Main Concept
In the era of easy access to AI tools — especially Generative AI — there is a tendency to want to solve every problem with Machine Learning. However, many problems are deterministic or can be solved with standard programming and mathematics, where using ML is over-engineering a solution that does not need it.
Two Clear Cases Where ML Is the Wrong Choice
Deterministic Problems
Problems where the correct answer can always be computed exactly using a formula, algorithm, or rule — not estimated or predicted.
In these cases, ML models are counterproductive because they produce approximations, not exact answers. Supervised learning, unsupervised learning, and reinforcement learning all learn statistical patterns — they will get you close, but not always exactly right.
Key Idea
Deterministic problem → there is one correct answer, always computable the same way.
ML output → a probabilistic approximation, not a guaranteed exact answer.
Rule → if the problem always needs an exact answer, write code — do not train a model.
Examples of deterministic problems
“A deck has 5 red, 3 blue, and 2 yellow cards. What is the probability of drawing a blue card?” → 3/10, always. A formula solves this perfectly.
“Calculate compound interest given an initial capital of $10,000 at 5% annual rate over 10 years.” → exact mathematical formula, no approximation needed.
A classic LLM failure: “How many letter r’s are in the word raspberry?” — many older models answered incorrectly. This happens because LLMs process tokens, not individual characters — they were never designed for exact character-level counting.
Cost-Benefit Analysis
Sometimes ML is technically possible but not economically justified. If the cost of building, training, and maintaining an ML solution exceeds the value it delivers — or if a simpler rule-based system works well enough — ML is not appropriate.
Key Idea
Ask first → does the complexity and cost of ML justify the improvement over a simpler solution?
If a simple rule works → use the simple rule.
ML is a tool, not a default → the goal is solving the business problem, not using the most sophisticated technique.
Example
A business wants to flag orders over $10,000 for manual review. This is a simple threshold rule — it is always exact, always explainable, and requires zero training data. Building an ML model for this would be over-engineering with no benefit.
Why This Matters for Generative AI Specifically
Although modern LLMs have increasingly strong reasoning capabilities, they are still probabilistic systems — they predict the most likely response, not the mathematically correct one. For problems that require precision, an LLM is the wrong tool regardless of how capable it has become.
Key Idea
LLMs → excellent for language, reasoning, summarization, generation — probabilistic by nature.
Exact computation → use code, formulas, or deterministic algorithms.
Hybrid approach → use an LLM to write the code that solves the deterministic problem.
Exam Scope
This concept is explicitly tested. The exam will present scenarios and ask you to determine whether ML is appropriate. Watch for:
- Scenarios requiring an exact, deterministic answer → ML is not appropriate.
- Scenarios where a simple rule or formula already works → ML is not appropriate.
- Scenarios where cost outweighs benefit → ML is not appropriate.
- Scenarios with predictions, patterns, or large data → ML is appropriate.
Exam Domain
- Domain 1, Task Statement 1.2: “Determine when AI/ML solutions are not appropriate (for example, cost-benefit analyses, situations when a specific outcome is needed instead of a prediction).”