Any IT leader who has tried to put advanced technology—such as AI/ML, data science, and IoT—into production for their company has most likely faced their share of challenges.
Just knowing where to start—let alone how to make it all run sustainably in a production environment—can be difficult. The fact that only 8% of companies have deployed an AI model into production (according to Gartner’s 2019 AI Maturity report) is testament to that.
But here’s the thing: Becoming part of that 8% is not impossible. What’s more, because there are so many advantages to AI/ML in a business setting, it’s worth the effort to make it stick. You just need to be willing to learn from your own—and others’—failed attempts.
How modern BI factors in
Domo spends a lot of time discussing and defining “modern BI”—and for good reason: It’s the next rung on the digital transformation ladder, which is to say it’s a data-driven approach that puts real-time data into the hands of business personnel, fostering innovation, better decision-making, and an ability to solve more complex problems, fast.
That’s an important rung to be on, mainly because we no longer have the luxury of waiting until the next review cycle, the next month, or even the next week to correct business strategies or initiatives that aren’t working.
And because the demand for faster decisions has increased, the complexity of the solutions that are out there has also increased—and placed additional onus on IT leaders to put the right kind of data at business consumers’ fingertips.
So, how do you overcome the challenges of deploying an AI/ML model into production and experiencing the true technological power of modern BI? By looking first at the three main reasons why deploying it is so hard.
1 – The process is different
While AI/ML for business is not a new concept, the ability to innovate and deploy a model in a business setting is relatively recent. And, unfortunately, one can’t just apply traditional dev ops practices to AI/ML problems.
One of the defining characteristics of the AI developmental process is its need to be iterative; because of the heightened need for data quality and data volume, and the interrelationships between the training data, the model, and the interface data, iteration is critical to ending up with a robust, scalable, and deployable model.
But that’s a tough mindset to adopt. Most IT organizations are measured by their level of churn. Churn equates to inefficiency, and iteration can often seem antithetical to reducing churn. So, at first blush, iterative approaches can appear to fly in the face of what productionizable AI/ML is all about.
2 – The steps in the process aren’t always linear
The process for modern dev ops is well defined. It’s a condensed series of steps that repeat throughout the agile development methodology and continually deploy to production.
In AI/ML development, the process involves five key pieces—knowing what you’re trying to achieve; having data (and AI/ML generally requires far more than traditional software); being willing and able to experiment with models; deploying the chosen model in a form that makes sense to the person consuming it; and monitoring that model for accuracy—but is much more multi-faceted than modern dev ops, and often requires more skillsets, as well.
To make matters worse, the pieces within the AI/ML development process are neither sequential or predictable—because we often find that as we complete a step, we learn something new about the data, causing us to go back to the beginning and address it.
The steps must all be treated as interwoven and interrelated components in order for the whole process to work in a seamless fashion. And that can be a head-scratcher to some.
3 – There are lots of teams to consider—and they must be aligned
As I said, a key piece to the AI/ML development process is knowing what you’re trying to achieve. That starts with defining the business problem, and continues with revisiting how that goal is being addressed on a regular basis.
However, even if everyone knows what the problem is, it’s no guarantee they are marching to the same drumbeat. Business initiatives aren’t always the same as the IT team’s initiatives, or the analytics team’s initiatives. And chances are you’ve also got individual systems and tools in place that can make it challenging to reconcile or streamline the overall process.
In other words, the process of productionizing AI/ML is not just a technical issue; it’s organizational, operational, and strategic, too. And underpinning all of that is data, which of course can be siloed, buried, messy, and full of holes.
Breaking the barriers to entry
While the roadblocks to productionizing AI/ML are real, you can overcome them. You just need to understand them—and then employ the right tools and methodology.
A big reason why I joined Domo more than a year ago was because its platform excels at addressing everything that makes AI/ML productionization hard.
Because it’s a solution designed to help organizations achieve Modern BI for All, it possesses features—including a dynamic 1-2 punch of MLOps and AutoML functionality—that not only help alleviate all three pain points detailed above, but free up smart people to focus more on projects that are interesting and add business value.
What’s more, it considers the inefficiencies associated with defining the business problem, gathering the data, developing the model, visualizing the results, and packaging and deploying to production through the following capabilities:
- Accelerating model management. Data scientists can set alerts that push notifications to users if production data science models are becoming less accurate over time. Monitoring model performance and model drift is critical—and needs to be implemented through a solution with automated alerts, notifications, and configurable thresholds to properly leverage the inferences models suggest for deriving business value and insights.
- Democratizing data prep and profiling. To achieve a performant, sustainable solution, speed, scale, and control are paramount. Automation and orchestration of data pipelines and complex workflows for business users such as data prep and exploration, data profiling, and arbitrary joins become key pillars to success. This accelerates the data lifecycle and greatly shortens time to value for models deployed in production.
- Aligning with current DSML processes. Make complex, algorithmic capabilities transparent to the business with the ability to embed raw Python or R code directly into the data platform and model pipeline while implementing row-level entitlement and governance. There’s no need to utilize third-party tools. This also allows you to track and visualize model performance, explainability, and cause-and-effect, seamlessly.
- Automating feature importance. Automatically generate and test ML models until a best fit is found. Expand your rapid prototyping capabilities in pursuit of a trustable answer, or perform sensitivity analysis against live production data.
Watch my recent talk at A2.A1 (directly below this paragraph) to learn more about these important processes and how to productionize AI/ML through personal lessons learned, telling anecdotes, and powerful case studies.