Definition
A valid AI use case is a digital process characterized by dynamic, high-dimensional, or unstructured
inputs and outputs where deterministic solutions or narrow, task-specific probabilistic models are
unlikely to achieve acceptable performance, reliability, or scalability.
Use Case Discovery Framework
- What problem are we trying to solve?
- What’s the business impact of solving it successfully?
- What is the user story (e.g., “As a user I want X so that Y”)?
- Is it (or can it be) a digital process?
- What does the typical input look like?
- What does the expected output look like?
- Can it be reliably solved using traditional software or business logic?
- Can it be solved with a task-specific ML model?
- What is the technical landscape (e.g., stack, vendor lock-in)?
- What is the data landscape (e.g., availability, quality, format)?
- Are there any constraints (e.g., legal, regulatory, privacy)?
- What is the readiness level (e.g., resources, access, budget)?
GenAI Solution Pathfinder
- Can a LLM solve it out of the box?
➡️ Use as-is *(zero-shot)
- Are instructions alone enough to guide the model’s behavior?
➡️ Use prompt engineering *(role-based, few-shot, CoT, etc.)*
- Can we solve it by providing relevant context?
➡️ Use Retrieval-Augmented Generation *(RAG)*
- Do we need multi-step decision logic or tool use (e.g., search, APIs, code execution)?
➡️ Use agentic frameworks (e.g., Agents SDK, LangGraph)
- Do we need to adapt the model’s behavior (e.g., domain adaptation)?
➡️ Use fine-tuning
- Is the problem still too vague or misaligned?
➡️ Rescope the problem