To ML or not to ML?

Learn how to determine the applicability of machine learning in addressing business automation problems

The barriers to adopting machine learning (ML) for solving business problems are getting lower and lower. Today, there are tons of low-cost cloud services with abstraction layers to facilitate data preprocessing and even identify and deploy the best algorithm without the help of a data scientist.

This reality creates great opportunities, but also poses some serious risks. On one hand, companies accustomed to traditional decision-making methods may lose competitiveness as they ignore ML approaches. On the other hand, companies seeing early success in their first ML projects may experience negative impacts as they start to apply it to every problem without first confirming it’s the best tool for the job.

When an organization is considering an automation solution, the suitability of ML applications will sometimes be apparent. For example, consider a company with a large e-commerce platform looking to optimize sales by offering relevant, user-specific product recommendations. In all likelihood, that company will significantly benefit from ML to process information on the fly, helping them provide personalized recommendations to the thousands (or millions) of users visiting their website.

Yet, for many business problems, there’s a risk of mistakes on both ends of the spectrum:

1. Using ML to solve business problems that require a different solution.
2. Failing to adopt ML when it’s the best solution to the business problem.

Let’s take a deeper dive into how to prevent these mistakes.

Issue #1: Using machine learning to solve business problems that require a different solution.

There are plenty of legitimate use cases for ML. However, as it becomes increasingly ubiquitous in organizations, a risk arises: treating every automation problem as an ML problem.

For instance, you may’ve seen this happen when a company hires a director of machine learning who starts proposing ML solutions for various automation initiatives without first confirming it’s the best tool for the job.

Many automation projects don’t require ML at all. For example, a conventional rules-based solution can easily automate a task to screen millions of incoming medical claims for apparent errors (e.g., the patient or diagnosis code doesn’t exist in the database) without ML.

Issue #2: Failing to adopt ML when it’s the best solution to the business problem.

This is also a common issue, as solution builders may be unable to recognize when ML is the most suitable answer to a problem. Some examples include>

• A small company acquired thousands of emails from interested parties at a trade show. With a small team, it’s struggling to prioritize which contacts to follow up with first to maximize returns.
• A midsize company is growing fast and needs to accelerate its hiring process. Despite doubling its spending on job postings and events to attract more candidates, it isn’t seeing results.
• An online marketplace is expanding and experiencing the negative impact of increased fraudulent activity, such as fake positive reviews used to boost the ranking of low-quality suppliers.

So how can ML contribute to the solution of the above problems?

1. By prioritizing leads.
   ML can segment previously won and lost deals via unsupervised learning and map the new contacts to specific clusters: best fit for our product, potential fit with costly customization, and poor fit. This classification will help prioritize follow-ups.
2. By finding the most effective recruiting initiatives.
   Feed a supervised learning model with a table where each row contains information about a job post, its placement (where the opening was advertised and for how long), and the outcome (how many qualified candidates applied). This model can be used to identify the best combinations of ad words and promotional activities to increase talent acquisition.
3. By detecting and deterring fraud in a marketplace.
   Instead of relying on fixed rules that fraudsters quickly adapt to, build an anomaly detection model that combines various signals, such as recency of the account posting and wording similarities to other reviews by the same account. The model can then be used to block or flag reviews that look fake, and you can periodically retrain the model to adapt to new fraud strategies.

How to avoid these common pitfalls

It’s up to organizations to identify the business problems they need to solve and assess whether the benefits of ML will outweigh the up-front investment, ongoing maintenance costs, and possible ethical and security risks.

The following steps can greatly minimize the risks of misusing ML technology:

1. Prioritize tech and data literacy. For businesses that want to thrive in the current landscape, this is a crucial skill that executives and employees at every level should develop. Topics like data-driven decision-making, Robotic Process Automation (RPA), and the art of the possible with AI/ML should be part of the professional development curriculum industry-wide, from marketing and sales to manufacturing and product management.
2. When dealing with an automation problem, ask yourself: “Can standard automation work well here? Or does our solution need an engine that learns from the data?”
3. If ML is indeed the right approach, don’t underestimate the need for human analysts. Human intelligence is still required to confidently establish that the data you feed your learning algorithms includes all essential aspects of the problem and excludes irrelevant or misleading data features that can lead to spurious results.

In summary, if you’re given a business challenge to work on, make sure you explore the problem and solution spaces before deciding for or against ML. Will it increase the precision, consistency, or agility of a business decision? Or reduce the time to decide on (as well as the cost of) the decision? Great. But if you’re adopting ML primarily because it’s such a hot topic in business today, take time to slow down and evaluate what set of tools and methodologies are going to work best for your business.

You may also like: Why AI/ML Projects Fail — and How to Fix Them

Slalom is a global consulting firm that helps people and organizations dream bigger, move faster, and build better tomorrows for all. Learn more and reach out today.