Not everything needs an LLM: a frame for the evaluation when AI makes sense

Ask: Which product should use machine learning (ML)?
Project manager answer: Yes.

Apart from the jokes, the accommodation of the generative AI has given our understanding of the best for ML. Historically, we always used ML for a fog Repeatable, predictive patterns In customer experiences, but now it is possible to use a form of ML even without an entire training data set.

The answer to the question “What customer needs does a AI solution require?” It is still not always “yes”. Great -speaking models (Llms) can still be unaffordable for some, and as with all ML models, LLMs are not always exactly exactly. There will always be applications in which the use of an ML implementation is not the right way. How do we, as an AI project manager, rate the needs of our customers for AI implementation?

The most important considerations that make this decision are:

1. The entries and outputs that are necessary to meet your customer’s requirements: The customer is provided to your product and the output is provided by your product. For a Spotify-ML-generated playlist (one edition), inputs can contain customer preferences and ‘liked’ songs, artists and music genre.
2. Combinations of inputs and outputs: Customer needs can vary depending on whether the same or other output is desired for the same or different inputs. The more permutations and combinations we have to replicate for inputs and outputs, the more we have to contact ML-opposite-based systems.
3. Patterns in entrances and outputs: Patterns in the required combinations of inputs or outputs will help you to decide which type of ML model you have to use for implementation. If the combinations of inputs and outputs contain patterns (e.g. review of customer anecdotes to derive a mood assessment), you should consider supervised or semi-supported ML models via LLMs, since they may be more cost-effective.
4. Costs and precision: LLM calls are not always cheap on a scale and the expenses are not always precise/accurateDespite fine -tuning and faster engineering. Sometimes they are better off with supervised models for neural networks, which can classify an input with a fixed set of labels or even rules based instead of using an LLM.

I have put together a short table below to summarize the above considerations in order to support project managers in assessing their customer needs and determine whether an ML implementation appears the right way forward.

Type of customer needs	Example	ML implementation (yes/no/dependent)	Type of ML implementation
Repeated tasks in which a customer requires the same issue for the same input	Add my e -mail using different forms online	NO	Creating a regulatory -based system is more than sufficient to help you with your expenses
Repeated tasks in which a customer requires different expenses for the same entry	The customer is in “recognition mode” and expects a new experience if he took the same measures (e.g. registration in an account): – Generate a new work of art per click –Dumpling (Do you remember it?) Discover a new corner of the Internet through random search	Yes	–Image generation llms – RecoMMendation algorithms (collaborative filtering)
Repeated tasks in which a customer requires the same/similar output for different inputs	–Arging essays – Creating topics from customer feedback	Depends on	If the number of input and output combinations is simple enough, a deterministic, regulatory-based system can still work for you. However, if you have several combinations of inputs and outputs, since a regulatory system cannot effectively scale, you should support yourself: – classifier –Topic modeling But only if these inputs have patterns. If there are no patterns at all, you should consider LLMs, but only for unique scenarios (since LLMs are not as precise as monitored models).
Repeated tasks in which a customer requires different expenses for different entries	– Questions about customer service support -Seek	Yes	It is rare for you to encounter examples where you can provide different outputs for different inputs in scale without ML. There are simply too many permutations for a rule -based implementation to scale effectively. Hold: –Ilms with access generation (rag) –Decision trees for products such as search
Non -repetitive tasks with different outputs	Evaluation of a hotel/restaurant	Yes	In advance, this type of scenario was difficult without doing models that were trained for certain tasks, e.g. B.: – Recurrent Neural Networks (RNNS) –Long short-term memory networks (LStMS) to predict the next word LLMS fits well with this type of scenario.

Conclusion: Do not use a lightsaber if a simple scissors can make the trick. Evaluate your customer’s needs using the above matrix, taking into account the implementation costs and the accuracy of the output in order to build precise, inexpensive products on a scale.

Sharanya Rao is a product manager for fintech group. The views expressed in this article are those of the author and not necessarily those of their society or organization.