The AI revolution, as a continuation of the BI revolution, caused companies to acquire, ramp up, centralize, and deepen their data analytics, data science, and data handling capabilities. As part of this revolution, companies are building some of their core capabilities around machine learning models as an automated mechanism to make business decisions.
In finance, for example, the economic boom together with low interest rates has resulted in a large surplus of lending and investment funds. This changed the competition from “who will get funded” to “who will fund.” Lenders now have to snatch up market share quickly and efficiently, while also making smart decisions when it comes to potential risk/fraud, interest rates, and expected ROI. In the travel industry, online travel aggregators and agents use machine learning models to personalize and match the best pricing, hotels, and flights to customers in order to ensure maximum sales while keeping competitive pricing.
Building a company’s core operation around ML models has helped many companies thrive across many other verticals — insurance, advertising, and eCommerce, to name a few. Companies that harness data and use it to make automated decisions like Uber, Airbnb, and Agoda have grown massive on this premise. These days almost every new technology-driven startup is built around ML engines.
But! And there is a big but! Data scientists are scarce and building the right set of models to sustain an entire business is a hard, time-consuming effort that requires almost all of the data science resources a company can get.
As these companies grow and their core business models and processes stabilize, training models and managing their life cycle in production become a streamlined process. So, naturally, they start looking to optimize additional areas of their business. Combine this with the clear ROI predictive models provide and the appetite for building more models across a business’ value chain skyrockets.
Still, most of these organizations wait to scale. Why? When it comes to ML I say start big and grow.
Why don’t companies use ML across their business?
Companies don’t scale their ML capabilities for a whole host of reasons. Generally, they can be broken down into three main categories.
- The perceived resource gap: You think your data science resources should remain focused on the core business rather than support additional functions. However, today’s AutoML tools, libraries, and capabilities let you access new data and build new models much faster than before.
- The structural gap: In many companies, data science is either spread across the company or located in a specific function. What was once a risk team, can now be seen as a potential centralized data science team. Once you start looking at this team in a new way, you open your company up to an endless world of models.
- The perfectionist conundrum: Any supporting models you create don’t have to have metrics as high and sound as your core models. There is room for compromise. As long as a model’s output test is consistently better than your current random/rule-based solution, it can still be a quick win.
Where can I deploy ML models?
Top of the funnel – the usual suspect
Leads are expensive, especially in a competitive environment. Whether they are acquired organically or paid, there are many ways to optimize your spend. For example:
- Outbound activities: Time and effort for outbound activities can be prioritized and optimized right at the top of the funnel. Some examples of outbound prioritization include:
- Direct mail marketing: prioritizing the right companies to spend money on to send physical mail.
- Messaging optimization: Adding a personalized touch to your emails and targeted efforts
- Campaign marketing: Getting the right leads into your funnel is tricky, companies are now spending more and more time, money, and effort in employing ML models for lead scoring, so their online and social media advertising spend results in higher conversion rates.
Specifically in the financial realm, top of the funnel models are useful as they offer a more focused and less regulated method for avoiding wasting marketing efforts on risky endeavors.
Moving down the funnel – use every drop of data
Every step in the sales cycle requires time and effort — whether that means calling the customer, meetings, writing content, or any other manual work.
As you move down the funnel you lose leads but the leads that survive gather more and more data and can provide a richness to your model. So, why not use this data?
ML models can still be used to prioritize and focus the sales team on the right opportunities, especially in the (always welcome) case where open opportunities surpass the resources to handle them. Building a model that will assign an actual probability to close (at all times) will help your sales team better prioritize their engagements.
The other side of the coin – churn predictions, success, and lifetime value
Building a customer success team is tricky. On the one hand, you want to make sure you have a sustainable ratio of team members to clients, and on the other hand, you want to make sure you are preventing churn and maximizing the value of each individual client.
Metric led CS teams are non-negotiable. There are many solutions today that help you measure your client interactions, engagements, clicks, and calls. These solutions help you gather product feedback and perhaps create a few rules of thumb regarding clients slipping away. In the era of BI, these rules were good but building a churn prediction model that utilizes both all the data you have gathered on the client and additional enrichments will help your CS team prioritize their efforts and identify “flight risks” ahead of time. Why wait for a client to start the churning process when you can take preventative measures?
Building and deploying ML models has never been easier. Where there is data, there is a model to be built. Service desks, HR capacity forecasting and even additional small, yet significant steps in the funnel.
Machine learning platforms, new libraries, and automated data discovery and feature engineering tools let data scientists build models, enrich them, and deploy them in a matter of hours or days. Utilizing the right set of capabilities and platforms maximizes the efficiency of your data science resources. You no longer need weeks of pre-processing, feature engineering, data selection, and model testing. It’s a simple solution for simple problems.
Your core models will still take up the most time. A good risk model is not a nice to have for an online lending company, it’s a necessity. Any AUC lift can mean a huge impact on the core business. But if you haven’t started focusing on the many more machine learning use cases across your pipeline, you are just leaving money on the table, or worse — throwing money away.