Skip to content

Chapter 0.7 — What AI does well (and what it shouldn't control on its own)

🎯 Chapter objective

Map capabilities and limits at a high level, with a practical criterion to separate a task that is "natural for AI" from one that "needs deterministic software around it".

🧠 Core concept

The right question is rarely "can AI do this?". The right question is "can the typical error of this task be absorbed by the system without becoming an incident?". When the answer is yes, AI tends to be the best productivity choice; when it is no, there is almost always something deterministic that must sit between the model and the action.

AI tends to do well:

  • Recognizing patterns in high-dimensional data.
  • Handling semantic ambiguity.
  • Interpolating within the training domain.
  • Personalizing based on signals.
  • Summarizing, classifying, translating, generating text, extracting entities.
  • Suggesting, ranking, recommending.

AI should not control on its own:

  • Authorization (who can do what).
  • Exact financial calculation.
  • Compliance rules.
  • Transactional consistency.
  • Concurrency control.
  • Auditing.
  • Irreversible decisions without review.

None of this is absolute: AI can assist all of these areas (review a calculation, suggest a compliance decision, propose the next action). The red line is letting the model decide on its own with immediate effect in the world, without deterministic mediation.

The practical rule:

Use AI for ambiguity. Use deterministic software for commitments, rules, permissions, calculations and side effects.

📚 References