What are the top challenges around introducing machine learning at your company?

There are two top challenges to introducing a machine learning feature at any organization (1) Lack of vision from product and technology leaders (2) Selection of business problem to solve. Leaders need to understand what problems ML can and can't solve. Based on this, they need to pick a problem ML can solve and deliver maximum values to the clients. Learning ML, gathering data and training an algorithm are minor tech challenges which any technologist can overcome.

The biggest hurdle by far is Adoption. If customers and staff are not sure of how to interact and use the algorithm created, then that can cause distrust in your analytics team along with the ML as being not valuable. Focus on leveling up your organization on AI and ML. Product managers are great at products, customer service teams are great at service. It's unfair to expect them to under AI and how to properly use it. Develop training programs that level up your team members who consume analytics and AI. Define what AI means at your organization and create a foundational understanding, so everyone is talking apples to apples. Dispel myths and unknowns about AI, and help them play through business case scenarios to get an interactive feel for how and when AI should be used, when insights should be used, or at best, when not to use these!

One challenge I’ve seen is ensuring that the underlying data being used for machine learning is not biased. There have been numerous examples in tech where this has been an issue. One example was Amazon’s recruiting engine that was supposed to help them figure out who to proceed to interviews. It turned out that so many of the resumes they trained the system with were male resumes that the system was unintentionally biased against women.? To combat this we need to ensure we are taking into account the potential biases in the data we use for machine learning.?

I've helped clients in implementing Machine Learning to predict machinery failure. The most difficult challenge is to obtain reliable results in such a way that maintenance work orders generated by the algorithm can be executed without the need for a "second opinion" from analysts. If an analyst's diagnosis is still required, Machine Learning isn't worth it.

Building an understanding of the role of the machines and the humans. Machines take us a long way to the end but humans are still vital to final outcomes. In most business applications, ML models are built to predict outcomes (make decisions) in large datasets. The challenging points come from determining the application of ML to a specific desired outcome (problem being solved), choosing the appropriate data to solve the problem and making the right design choice (training dataset broad enough contain the answer and narrow enough for relevant inference) and importantly constantly training the models to “see” things that they previously didn’t recognize so that new anomalies become adopted, and the efficacy of the model remains viable.?Finally, having the best humans who can work through how to engineer the desired outcome consistently.