Almost every single company is being pushed to innovate, and the favorite flavor of innovation currently is Artificial Intelligence (AI). In this article, we will cover some common challenges a company faces in the journey to becoming an AI-Driven Enterprise.
- Hiring a team
- Finding the right project
- Subject Matter Expertise and Communication
- Transparency and Trust
- Project Management
Challenge #1: Hiring a team
Data scientists are difficult to hire. They are scarce and in hot demand, have many options at their feet, which translates to a bidding war with a high risk of churn. For most companies, they will spend about 1.5 years just hiring the team. The next step usually will take another 1.5 years finding out you hired the wrong people, firing them, and starting over. In addition, another difficulty in hiring data scientists is that most companies are not well equipped to interview skills necessary for their businesses.
Challenge #2: Finding the right project
Now let’s assume it is a few years later and you now have a good data science team. Most likely this would be a centralized team – a central team of excellence that serves the whole organization. Based on my experience, I had always claimed that on average 90% of data science projects never fulfill their promise of bringing real value to a company. I later saw statistics from Gartner saying the rate is around 80%, VentureBeat claiming 87%. This is independent of the quality of data scientists you might have. The cause of this high rate of failure is not technical in nature. About half of those projects fail because of bad problem formulation. Let’s take a look at an example.
Imagine you sell t-shirts online. One day, Tom, a logistics manager, notices that every year the company spends $15 million in returns. He thinks that maybe AI can help and goes talk to Sarah, a data scientist. Tom will probably collect some historical data on which orders were returned and ask Sarah to build them a model to predict which orders will be returned. Sarah goes hard at work and after a couple of months, comes back with a very accurate model that she is proud of. She presents the results to the whole company, and now you ask how you can use this to impact your business.
You see it in use. When someone makes an order on your website and the order is shipped, you now can perfectly predict if that order will be returned or not. Now what? That’s right. Now nothing. There is not much you can do and this model will not be much help to your business. Would you consider to have become an AI-Driven Enterprise? Probably not.
This sounds exaggerated and nonsensical, but this happens more than often. The details and use case has been changed but I have seen this same problem a dozen times in companies of all sizes, including Fortune 50 Companies. This stems from the lack of communication between the data science team and the business team.
Challenge #3: Subject Matter Expertise and Communication
If Tom had the tool to do the AI projects himself, he would have the time to test different scenarios and pivot appropriately as needed to find a model that fits the problem. In fact, he would learn valuable insight that most returns are due to customers ordering wrong shirt sizes. As a matter of fact, Sarah couldn’t possibly have found the same fact just from a dataset. Only If Sarah and Tom were communicating more closely, they would have revealed this problem faster, leading to building a model to predict shirt sizes. As a result, when a customer puts a t-shirt in their cart, your model will run and recommend a more fitting size for that particular shirt.
Challenge #4: Transparency and Trust
Another reason why projects fail is because of a lack of trust and/or understanding. Even if you build a state-of-the-art model with significantly accurate results, if the users don’t understand how the model is making such predictions, they will not use the model. This happens extremely often in regulated industries, such as medical and healthcare.
Moreover, high accuracy does not really help a business user. For example, you may have a model that predicts your weekly sales and the model says the sales will be down by 15% next week. Great, now you know what will happen, but would this be an answer that you actually want to hear? You would want more, which brings us to the transparency, or explainability. If the model could explain itself and the users understand it, then it would help the users to dive deeper into the WHY. Instead of seeing your sales go down by 15%, you could reverse the prediction and see the sales go up by 15%!
At the end of the day, a model has a higher chance of being used when it’s built by a team that is going to use it because they are responsible. In other words, the team will build a model that is trustworthy and understandable so that it impacts the business without hurdles. Once you’ve impacted the business, you can claim victory as an AI-Driven Enterprise.
Challenge #5: Project Management
Finally, most data science teams are backlogged with projects. Unfortunately, the data science team has to prioritize and pick projects that would impact the business soon without much hassle. As a result, the turnaround time is high with no guarantee that your project will be completed within the timeframe.
In fact, there is a tool that is easy for business analysts to use. Because the project is led by the person that understands the business and is the final consumer of the model, the project will be developed accordingly.
As we covered different challenges to becoming an AI-Driven Enterprise, there are more than enough reasons to believe that an AI tool for business analysts is the future of making Artificial Intelligence ubiquitous. The tool could be created internally or externally and could be a SaaS or PaaS depending on your business. If you want to learn more about different options, read the article I wrote with VentureBeat here.