Get the latest insights and exclusive content on cutting-edge AI coverage by subscribing to our daily and weekly newsletters. Find out more
Query: Which products can benefit from machine learning (ML)?
Response from a project manager: Definitely.
With the rise of generative AI, our traditional understanding of suitable ML use cases has been disrupted. Previously, ML was primarily used for predicting patterns in customer interactions, but now, even without a complete training dataset, ML can be utilized.
However, the decision to implement an AI solution for customer needs isn’t always straightforward. Large language models (LLMs) can be costly for some, and their accuracy is not guaranteed. There are instances where opting for an ML implementation may not be the most appropriate choice. How can AI project managers assess the necessity of AI implementation for their customers?
Key factors to consider when making this determination include:
- The input and output requirements to meet customer needs: An input is what the customer provides to your product, and the output is what your product delivers. For example, in a Spotify ML-generated playlist (output), inputs could consist of customer preferences and liked songs, artists, and music genres.
- Variances in inputs and outputs: Customer requirements may differ based on whether they expect the same or different outputs for the same or different inputs. When there is a need to replicate numerous input and output permutations, ML becomes more favorable over rule-based systems.
- Patterns in inputs and outputs: Identifying patterns in input-output combinations helps determine the type of ML model to employ. If there are discernible patterns in the combinations (e.g., analyzing customer anecdotes for sentiment scoring), consider using supervised or semi-supervised ML models instead of LLMs due to potential cost savings.
- Cost and Accuracy: Large language model calls can be expensive at scale, and their outputs may not always be precise, despite optimizations. In some cases, opting for supervised models for neural networks that can classify inputs using a fixed set of labels or rule-based systems may be more effective than utilizing LLMs.
Below is a summarized table outlining the considerations mentioned above, aiding project managers in evaluating customer needs and determining whether an ML implementation is the right course of action.
| Type of customer requirement | Example | ML Implementation (Yes/No/Depends) | Type of ML Implementation |
|---|---|---|---|
| Repetitive tasks where a customer requires the same output for the same input | Add my email to various online forms | No | Creating a rule-based system suffices for generating outputs |
| Repetitive tasks where a customer requires different outputs for the same input | The customer is in “discovery mode” and anticipates a new experience with each action (e.g., signing into an account): — Generate a new artwork per click — StumbleUpon (remember that?) exploring the internet through random searches | Yes | – Image generation LLMs – Recommendation algorithms (collaborative filtering) |
| Repetitive tasks where a customer requires the same/similar output for different inputs | – Grading essays – Extracting themes from customer feedback | Depends | If the input-output combinations are straightforward, a deterministic rule-based system can suffice. However, for complex combinations, consider: – Classifiers Only opt for LLMs if patterns are present, and if not, reserve them for unique situations as they may lack precision compared to supervised models. |
| Repetitive tasks where a customer requires different outputs for different inputs | – Responding to customer support queries – Search functionality | Yes | It’s challenging to cater to varying outputs for distinct inputs at scale without ML. Rule-based systems struggle to handle the numerous permutations efficiently. Consider: – LLMs with retrieval-augmented generation (RAG) |
| Non-repetitive tasks with diverse outputs | Reviewing a hotel or restaurant | Yes | Before LLMs, accomplishing such tasks without task-specific models like: – Recurrent neural networks (RNNs) LLMs are ideal for these scenarios. |
Ultimately, avoid using a lightsaber when a simple pair of scissors will suffice. Utilize the matrix provided above to assess your customer’s needs, factoring in implementation costs and output accuracy to develop efficient and precise products at scale.
Sharanya Rao serves as a fintech group product manager. The opinions expressed in this piece are those of the author and do not necessarily reflect those of their respective company or organization.
