How to start an AI Project – Baby Steps or Giant Leap
Speakers: Srinivas Atreya, Chief Data Scientist, Cigniti
Here is the Transcript
Srini: Hey, this is Srini and great to have you all here.
It’s that time of the year again and everyone is always talking about all the great things they’re going to do
starting in 2020. I will not lie. I’ve been thinking too, and hence today I want to do a small podcast on how to start an AI project for everyone that’s trying to do that in the new year. It’s amazing how quickly the conversations that I’m having with my data science colleagues and clients about artificial intelligence and machine learning have shifted in the recent months. I’ve gone from hearing many say that taking a wait-and-see approach to many suddenly wanting to get a jumpstart on the promise of these technologies. There’s a lot of excitement about how much value can be garnered from an AI ML implementation increasingly CIOs, and IT leaders are chomping at the bit to implement something to show what they can do. But implementing a companywide AI strategy is challenging, especially for legacy enterprises. The problem is many of them don’t know where to start. Or worse yet, they want to start with an Audacious project to show off what’s possible.
When asked my opinion, I always say start with your data. But once your data is in order, at least in the past, that you care about, what should you do? How can you get started with AI or ML? Should you jump in and take a large project? Or should you find some smaller opportunities to prove your data, processes, and people? Should you build internal capability or calling the big guns externally? Do you need big data for AI projects? Does this need to be a part of the IT.org or the Business.org or something else? We will try and explore some of these questions over the next few minutes.
My advice for executives in any industry is to start small. The first step to building an AI strategy is to choose one to two company-level pilot AI projects. These projects will help your company gain momentum and gain firsthand knowledge of what it takes to build an AI product. Technology is an issue, of course, but the bigger challenge is always change management. AI is different from other software projects in the sense that it will directly attack. Maybe attack is the wrong word. Directly challenge the decision-making culture of your organization. People will generally not trust the outcomes predicted by the AI platform and will erect all types of Chinese walls to discredit them. The best way to achieve this will be to start small and work with the teams that are relatively more positively inclined towards new-age tech like AI. I know this is easier set than down, but more AI projects failed due to this one reason than every other reason. It’s extremely critical to pick up the right team for your AI pilot. When you’re considering a pilot AI project, ask yourself the following questions does the project give you a quick win? Use your first AI pilot project to get the flywheel turning as soon as possible. Choose initial projects that can be done quickly or at least relatively quickly, six months or less, and have a decently high chance of success. Instead of doing only one pilot project, maybe choose two to three, not more than three, typically to increase the odds of creating at least one significant success. Is the project too trivial or too unveiling in size? Your pilot project does not have to be the most valuable AI application as long as it delivers a quick win. But it should be meaningful enough so that a success convinces other company leaders to invest in further AI work. It will help immensely if these projects attack problems that you already know well. This may sound like common sense, but trust me, it’s actually not that common. Is your project specific to your industry? By choosing a company-specific project, your internal stakeholders can directly understand the value.
For example, if you run a medical devices company, building an AI recruiting project to automatically screen resumes is a bad idea for the following two reasons. Reason one, there is a high chance that someone else will build an AI recruiting platform that serves a much larger user base and will outperform what you could do in-house and or undercut you. This project is likely to convince the rest of your company that AI is worth investing in. Then, if your pilot project applied AI to medical devices, it is more valuable to build a health care-specific AI system in this case anything ranging from using AI to assist doctors with crafting treatment plans, to streamlining the hospital check-in process through automation, to offering personalized health advice. So once again, it’s kind of important to take a project that makes sense for your industry and even more importantly to your organization.
Even when I was working with a large bid vested retailer, one of the first AI use cases we picked up was in the area of customer uplift modeling. This is one of the core problems for a retailer in the area of campaign management and marketing. One which is small and contained enough, but important enough too for the organization at large to care about the success of the same. Right? So that’s kind of where we’re getting with this.
The another very important aspect is really are you accelerating your pilot project with credible partners? If you’re still building up your AI team, consider working with external partners to bring in AI expertise quickly. Time is of essence here. Eventually you will want to have your own in-house AI team. However, waiting to build a team before executing it might be too slow relative to the pace at which AI is currently moving across industries. Actually, AI is largely IT initiative, but it is very different from other software projects. For one, the success is never really guaranteed. We can never be sure if a particular decision-making boundary can be learned or not.
Secondly, to be successful, AI projects need to build trust with humans. And this requires constant working with the end users and creating an environment where they can appreciate and understand the underlying aspects of the AI being built. And this point is extremely critical. Unfortunately, today AI is marketed as a black box, as some sort of panacea, some sort of magic. And this, by its very nature of the way it’s protected, creates a lack of trust, creates mistrust, because people don’t generally like something they can’t understand. So it’s extremely important that the target user groups are brought into conflicts. They kind of should be so-called bought into the inner circle so they start understanding how this thing is built and that it’s not just magic. And it’s actually a very very realistic and practical way to approach and solve a problem. In my experience, I’ve seen that helps a lot.
Thirdly, the marginal cost of an AI project does not approach zero like traditional IT projects. For example, if you build an accounting software, once it’s deployed in production, it keeps performing at the same level forever. We’re just a very lean team to support any potential bug fixes, et cetera. Now, in the case of an AI project, its performance will invariably degrade once put into production. And it will need constant handholding from experienced data scientists. This makes AI projects very different to traditional software project. Considering all the above, my suggestion would be to start AI projects as part of a separate small group. Neither IT nor business but having strong ties to both. over a period of time what I have seen is, depending on the culture and size of the organization, this group will drift to either side. But it’s important that at the beginning we do not try and stiff all that freedom, right? Let it be free and it will naturally move one way or the other.
Now, AI can also be taught also about AI, kind of is like automation on steroids. A rich source of ideas for AI projects will lie in automating tasks that humans are doing today using a technology called supervising them. You will find that AI is good at automating tasks rather than jobs. Try to identify the specific tasks that people are doing and examine if any of them can be automated. For example, the tasks involved in a radiologist job may include reading X rays, operating imaging machinery, consulting with colleagues and surgical clients. Rather than trying to automate their entire job. Consider just one of the tasks could be automated or made faster through partial automation. Additionally, starting small allows you to better understand the risks involved, of course, of which there are too many to enumerate here.
For example, if you start with that big project and realize that most of your data is grossly biased, you’re going to have a hard time building accurate models with a small project, you will still find that bias in your data, but the investment and resources to find it and correct it, if possible, will be much smaller and easier to manage. All in all, AI can give your enterprises tremendous value. It’s, been differently called as a new oil, the new electricity, the new data. But according to me, I think it’s much beyond all of this. AI will fundamentally change the way we live, and we work. And so, no organization, according to me, can actually really not start their journey, not start their AI journey. It’s absolutely imperative. But it’s equally important to kind of start the right way so that we can all reap the benefits without necessarily having to go through the pain. And my suggestion for that would be start small, start more than one of them, and start it in an area or in a place in your business where people are a little more receptive to these types of ideas. And also try and make the first couple of projects very very relevant to your particular industry and company. It’s not important that you solve the most glamorous or the most valuable one, just that the one that will probably have the highest probability of giving you a quick win and impacting the entire organization positively.
Hope this helped. So, once again, thanks for listening and bye. for the next time. See you, maybe. Hope to have you for my next podcast. Thank you very much. Bye bye.