How to implement an AI project – from business problem to action

Artificial intelligence (AI) and its use in the business has given rise to many debates lately. Thanks to cloud computing platforms, handling large amounts of data is now much more cost-effective. AI is promised to automate processes, imitate human intelligence, predict the business and customer behavior and perceive the complex business environment around us.  

The retail sector can benefit from AI in many ways. It's no longer a question of whether to use it or not, but how and when to use it.  

AI project typically involves a few common steps, no matter what use case you might have. It always starts with 1) business understanding and then proceeds to 2) data understanding.  

After turning business understanding into data language, it is necessary to 3) find the correct data and prepare it for modelling.  

The next step is to 4) model and apply the algorithms – that is, to teach the model.  

The final step is to 5) take the model into use in daily business  

and then 6) see how it works and what kinds of results it gives.  

The first three steps: business understanding, data understanding, and data preparation take most of the time.  

When using AI, business understanding is essential. You must first know what you want to solve. That guides you through the entire cycle. If you cannot find a problem you want to solve, it's not even worth starting the process. 

Price optimization case  

When you want to build a price optimization model for some of your products, your objective is to increase profit. You want to see how your customers react to different price strategies and price changes and then set the best prices for the products.  

The aim is to:  

• predict demand for certain products 
• get an idea of ​​the optimal price at a certain demand level  
• estimate how the profit evolves. 

Business understanding, data understanding and even data preparation all require lots of effort from business experts. Projects are far too often technically led, and business expertise is only asked when needed. One team where everyone works together – others with the business problem and others with the technical issue – would benefit the whole project.  

Once you have set up a team around you, it's time to collect the correct data. The problem here is not to find the data but to find the relevant data. In addition to internal data sources, it's possible to acquire data outside the company.  

In the case of price optimization, you start with sales data. However, many different factors affect the demand, so you need other data types. Item information (such as categories, sub-categories, and descriptions), holidays, events, calendar days and so on affect the demand, as do competitors’ prices and macroeconomic information, such as unemployment rates.  

Only relevant data matters 

It is essential to validate merely the relevant data parameters for each case. The importance of cooperation between data engineers and business experts is emphasized here, as business experts are those who can say which data is relevant.  

Note that you don't have to have all the data currently stored in your company. You can also acquire some of the data outside.  

Once you have identified which data answers your question, you need to store it on a data platform. A data platform integrates different data sources and creates a modelling data set that is easy to use for data scientists.  

Before proceeding, you'll probably have to do some pre-analysis, restrictions, and decisions regarding the data. You might, for example, have products that sell very differently. Some products might have smooth and steady sales, while others have intermittent or irregular sales. In that case, it would be a good idea to start with those who have stable and smooth sales over history. 

For an initial data analysis, ask yourself: 

• Which are the items to be price optimized? 

• How often to do it? How have the prices changed in history? 
• How has the demand evolved in history? 

Testing the model in real life 

Once you have the data set ready and you have applied the required decisions and restrictions, you are ready to apply the model. The idea is that you teach the model and train the algorithms. You can apply ready-made algorithms with your data. You can, for example, use different deep learning models and even reinforcement learning techniques depending on the volume of the data.  

The purpose of the first results is to validate the model. You want to see whether you have the necessary parameters and factors that affect the demand. The aim is to build a demand curve – to see how different factors affect the demand. How much different prices for a particular product in a particular time window, for example, change the demand and the expected profit. Through modeling, you should also get the top categories and factors that affect the demand.  

Storytelling helps you to understand the results better.  

Once all this is clear, you are ready to proceed and test the model in real life. You can set the prices for your products and see how the profit evolves. You can do that regularly, either dynamically or manually.  

By doing this, a bunch of new questions will undoubtedly come up. How, for example, does selling products with these new prices affect the demand for related products?  

With these new questions, a new cycle starts all over again.  

Want to learn more? Can we help you with your AI project?