The Cold Start Problem with Artificial Intelligence

Why companies struggle with implementing AI and How to overcome it.

If you have become a Data Scientist in the last three or four years, and you haven’t experienced the 1990’s or the 2000’s or even a large part of the 2010’s in the workforce, it is sometimes hard to imagine, how much things have changed. Nowadays we use GPU-Powered Databases, to query billions of rows, whereas we used to be lucky if we were able to generate daily aggregated reports.

But as we have become accustomed to having data and business intelligence/analytics, a new problem is stopping eager Data Scientists from putting the algorithms they were using on Toy Problems, and applying them on actual real-life business problems. Other wise known as the Cold Start Problem with Artificial Intelligence. In this post, I discuss why companies struggle with implementing AI and how they can overcome it.

Start with Data

Any company either startup or enterprise, who wants to take advantage of AI, needs to ensure that they have actual useful data to start with. Where some companies might suffice with simple log data that is generated by their application or website, a company that wants to be able to use AI to enhance their business/products/services, should ensure that the data that they are collecting is the right type of data. Dependent on the industry and business you are in, the right type of data can be log data, transactional data, either numerical or categorial, it is up to the person working with the data to decide what that needs to be.

Besides collecting the right data, another big step is ensuring that the data that you work with is correct. Meaning that the data is an actual representative of what happend. If I want a count of all the Payment Transactions, I need to know what is the definition of a Transaction, is it an Initiated Transaction or a Processed Transaction? Once I have answered that question and ensured that the organization agrees on it, can I use it to work with.

With the wide adoption of SCRUM and frequent releases, companies have to devote resources to ensure that the data is correct. Companies could, add new sources of data, changes in the code that can have an impact on the logged data or even outside influences like GDPR or PSD2, that can cause the data to be altered because it needs to be more secured or stored in a different way. By ensuring that during each process the correctness of the data is ensured, only then can you move on to the next phase of analytics.

Turning Data into Analytics

Even though AI is currently what everybody talks about, before we get there we still have to take an intermediate step, which is Analytics. What I mean by Analytics, is the systematic computational analysis of data or statistics. In most companies, the process to get to the visualization of the data might be known to few, but the impact it has on each department is tremendous.

Analytics is the systematic computational analysis of data or statistics.

Companies need to determine which Key Performance Indicator’s (KPI’s) actually drive the business. Working in the Payments Industry, my KPI’s include Processed Revenue, Transaction Costs, Profits, Authorization Rates, Chargeback Rates, Fraud and many others that provide me with the information to manage the performance of the business. For a Taxi App, KPI’s might include, Revenue, Profit, Average Pick-up Time, Average Ride Time, Active Users and Active Drivers.

From those KPI’s, a company can then decide what type of reporting or dashboards are necessary for the business users to make informed decisions and work on automating the systematic computational analysis of the data or statistics.

But as the volume of data increases from KB’s to TB’s, and business users are looking more and more at aggregated reports and visualizations of the data, the chances of detecting smaller issues reduces significantly. It is only then, that implementing AI can become a worthwhile investment of time and resources.

From Analytics to AI

Having determined the KPI’s that help steer the business, Artificial Intelligence can be used to improve these KPI’s. In this case AI is the rational agent that uses algorithms to achieve the best outcome for a specific scenario.

For example, as the Authorization Rate in the Payments Industry is very important for large merchants due to the millions of transactions they process everyday, being the PSP with the highest Authorization Rate can make or break a deal or in the case of big merchants getting traffic or not.

By tracking the Authorization Rate on a day to day basis, we are able to figure out what our performance is over time. As our analysis has determined that Routing transactions through different Acquiring Banks leads to two different KPI’s. We can use AI to learn in which scenario’s one Acquiring Bank performs better than the other, and automatically route transactions to the acquirer that performs best.

As our implemented AI model and logic positively impacts the Authorization Rate, the related KPI’s are affected as well. As a higher Authorization Rate leads to higher Processed Revenue which leads to higher Profits.

Reasons why companies struggle with AI

So why do companies struggle with AI? I believe that the rapid waves of technology has caused many companies to be a little clueless what they should focus on. The move from data centers to cloud, from web to mobile web to native apps, and from big data to AI, has not made it easy for enterprise companies. Especially the discussions between the CFO and the CTO, who on one side wants to reduce costs and on the other side wants to have the best possible technology at his/her disposal, can lead to indecision, which leads to no decision.

Another reason why I believe companies struggle with AI is that the lack of evidence that AI can have an impact, leads a lot of people to believe that it’s just a hype and will eventually go away. Many industries including Financial Services, Transportation and Insurance have used data and computerized decision making to have an impact on their business, but many other industries who don’t have the same level of data are harder to convince that AI can have an impact on their business.

How to get over the Cold Start Problem with AI

Just like the Digitization of business required a different mindset, the Algorithmization requires one as well. With technology being available to everybody, competition has increased and commoditized many industries. By adopting an AI mindset as a company, and acting accordingly, is the first but most crucial step in overcoming the Cold Start Problem with AI.

New Technology startups have it the easier, because there is no history and nobody in the company needs to be convinced that it is necessary. So from the moment a startup designs and writes their application, adopting the AI mindset, will have a huge impact on the decisions that will be made and how data will be used to build a better business/product/service.

Incumbents have a bigger challenge ahead of them. Because even though top management might be convinced that going the Google, Amazon and Microsoft route in becoming AI-first, is the way to go, most employees will not be easily convinced. And as with any project, ensuring that everybody is onboard is a crucial factor in succeeding.

So the real question than becomes, what do we do?

Learning from past transformations in business, we know that true change comes from within. By adopting the AI-mindset management can have a big impact on the rest of the organization. But actual AI projects that focus on using AI to change a process, is where you will be able to change people’s opinion. Of course, starting a one-off AI project that doesn’t impact the way you do business, is not going to cut it. It is best to start multiple AI projects that tackle different parts of the business, hedging your bets but also encouraging sharing best practices across the organization.

The Future of AI and Data Scientists

As Data Science and the profession of Data Scientists begins to settle within the scope of business, I see a lot of opportunities. With the access to Cloud Computing from companies like Google, Amazon and Microsoft getting cheaper, the capabilities and the tools on top of this infrastructure keep on increasing, and it becomes extremely easy to train, deploy and develop AI models, the true role of a Data Scientist will be to really understand what the problem is that you are trying to solve and then figure out what tools and models are necessary to solve it. This will lead to Data Scientists, that will have to become a Domain Experts instead of only focussing on the “Data Science” part of the solution they are proposing.

This does not mean that a Data Scientists should not know what algorithm’s are best suited for a particular problem, it only means that they need to fully understand the problem, how it relates to the customers, the objectives of the business and the unique composition of their data set, and than bring together the Data Science and their Domain Expertise to improve the KPI’s and the underlying business.

Thanks for reading 😉 , if you enjoyed it, hit the applause button below, it would mean a lot to me and it would help others to see the story. Let me know what you think by reaching out on Twitter or Linkedin. Or follow me to read my weekly posts on Data Science, Payments and Product Management.

Recent Posts