Why is it so Hard to Accelerate Time-to-Value in Machine Learning?

The benefits of Machine Learning (ML) for automation and efficiency have been obvious for some time now. It isn’t difficult to find those who have touted ML’s transformational power for business, science, and manufacturing. Reid Hoffman recently pointed out at a CNBC Technology Executive Council, “You are sacrificing the future if you opt-out of AI.” But a survey of executives found that 65% had not seen value from their investment in AI, because so few projects have been put into production. Increasingly, in businesses of all sizes, everyone from the CFO to product leads have been asking Data Science teams the same question: When will this AI/ML experimenting pay off? And what can be done to accelerate the time-to-value of Machine Learning efforts?

Increasing Machine Learning accessibility

Machine Learning, as it is thought of right now, is the domain of Data Scientists and MLOps experts. This is understandable with the previous requirements of building ML systems and reasonable if your company can afford to staff these types of teams. Even then, those teams need time and patience to create a Machine Learning infrastructure that can support an overall organization's needs.

Ok, so what about everyone else? Most companies don’t have the budgets or resources to build teams internally or to hire an army of external consultants to replicate these structures. The argument goes that perhaps they should upskill their teams in Data Science. That’s missing the point – empowering organizations with Machine Learning means empowering teams with Machine Learning. Product teams, Business Analysts, Data Teams, and Operations Teams. Because ultimately, every team can benefit from the efficiency and power of data-driven decisioning.

You shouldn’t need to learn Linear Algebra to help your team succeed with Machine Learning, in the same way you don’t need to be a Computer Scientist to use a computer. You just have to understand your use case and business goals to design an ML automation to help your team succeed.

Here is where better tools come in. Not just tools that help teams tweak models (important, but again requires expertise to use), but truly intuitive tools that any team could use. And the organization itself benefits - a single AI project will help some of the business, but hundreds of Machine Learning projects will create a distinct, ongoing competitive advantage for the business. Accessible ML will transform a company's operations, unlocking value and productivity across the organization.

Lowering the data requirements of getting started

Ask any Data Scientist what the first step of any successful Machine Learning project is and they’ll say, “Data.” Ask them the most critical step in successful Machine Learning, and they will also say, “Data.” It begins and ends with data and, not surprisingly, the initial data gathering project is where many ML projects start to go sideways. Why? Because inevitably, the data piping and warehousing project becomes a monolithic task in and of itself.

I recently spoke to a technology leader who shared that he spent over a year-and-a-half just gathering requirements for his company's data repository project. Admittedly, that data will have other valuable uses for the organization, but when it takes 12 - 18 months just to prep for collecting data (never mind the subsequent year plus of building the pipelines) it’s understandable that businesses are becoming increasingly impatient to see the outcomes of these efforts. Oh, and let’s not forget that some of these data projects often ignore collecting event-driven data, the very causality data needed to find the all-important patterns critical for the success of Machine Learning. 

"It is a mistake to deprive teams and organizations of the power of Machine Learning while waiting for a large data planning project to complete."

So what can be done here? Plainly put, there needs to be a way to collect data for an ML use case without re-architecting entire data collection processes. There needs to be a way for teams to look at business use cases and quickly collect the data they need for Machine Learning in that specific use case without making a requirement that the entire system is rethought. Some Data Architects will balk at this approach, arguing it will create pockets of data throughout the organization. However, these two concepts of centralized data management strategy and speed-to-value of Machine Learning need not be mutually exclusive. Meaning, you can get started on one while working on the other. It is a mistake to deprive teams and organizations of the power of Machine Learning while waiting for a large data planning project to complete. It’s a bit like that old adage of “sacrificing the good for the great.” 

Accelerating low-cost experimentation and use cases

A common mistake companies make when they build or hire Data Science teams, is thinking they need to “swing for the fences,” in terms of moonshot goals. Addressing high-visibility,  high-risk use cases business leaders see as most impactful is costly, time-consuming, and fraught with risk. Often, these core use cases are central to how the organization conducts business daily - such as Fraud Prevention in Banking - and as such, require a lot of planning, stakeholder participation, and long timelines to execute. And in the meantime, the rest of the organization has to wait until these projects are complete for their own use cases to be addressed. 

The savvier strategy is to increase the amount of low-cost, rapid experimentation across the entire organization. Rather than just putting all the organization's ML eggs in one basket, a rapid test-and-learn strategy allows for the discovery of other valuable, lower-risk workflows that can be enhanced with Machine Learning. One project will not increase the efficiency or profitability of a company by 30% but a lot of 2 - 5% improvements can drive massive change in aggregate. An example of this success was found in the famous case study of the British cycling team under Sir Dave Brailsford, with his big revelation was that the team was going to find 1% improvements in hundreds of places. This “aggregation of marginal gains” to one of the most successful runs in British cycling history. Quick, easy wins, lead to significant and sustainable, long-term success. 

Speeding up Machine Learning's time-to-value

With companies increasingly re-examining technology spend and efficiency, Machine Learning promises to accelerate efficiency and value. However, many companies have struggled to find reasonable time-to-value on their internal Machine Learning projects. By empowering all teams to access and deploy basic Machine Learning capabilities, lowering the data requirements of getting started, and accelerating low-cost experimentation, more businesses can unlock the value of Machine Learning’s promise of efficiency, productivity, and profitability - in a relatively short time.