This Q&A on the future of AI is based on a data chat podcast episode: How to Avoid an AI Project Failure with Harish Krishnamurthy
Harish Krishnamurthy is president at Sciata and has held leadership roles across P&L, sales, marketing and strategy during his tenure at IBM, Insight Enterprises and Spear Education.
Where do early mistakes happen when companies decide they want to be data-driven?
Most mistakes fall into three buckets. The first is isolating data rather than seeing it as an element of a broader strategy toward dealing with the problem. The second is not anticipating the change management required to have a successful implementation. The third is not anticipating cultural aspects of the organization in terms of getting buy-in and getting people to understand the purpose.
How should companies begin the process when launching data projects?
The first step is to always take a holistic view of what needs to be done.
I’ll share an example of a client who was a well-respected CIO for a financial services organization He came to us and said, “we’ve got to leverage AI.” He then told us the most important thing they needed was a data lake, and from there they’d figure out the use cases. This approach was taking a deep dive into one particular aspect of a data project without looking at the overall solution.
Ultimately, we ended up having a broader conversation to look at the use cases and how they lined up with the strategic objectives. We answered questions like, “Where do we want to focus and what do we want to prioritize?”
And once we got an alignment on the use cases that made the biggest impact, we backed up and examined infrastructure options.
Walk us through the process of starting a data project.
Let’s use the example, “I want to improve customer satisfaction.”
But does that really mean? How can we translate this to something specific? What do we want to get done? What is the financial number that needs to change? What is the operational objective? Do we need to acquire more customers? Does it mean improving NPS scores from 85 to 87? It’s about getting down to those specifics. Even more specific, what are the objectives within those goals?
Then, you define the use case on how we can improve. There could be two, three different use cases. You’ll choose the one you think will have the most significant business impact.
Next, you find all the available data and measure its quality. The goal is to understand is the data available? Is the data relevant? How predictable is the data toward the outcomes that you’re looking to tell? To do this, before you invest in massive models, you’ll want to do a quick, small proof of concept.
Who Should be on a data team?
I would take a step back to look at what are the things that need to change at different levels. I would think about who needs to make the decisions. Who is influencing the execution, and who is actually implementing that?
If I use the same customer success or customer service example, it’s going to be important to include people from customer service. Not to mention, they are the people who would need to be making a change in behavior in order to get value or benefit from the data insights.
There are different levels of inclusion. At one end, it’s about getting their buy-in, having them be prepared to make a change and having them understand the rationale.
On the other end of this spectrum, it could be the data analytics team who need to be involved every step of the way, so they hear the feedback, they can secure the data, and they can build a model that’ll integrate that with the rest of the business.
What works when it comes to getting buy-in from the organization for a data project.
It starts with identifying the reason for the change. What is the issue? What’s the problem? Why are we making a change? And why do we need to do it now? Having them understand the status quo is not working, or there’s a reason for making a change is going to be critical.
Once the team is convinced that there is a need to make a change, then it’s about outlining the different approaches and pointing out what makes the most sense. This will help you gain credibility.
Once everyone agrees something needs to change. Then having case studies, having examples of where this has been done, having credible data points is going to go a long way toward getting them to get comfortable with making this change.
Then, showing them what they would need to do differently going forward. It’s important to emphasize that it’s not about completely throwing out the old and bringing in the new.
There is a study indicating that $50 billion is being allocated by corporations on AI systems, but many of these projects fail in some way. Why do you believe AI is not immediately solving the problems that organizations are hoping they’d solve?
There are typically five reasons why projects fail. It starts with not having a clear strategy and understanding of what is achievable. Companies can dive too quickly into the details without making sure there is a holistic approach.
The second is around design. There are governance issues, making sure the model doesn’t have algorithm bias and it’s designed to actually produce the outcomes that we’re looking for.
The third is the implementation. Failure will happen if a data project doesn’t have internal support for full implementation.
The fourth is around trust. It’s about believing in the insights that the model is providing and having that credibility with the team so that they actually go and implement it.
The last is actually changing behavior. One of the main reasons I’ve seen projects fail is the inability to track insights into action. It’s great that you get insights, but if there’s no action, it’s going to be very difficult to get any value from that.
How does an AI project differ from other data projects?
As the models get more sophisticated, the key thing that I would look for is there a mathematical or analytical approach to analyze all of the data to identify insights so that someone can take an action. I think that at a high level defines that there’s an AI component.
You can get much more sophisticated and look at deep learning convolutional network models, neural networks and machine learning models. These are all different variations of more complex analysis for identifying patterns and capturing insights from those patterns so that you can take an action.
Who should decide if AI is the right strategy?
I think it’s going to be a combination of folks looking at this from an analytical perspective.
It could be executives who have some level of experience in data and analysis. Specifically, they should be able to identify if the project cannot be done with simple statistical or analytical approaches.
It could be folks in the IT department. It could be someone in the finance department or an executive marketer who determines that this is a complex project that cannot be handled by normal analytical techniques.
Overall, it’s about having a good understanding of the challenge that needs to be addressed.
When creating a proof of concept, what are you looking for when it comes to scaling and moving forward?
There is a discovery effort at the beginning where you identify the data flow, how value will get created and where decisions will need to be made.
You want to start with a static model. Something standard that is built in excel or SPSS-type models. Most importantly, you want to be able to prove that the data predicts outcomes. It might not cover all the analyses you are looking for, but you want to be able to test several models.
Even before you build the final model, you might start to gain initial insights that can inform where you want to make investments to get more data.
Every situation is different to know when to move from a proof of concept to building the model, but typically it happens over a couple of months. It heavily depends on complexity.
Additional Resources:
Harish has written a series of white papers to help data professionals begin leveraging AI effectively in their businesses:
- Making the Leap from AI Investments to Business Results
- Aligning IT and Business Strategy for Project Success
- Using AI to Maximize Customer Lifetime Value
- Transforming AI Insights into Actions
- Designing AI Models to Extract Insights
Continued Learning
Business-Driven Data Analysis
Do you often find yourself asking, “How do I make data analysis more strategic for the business?” If so, we have the perfect course for you.
Business-Driven Data Analysis teaches you to use data to solve business problems. You’ll learn how to tell a cohesive story from your data and get your stakeholders excited about the insights and strategies it can deliver. This course includes 5 modules of training and a downloadable workbook of exercises. >> Learn More
Data Science for Business Leaders
Data is used to make predictions, understand human behavior and make decisions. The world of data is moving fast; it’s easy to get left behind. Data Science for Business Leaders shows you how to partner with data professionals to uncover business value, make informed decisions and solve problems. >> learn more
