Predictive and Generative AI: A comparative study of its applications in Data Cloud

In this blog, we we will share the key use cases which harness the power of Predictive and Generative AI (Gen AI) and their usage in Data Cloud.

Let’s look at an example…

Consider an automotive company that can leverage both types of AI – Predictive and Generative, to enhance sales and customer service.

With Predictive AI, the automotive company can predict what parts of the car will have a higher probability to be sold in a certain region for a certain customer based on past data. This will help them target the right set of customers.

With Generative AI, the automotive company can:

1. create an assistive service agent, that, when a car breaks down in the middle of the road, is able to trigger a service case against the user’s profile to seek help. The trigger can be associated with any breakdown event.
2. look up the customer’s past service cases and profile data, to determine if the repair service is covered by warranty.
3. help the customer with personalized suggestions to schedule a service based on past service history, and likelihood of part failures.
4. help with basic quick repair suggestions based on old manuals from the company database.

Predictive and Generative AI – Understanding the underlying Models

Predictive and Generative AI are distinct, and the differences lie in the underlying algorithms and architectures behind them.

Machine Learning / Deep Learning models are based on Discriminative techniques. These techniques predict what is next based on conditional probabilities.

Generative models understand and replicate the complete underlying probability distribution of a dataset, learning how data is generated and allowing it to generate new data points similar to the training data.

Predictive AI in Salesforce Data Cloud

Salesforce Einstein Studio is a tool within Data Cloud used to implement Predictive AI. Let’s build a model using it.

Use Case: Continuing on our automotive company example, lets help predict the Probability of a Deal Win for a car part based on past data of deals won and lost based on brand, vertical and promotions it was sold with.

Training Data looks like this:

When test data comes in, we can now predict what is the Probability of Winning a deal for a Car Part with different input parameters. Here are the steps with Einstein Studio:

1. Make sure to collect the relevant and complete data. Einstein Studio needs atleast 400 records to build a model.
2. Ingest data via a connector (Data Cloud Setup > Other Connectors, a blob store like S3, GCS, Azure) or through data in Sales or Service Cloud)
3. Map the ingested record DLO to the appropriate standard or custom DMO.
4. Go to Einstein Studio > Add Predictive Model . Choose to build your model or bring your own model from Amazon Sagemaker, Google Vortex AI, Databricks.
5. Select the dataspace to build the model in and the DMO on top of which to build the model.
6. You can choose to bring in all of the data or filter incoming data.
7. Select a goal – In our case, it is whether we will “win” or “loose” the deal for a car part.
8. Select the variables which the model will use to predict the deal win probability.
9. Select an algorithm or have automatic selection pick the one that suits best.
10. Save and Train your data.
11. View your training metrics and use the win prediction for a sample customer.

Output: Below are some metrics that “Training Metrics” in Einstein Studio will show, to depict how well your model was trained:

1. Model Accuracy – for example a model accuracy of over 80% indicates the model is sufficiently well trained to predict on future outcomes.
2. False Positive / False Negative, True Positive / True Negative Percentages
3. Top predictors and recommendations- These help you take actionable next steps if you want to convert a deal that has a high probability to win.

Generative AI and it’s use in Salesforce Data Cloud

While Predictive AI focuses on forecasting, Generative AI takes a more creative approach. It learns patterns within data and generates new content, such as personalized marketing messages, customer service responses.

How do you use Generative AI in Data Cloud?

One of the ways to utilize Gen AI in Data cloud is to build an agent that answers questions based on your database.

Let’s say the same Automotive Company wants to answer basic customer queries automatically if the car breaks down and the drivers need immediate help. Using Data Cloud, we can build a Help Agent to answer such customer queries.

Here is a sample of how the pdf for an oil change for the car might look like:

Below are the steps to build this agent:

1. Collect the car repair manuals (Docs, PDF, Image etc.) and pick a blob store (S3, GCS, Azure) of your choice that you will use to bring these manuals into Data Cloud.
2. Create a connected app in Salesforce and set up a file notification pipeline. This helps Salesforce Data Cloud know, whenever new data gets added to the blob store destination.
3. Create a Data Lake Object in Data Cloud that points to the bucket/folder in your blob store.
4. Add your unstructured data (repair manuals) to your source blobs now and refresh your UDMO, to see records coming in.
5. Alternatively you could bring your data in via Agentforce Data Library, a search index and retriever will be automatically created within some time.
6. A directory table, vector and search index UDMO gets created for the ingested object, once you finish UDLO creation.
7. There will also be a RAG retriever set up automatically for your ingested data. Einstein Studio> Retriever is where you can find them.
8. You can now go ahead and use this retriever in your prompt, then build an action that utilizes the prompt.
9. Go to an agent > add a topic and give instructions to it > add the action you created in the last step.
10. This agent can now be activated and is good to go! Fire away queries related to car breakdowns and the agent will be able to answer your queries based on the information it learnt from the manuals you imported.

Cheatsheet for when to use what

Below is a table that highlights differences in where Predictive and Generative AI are used.

Conclusion

In this blog, we talked about the difference between Predictive and Generative AI. We also talked about what is possible with Salesforce Data Cloud, Agents and Unstructured data, in terms of these two facets of AI. As a next step, we recommend that you to try out building your model on Einstein Studio, explore your predictive AI use cases – propensity to convert/win/churn models and also build your first agent with our Agentforce!

I want to thank Chandrika Shankarnarayan, Anurag Juneja, Bobby Brill, Pradeepa Sankara Subbiah and Maanasa Gottipati for going above and beyond to review and provide their inputs to this blog.

We will follow up with another blog that talks more about Unstructured Data and Agents, and their immense possibilities in Salesforce Data Cloud.

Links to other useful blogs that were referenced above:

• Read more about how unstructured data makes AI more intelligent here in this article.
• Learn how to ingest Unstructured Data into Data Cloud here.
• Build an AI model with clicks in Data Cloud here.
• Find out more about RAG here.
• For more information on executing actions based on predictions in Data cloud, refer to this article.
• For a comprehensive list of Predictive AI based use cases, refer to this help page.

Written by