the calculator and keep user inputs distinct from measured usage.

More posts