What needs to happen at the end?
“We need AI” is the start of a conversation. To choose a solution, describe the work in one sentence: an employee looks up a procedure, a manager approves a request, or a team summarizes a document. Then ask which part requires understanding information and which part requires taking action in a system.
Search, answer generation and triggering an action are different capabilities. They can be combined, but each one has its own way of measuring quality and its own permission boundaries.
When you need an answer grounded in knowledge
RAG supplies the model with information retrieved from knowledge sources to help it generate an answer grounded in that information. The challenge also includes retrieval quality, source freshness and access permissions. A source that was never retrieved cannot support an answer.
Start with a set of real questions and data you are allowed to use. Check separately whether the right sources were retrieved and whether the answer actually relies on them. If the requirement is only to find a document, a good search interface may be enough.
- Can you identify the source and when it was last updated?
- Does each user see only the information they are authorized to see?
- What is shown when no source supports an answer?
When you need to move a process forward
If every step of a process is known in advance, you can start with explicit automation: conditions, input validation, action and logging. A model can help with a well-defined part, such as classifying text, while the system stays in control of the actions.
The ability to choose tools and steps flexibly can be useful when there are several ways to reach the goal. Before handing that ability to an agent, define which actions are allowed, what requires approval and how a failing process is stopped.
A small experiment with clear measurement
Choose a bounded process and a set of examples that includes edge cases. Measure an outcome tied to the work: did the user reach the information, was the action performed correctly, and which errors require human handling.
Set stop and continue criteria in advance. Deciding to stay with a simple process is also a useful outcome of a feasibility check. The goal is to improve work that can be defined and tested.