Not everything needs an LLM: A framework for evaluating when AI makes sense
Join daily and weekly newsletters to obtain the latest updates and exclusive content to cover the leading artificial intelligence in the industry. Learn more
a question: What product should use machine learning (ML)?
The answer of the project manager: Yes.
Bank, aside, the emergence of the Truidian artificial intelligence has raised our understanding of what expansion cases that give the best in the ML. Historically, we have always benefited from taking advantage of ML for repeated predictive patterns in customer experiences, but now, it is possible to take advantage of the ML even without a full training data set.
However, the answer to the question, “What does the customer need require a solution to artificial intelligence?” It is still not always “yes”. LLMS models can still be expensive for some, and as with all ML models, LLMS is not always accurate. There will always be cases of use where the use of ML implementation is not the right path forward. How do we evaluate the managers of Amnesty International Projects, the needs of our customers to implement artificial intelligence?
The main considerations of helping this decision include:
- The inputs and outputs required to meet the needs of your customer: Entrance is provided by the customer for your product and the output is provided by your product. Therefore, for the created Spotify ML playlist (exit), the inputs can include customer preferences, “admiration” songs, artists and type of music.
- Groups of inputs and outputs: Customer needs can vary based on whether they want the same output or different for the same input or different inputs. The higher the number of exchanges and groups we need to repeat the inputs and outputs, the more we need to switch to ML against the rules -based systems.
- Patterns in inputs and outputs: The patterns in the required groups of inputs or outputs help you determine the type of ML model that you need to use for implementation. If there are patterns of input and output groups (such as reviewing customer stories to extract the degree of feeling), consider ML models under supervision or semi -subject to supervision of LLMS because they may be more expensive.
- Cost and accuracy: LLM calls are not always widely cheap and outputs are not always accurate/minute, despite installation and instant engineering. Sometimes, they are better with models under the supervision of nerve networks that can classify insertion using a fixed set of stickers, or even rules -based systems, instead of using LLM.
I have collected a quick schedule below, to summarize the above considerations, to help project managers assess their customer needs and determine whether the ML application looks like the right track forward.
| Customer type needs | example | ML implementation (yes/no/depends) | ML implementation type |
|---|---|---|---|
| Repeated tasks where the customer needs the same output for the same input | Add e -mail across different forms online | no | Create a system based on the rules is more than enough to help you in your outputs |
| Repeated tasks where the customer needs different outputs for the same input | The client in “Discovery Status” and expects a new experience when they take the same procedure (such as signing at an expense): Create a new artwork for each click —Sumbleupon (Remember that?) Discover a new angle from the Internet through random search | Yes | – The image of the LLMS generation – Recommendation algorithms (cooperative liquidation) |
| Repeated tasks where the customer needs the same directing/similar to the various inputs | Translation of articles Generating topics from customer notes | Dependent on | If the number of input and output groups is simple enough, the inevitable system on the rules can still work for you. However, if you start obtaining multiple groups of inputs and outputs because the rules -based system cannot expand effectively, then think about: relying on: Classifiers But only if there are patterns of these inputs. If there are no patterns at all, think about taking advantage of LLMS, but only for scenarios for one time (because LLMS is not as accurate as models subject to supervision). |
| Repeated tasks where the customer needs different outputs for different inputs | Customer support questions answer -Search | Yes | It is rare to encounters examples where you can provide different outputs for different inputs on a large scale without ML. There is a lot of exchange for the rules -based implementation to expand an effective scope. It is considered: -Lms with a generation for retrieval (rag) |
| Unconscious tasks with different outputs | Hotel/Restaurant Review | Yes | Before LLMS, this type of scenario was difficult to accomplish without models that were trained on specific tasks, such as: Repeated nerve networks (RNNS) LLMS is a great occasion for this type of scenario. |
The bottom line: Do not use Lightsaber when a simple pair of scissors can do the trick. Evaluate your customer’s need to use the matrix above, taking into account the costs of implementation and the accuracy of the output, to create accurate and effective products on a large scale.
Sharanya Rao is the director of the Fintech Group Product. The opinions expressed in this article are the views of the author and not necessarily the effects of their company or institution.
Don’t miss more hot News like this! Click here to discover the latest in AI news!
2025-05-03 19:35:00
