How can Artificial Intelligence help me?

We asked this question on LinkedIn recently, and the answer was: First, help yourself by preparing your business, that is, your data.

But this needs to be worked on and contrasted with the types of projects that can be improved or automated with AIs. That’s why we’re listing a few types of projects… to give ideas and start pulling the thread.

Key Idea 1: You have to work on the data. What you have, how you have it, how to enrich it, what you want to achieve with it.

Key Idea 2: General models don’t solve your specific problems well. Make your own model, for you, with your data.

Let’s get to it:

Some types of projects that benefit from AI

1. General Content Generation:

This is the most well-known, and it’s what LLMs like GPT-4, Bard, etc., do. They generate content from a prompt. It can be text, images, music, etc.

Techniques: GPT-4, Bard, right out of the box.

• Example: Generate product descriptions for an online store.
• Example: Generate text or images for marketing.

2. Specialized Content Generation:

This one is more interesting because you “augment” the efficiency of one of the models with your data.

Techniques: LLMs plus RAG, fine-tuning, or distillation.

• Example: Generate product descriptions for an online store with your data.
• Example: Generate documentation for your projects according to your standards.

3. Data Classification:

Classifying or labeling data is a task that can be very tedious and very expensive if done by people. Labeling data with an LLM (GPT-X) is cheaper, but keep Key Idea 2 in mind. If you have your own dataset (customer complaints from your bank), you augment it, refine it, and manage to label it well, you have a treasure with which to train your own classification system.

Techniques: BERT, DistilBERT, Pretraining in a domain, rule-based labeling, RAG, fine-tuning or distillation, mechanical turk.

• Example: Classify your customer complaints based on their content.
• Example: Classify your store’s products based on their description.
• Example: Fraud detection in transactions.

4. Recommendation Systems and Shopping Baskets:

Recommending products to your customers based on their behavior, tastes, and needs. If you’ve bought potatoes and beer, I recommend you buy olives, based on the behavior of other customers.

Techniques: Shopping basket algorithms, clustering, etc.

• Example: Recommend asset purchases to a broker’s clients.

5. Identification of Winning Features:

Identifying the features that make one product better than another.

Techniques: Regression models, classification models, clustering models, UMAP, PCA.

• Example: Identify from a pool of debtors who are most likely to default or repay their credit.
• Example: Identify the features that make one financial product better than another.

Other examples of AI projects:

1. Anomaly detection in time series.
2. Time series prediction.
3. Conversational agents (supported by some of the other systems mentioned).
4. Adaptations of general systems to the individual.

Conclusion

The future is generative AIs, okay, but:

1. As part of a whole. To multiply numbers, the tool is a calculator, not Bard.
2. Specialized with your data and for your needs. RAG or fine-tuning, or both.
3. Especially for those who work well with their data and know how to label and classify it.
4. CHOOSE AND DEFINE YOUR USE CASE WELL AND WHAT IT WILL BRING TO THE COMPANY.

Photo by Elaine Bernadine Castro