Join our daily and weekly newsletter for the latest updates and exclusive content on industry-leading AI coverage. learn more
question:What products use machine learning (ML)?
Project Manager Answers: yes.
Joke aside, the emergence of generative AI has encouraged an understanding of which use cases are best for ML. Historically, we have always used ML for repeatable predictive patterns in customer experiences, but we can now leverage the format of ML without the entire training dataset.
Nevertheless, the answer to the question “Does a customer need an AI solution?” Still, it’s not always “yes.” Large-scale language models (LLMS) can still be very expensive for some, and like all ML models, LLM is not always accurate. There are always use cases where leveraging ML implementations is not the right way. As an AI Project Manager, how do you assess the customer’s needs for AI implementation?
Key considerations for making this decision include:
- Inputs and outputs required to meet customer needs: Inputs are provided by the customer to the product, and outputs are provided by the product. Therefore, for playlists (outputs) generated in Spotify ML, the input includes customer preferences, “Liked” songs, artists, and music genres.
- Input and output combination: Customer needs may vary depending on whether they require the same or different outputs for the same or different inputs. As permutations and combinations increase, the more largely it is necessary to rely on ML for rules-based systems to replicate inputs and outputs.
- Input and output patterns: The required combination patterns of inputs or outputs can help you determine the type of ML model that you need to use in your implementation. If there is a pattern in the input and output combination (such as checking the customer’s anecdote to derive sentiment scores), consider a monitored or semi-surveillance ML model as it may be more cost-effective to monitor or semi-surveillance ML model via LLMS.
- Cost and accuracy: LLM Calls are not always large and cheap, despite their fine tuning and fast engineering, and the output is not always accurate/accurate. Sometimes, instead of using LLM, it may be better to use a monitoring model of neural networks that allows you to classify inputs using fixed labels or rules-based systems.
A brief table is summarized below, summarizing the above considerations to help project managers assess the needs of their customers and determine whether ML implementation appears to be the right path.
| Types of customer needs | example | ML implementation (yes/no/dependencies) | Types of ML implementations |
|---|---|---|---|
| Repeated tasks where customers require the same output for the same input | Add your email online to various forms | no | Creating a rule-based system is more than enough to aid in output |
| Repeated tasks where customers require different outputs for the same input | The customer is in “discovery mode” and expects a new experience when performing the same action (such as signing an account): – Generates new artwork with every click –StumbleUpon (Remember that?) Discover new corners of the internet with random search | yes | – Image Generation LLMS – Recommended algorithm (co-filtering) |
| Repeated tasks where customers require the same/similar output for different inputs | – Essay rating – Generate themes from customer feedback | It depends | If the number of input and output combinations is simple enough, a deterministic rule-based system will continue to work. However, if you start starting multiple combinations of input and output because your rule-based system cannot scale effectively, consider tilting to the following: – device However, only if these inputs have patterns. If there are no patterns at all, consider leveraging LLMS, but only one-time scenarios (as LLM is not as accurate as monitored models). |
| Repeated tasks where customers require different outputs for different inputs | – Avoid customer support questions -search | yes | It is rare to come across examples where you can provide different outputs for different inputs of scale without ML. There are too many permutations for rule-based implementations to scale effectively. Consider: –llms with retirementval-augmented generation (rag) |
| Non-repetitive tasks with different outputs | Hotel/Restaurant Reviews | yes | pre-llms, this type of scenario was difficult to achieve without a model trained for a specific task, such as: – Iterative Neural Network (RNN) LLMS is perfect for this type of scenario. |
Bottom line: Do not use a lightsaber when simple scissors can do the trick. To build accurate and cost-effective products at scale, we use the above matrix to assess customer needs, taking into account the cost of implementation and output accuracy.
Sharanya Rao is the product manager of the Fintech group. The views expressed in this article are those of the authors and not necessarily those of their company or organization..
