Click to learn more about author Jon Reilly.
As a co-founder of an AI startup, I spend a lot of time talking to companies working to implement AI in their organizations. More often than not, they start with a vision, excitement, and momentum — but have a hard time getting out of their own way.
The issues are made clear in the New Vantage Partners 2021 AI Survey – while 99 percent of companies are investing in AI, only about a third of them report seeing transformational business results. What’s more interesting is that 92 percent attribute the “principle challenge to becoming data-driven” to people, business process, and culture. Only 12 percent of companies surveyed reported having widespread adoption of AI in production.
So what causes the big disconnect between the will to adopt artificial intelligence technology and its successful deployment in a business? Chasing perfection and biting off more than you can chew.
Chasing Performance Perfection: A Recipe for Failure
Let’s start with chasing perfection — a natural consequence of human behavior. If you are building a predictive model to guide critical business decisions, you want it to be as accurate as possible. That instinct for best performance is compounded with the state of academic Data Science. A great example of that is the Kaggle competition — where the best data scientists and machine learning engineers build models that eke out tiny performance wins, and in exchange, land football player-like salaries at Facebook and Google.
The problem with chasing model perfection is that it requires a lot of organization and infrastructure work to get the last mile of performance. You need super clean data — and in the real world, almost no data is clean. Every data hygiene task you apply to your training set requires a reflective systems task to make the data you run against the model clean. You end up with two scenarios — the best performing model that does not work in production because the information you feed it is not as clean as the training set or a considerable infrastructure barrier to adoption. Either of these scenarios will stop progress in its tracks.
Biting Off More Than You Can Chew
The other issue that derails AI initiatives in businesses is when they attempt to wrap models around big and complex processes. The typical thought process here is, “Let’s build AI models on all of our business data to predict our key performance indicators.” Over-scoping occurs most often when leadership is driving adoption from the top down. The problem is, KPIs are usually the roll-up of hundreds or thousands of sub-business-processes.
Take, for example, the sales and marketing funnel. Your first instance may be to predict revenue for next year. So you train a model on last year’s revenue and use it to forecast next year’s results. The results are about what you would expect with predicting the weather — you do well in the very short-term, but the longer out you forecast, the more likely you are to be wrong.
Why? Because the underlying patterns in your business that a big-picture model is tracking are continually changing. You are launching new products, changing your marketing campaigns, and optimizing your sales funnel. Every change you make has an impact on downstream revenue but cannot be captured by the model until it has had time to work through the system. By that time, you will have made a lot more changes to your business. It’s a harsh treadmill that keeps you running without progress.
Here Is How to Get to Production
The first important thing to realize is that this is a perfect example of the 80/20 rule. With 20 percent of the machine learning adoption effort, you can achieve 80 percent of the value capture. How do you practically apply that rule? The first step is to stop chasing the perfect model. It’s okay to train with the messy data you have, assuming that your future data will be just as bad. The best AutoML engines are built to be robust to messy or missing data, so their models actually work in the real world.
Your less than perfect model will almost certainly be more efficient than not using AI. Even better, you can start taking advantage of AI right away. No need to wait for massive overhauls of legacy systems and extensive data hygiene work. You can capture 80 percent of the transformational business results right away, with 20 percent of the effort. Once you see the value, it’s easy to run an ROI analysis on future data work and decide if it’s worth making incremental improvements to your models.
The second important approach to adopting AI within your organization is to start small. I call it bite-sized AI (or daily AI). Pick a sub-process of your larger growth engine and make that more efficient first. Another advantage to starting with a bite-sized application? The complexity for the model to learn is lower. You will have more accurate models with better performance when you focus them on individual process flows. Score your leads as you hand them off to the sales team so they can prioritize who to call first. Route open-ended text inquiries to the right place automatically. The more specific the problem, the easier it will be to improve performance with machine learning.
As small sub-processes get updated over time, it’s equally easy to adjust, retrain, and redeploy new models. The sum of efficiency gains will roll up to your KPIs, and the results will be noticeable. And by starting small and growing over time, you also gain confidence and learn how to drive your business results with AI efficiently. Better performance and easier adoption — I call that a win-win.