We use cookies to make interactions with our websites and services easy and meaningful, to better understand how they are used and to tailor advertising. You can read more and make your cookie choices here. By continuing to use this site you are giving us your consent to do this.


I often meet with businesspeople who are excited to get started using AI, but they’re confused about what it can and cannot do. AI is being built into business software with incredible speed. Some examples of AI-powered applications available today are lead scoring, sales forecasting, marketing, service, and many more.

For all the ways in which this intelligence is exposed, the AI that powers business applications really boils down to a few discrete functions which, when combined in innovative ways, can deliver real value. It is, in a way, like Taco Bell.

Taco Bell doesn’t have that many ingredients: It’s mostly beef, refried beans, cheese, tortillas, salsa, lettuce, and tomatoes. However, they can combine these ingredients to make many different menu items according to their customers’ tastes. Everything from a soft taco to a Mexican pizza to a quesarito are just different combinations of the same ingredients, and yet they each serve different segments of Taco Bell’s customers.

Now, as AI-powered business software matures, it’s allowing businesses to make their own custom predictions — and that means it’s more important than ever to understand the “ingredients” that AI provides.

By and large, a good AI platform comprises four different ingredients. We’ll take a look at them separately (many times they’re useful by themselves — sometimes it’s nice just to eat cheese), and then we’ll have a look at some examples of how these ingredients can be combined to make for a robust menu of AI applications.

Yes and No Predictions

The first “ingredient” of AI is a yes and no prediction. This allows you to answer questions like, “Is this a good lead for my business?” or “Will this prospect open my email?” AI learns how to answer these questions from the historical data you’ve stored in the system, and the prediction generally comes in the form of a probability (for example, “Mary Smith has a 67% chance of opening an email from you”).

Sometimes these probabilities are converted into scores. Scores are just a different representation of the likelihood of “yes”; they can be represented as numbers from 0 to 100 or 1 to 10, or they could be shown as 1 to 5 stars, Yelp-style. Fundamentally, though, scores are just showing the same prediction in a different way.

Numeric Predictions

The next ingredient is a numeric prediction. Numeric predictions often power predictive forecasting solutions (for example, “How much new revenue will this new customer bring in?”), but they are also used in other contexts like customer service (for example, “How many days will it take us to resolve this customer’s issue?”). Here, too, the AI is learning from your historical data to arrive at these numbers.


Next we have the notion of classifications, which use deep learning capabilities behind the scenes and generally operate on unstructured data like free text or images. The idea of classification is to extract useful information from this unstructured data and answer questions like, “How many soda cans are in this picture?” Or it can take a person’s utterance, like, “I’d like to buy another pair of the same shoes I bought last time,” and use that to kick off the workflow required to look up the last shoe order and place another of the same.

Classification using deep learning is very robust even when the unstructured data is arriving in different forms. To continue my above example, there are lots of ways I might be able to tell a vendor that I want another pair of shoes. I could just say “I want another pair of those shoes” or “Give me another one of those” — there are a theoretically infinite number of ways I could phrase that request, but the underlying deep learning engine built into an AI engine generally can understand them all, in much the same way that your brain can.

A third type of classification — which may or may not use deep learning — is called “clustering.” This type of AI ingredient learns insights from your data that you may not otherwise have noticed. For example, if you are a clothing vendor, it might learn that both rural older men and urban hipsters like to buy a certain type of sweater. Where your intuition might tell you that these are two totally different groups, the data shows that in fact they behave similarly with respect to the products they buy, and so you may want to market to those two disparate groups in a similar way.


Recommendations are key when you have a large set of items that you’d like to recommend to users. Many ecommerce websites apply recommendation techniques to products; they can automatically detect that people who bought a certain pair of shoes also often buy a certain pair of socks. Then, when a subsequent user puts those shoes in her cart, it can automatically recommend the relevant socks.

Recommendations are not just for products, though. In marketing, one can use the same technique to recommend content like whitepapers to business users. One can imagine many other use cases, too. For example, customers might use this technique with their HR recruiting system to recommend job postings to job candidates and to find the best job candidates for a given job posting.

Workflow and Rules

“Wait,” you’re probably thinking, “I thought you said there are only four ingredients to AI.” Workflow and rules are not part of AI per se — they often live in the business software itself and not in the AI system — but they form the key to how the outputs of AI are consumed. Just seeing that a given customer has a 25% likelihood to attrit, for example, is not enough — you need to do something about it, and workflow and rules give you that capability. In that example, your workflow might automatically detect that a customer is now predicted to attrit, and would kick off a retention campaign to try to retain that customer. Workflow and rules are like the tortillas and taco shells: They allow you to consume the ingredients without making a mess of it.

What’s on the Menu?

With these four fundamental AI ingredients and workflow, we can produce AI applications that solve real-world problems. Increasingly, as business AI systems mature, businesses are becoming able to use these ingredients themselves to make their own custom AI solutions.

For example, consider the case of predictive forecasting. I mentioned earlier that numeric prediction is a big part of predictive forecasting, but that’s not all of it. To accurately forecast a deal, you do indeed want to predict the amount of revenue it’ll bring in, but you also want to predict whether it will actually close or not, which is a yes and no question. These two predictions put together will give you the actual predicted forecast for a given sales opportunity.

Many good business AI systems already offer a rich set of capabilities including many of the ingredients illustrated here, and there exist numerous prebuilt applications that cover some of the use cases I used as examples above. With the recent ability to create customized AI applications, businesses will surely make thousands more applications that are customized to their own businesses. The more you understand the ingredients available to you with business AI, the better you can whip up a tasty dish for your business.

The more you understand the ingredients available to you with business AI, the better you can whip up a tasty dish for your business.”

Marco Casalaina | Vice President, Product Management, Salesforce



Created by Salesforce