LLMs are often seen as a universal solution, but in many cases they are the wrong tool for the job. As companies rush to integrate AI into their products, they often overlook situations where deterministic systems are more reliable, faster, and cost-effective. This webinar explores how to distinguish between problems that truly require AI and those where it only adds unnecessary complexity.
Speaker
Nikolay Dolgov is an AI Engineer at The Lightning Group, where he focuses on applying large language models in real product environments and making them work under production constraints.
Not Every Problem Needs AI
Product problems can be divided into two categories: those with a clearly correct answer and those that require judgement.
AI is most effective in the second category, where ambiguity exists. When tasks require precision, predictability, and strict correctness, traditional deterministic systems are often the
Where LLMs Underperform
LLMs struggle in scenarios that require exact computation, structured data lookups, fast response times, compliance and auditability, or repetitive logic with low variation.
In these cases, using AI introduces additional risk and complexity without delivering meaningful value.
Where LLMs Work Best
LLMs perform well in tasks involving ambiguous natural language, summarisation and synthesis, soft classification, content generation, and handling edge cases that are difficult to define with rules.
These are problems where flexibility is more important than strict correctness.
A Framework for Decision-Making
To determine whether AI should be used, four key questions can be applied:
Is the answer deterministic?
What is the cost of a mistake?
How fast does the response need to be?
How variable is the input?
These questions help evaluate risk, performance requirements, and the suitability of AI.
Three AI Usage Zones
AI systems can be structured in three main ways:
Zone 1 — Fully deterministic systems, where rules, functions, or SQL handle all logic without any AI.
Zone 2 — Model at the edge, where AI processes unstructured input, while core logic remains deterministic.
Zone 3 — Model at the core, where AI handles most of the logic and human validation is applied afterward.
Choosing the right zone depends on the nature of the problem and the acceptable level of risk.
Why Hybrid Systems Work Best
In practice, the most effective approach is often a hybrid architecture. AI is used where it adds value, while deterministic logic ensures reliability and control.
Rather than replacing existing systems, AI should complement them.
Conclusion
LLMs are not general-purpose solutions — they are tools suited for specific types of problems.
Understanding when not to use AI is just as important as knowing when to use it. By evaluating determinism, risk, latency, and input variability, teams can build systems that are both efficient and reliable — often by combining AI with traditional approaches.