We all want to use AI, right? But in which cases does it truly add value? Is it only good for building chatbots? In this talk, we’ll examine situations where deterministic code works perfectly and those where it reaches its limit or its cost outweighs the benefit. We’ll show how, in these scenarios, generative AI can integrate with what already works, simplify your product, and enhance the quality of existing features. We’ll also explore how some current UX patterns lose relevance and force us to rethink new ways of interacting in our applications.

We’ll present three use cases that any modern app can adopt by 2025:
- Search interfaces with filters and semantic search: when to stick with deterministic filters and when to migrate to prompts and RAG for more contextual results.
- Data entry in forms (or bulk/batch processing): how AI can autocomplete or validate large volumes of data, improving the user experience while leveraging existing validation rules.
- Predefined task generator from natural language: how AI can interpret natural-language instructions to generate workflows that are then executed deterministically on the backend.

In this talk, we’ll begin by reviewing scenarios where deterministic programming works optimally: simple filtering, precise calculations, and straightforward validations. Next, we’ll identify its limitations: tasks requiring semantic interpretation or where the cost of maintaining rigid logic becomes prohibitive.

From there, we’ll demonstrate how generative AI excels in those complementary areas. We’ll present three ready-to-implement use cases for 2025:
- Search interfaces with filters and semantic search: We’ll explain the traditional filter flow (checkboxes, ranges, dropdowns) and why, in some contexts, it becomes insufficient. We’ll show how to design a prompt for semantic search and how to scale it into a RAG pipeline—combining it with a vector index (e.g., FAISS or Pinecone). We’ll include OpenAI-format code snippets and examples using the Google Gemini API to enrich queries.
- Data entry in forms (or batch processing): We’ll illustrate how, when entering data manually or in bulk (e.g., via CSV), deterministic validation can be tedious and error-prone. We’ll propose ideas to simplify and speed up that entry, letting humans focus on validation rather than manual “data entry.” We’ll show how AI can leverage existing deterministic validation code, which must still run in the backend to ensure correctness.
- Predefined task generator from natural-language instructions: We’ll present an example where a user types a complex instruction (e.g., “Create a launch plan for the new version, assigning resources by role and deadlines”). We’ll analyze how AI translates that instruction into a structured set of tasks (in JSON or YAML) that is then executed deterministically to create events, notifications, or pipelines. We’ll demonstrate using the Google Gemini API and show how to orchestrate execution in a real environment.

Throughout each section, we’ll discuss decision criteria: when it’s preferable to stick with deterministic code, when to use a simple prompt, and when to evolve to a RAG pipeline. We’ll compare performance, usability, maintenance cost, security considerations, and data-consistency metrics.

Finally, we’ll share code examples in JavaScript/TypeScript that illustrate integration with the OpenAI API (https://platform.openai.com/docs/libraries) and Google Gemini for more advanced and large-scale data use cases.