2023 was a banner year for AI with Chat GPT’s debut. Generative AI sparked a firehose of AI articles, commentaries, and even dire prophecies and, as Eric Siegel — the bestselling author of The AI Playbook: Mastering the Rare Art of Machine Learning Deployment — notes, still gets all the airtime.
According to Siegel, “The majority of generative AI’s spotlight is predictive AI’s value proposition, which has been around for decades.” He believes that organizations should invest at least as much in investigating or piloting projects with predictive as with generative AI.
It’s time we asked ourselves why so few AI dreams — the models or “golden eggs”— fail to hatch or deploy.
Because all too often, AI projects fail to deploy. They create potential value by way of generating the model but fail to capture that value because, “They don’t get put into the field; they don’t change operations.”
How can organizations remedy this?
AI Dreams Vs. AI Deployment
Siegel, using UPS as an example, illustrates the gains to be made from actual AI deployment.
UPS wanted to optimize package deliveries by predicting demand and planning routes more efficiently. They tried to decrease the hours the trucks were out and the miles they needed to drive. Except each evening when they began to assign and load the trucks, they didn’t have complete information because some packages expected for next-day delivery would have yet to arrive.
To bridge this gap, UPS predicted likely deliveries based on historical data. With that more complete, albeit imperfect view of tomorrow’s deliveries, they were able to do a better job. This system, called Package Flow Technology, paired with another system, Orion, provided each truck with a relatively optimal route, telling the driver exactly where to go, saving UPS 185 million driving miles, $350 million and 185,000 metric tons of emissions. While a resounding success, the deployment of this project was not clockwork.
The tech team had to convince senior executives to make this change to large-scale operations based on predictions. Predictions are probabilities. So, that idea of acting on probabilities, of embracing probabilities, was central to the challenge of deployment and critical to driving better decisions. The exciting “rocket science” portion of any predictive AI project might be in training the model. But without a focus on deployment, the rocket goes nowhere. Instead, it ends up in a “no man’s land” between business and data teams, each saying it’s not their role.
The data scientist, Siegel says, believes the model will be deployed because its value is obvious. Meanwhile, the business stakeholder thinks, “That’s a technical matter; I’ll leave that to the data scientists.”
The solution to this conundrum lies in two parts. One, all business stakeholders should cultivate a semi-technical understanding around these predictive projects. And the other is a paradigm that Siegel calls “BizML,” a framework or business practice, to turn these machine learning models into business-building realities.
[Watch Eric Siegel’s full keynote and presentation here: The AI Playbook—How to Capitalize on Machine Learning]
BizML Makes AI Actionable
BizML, a 6-step practice, is foundational to any AI deployment. He argues machine learning (ML) projects go astray because their stakeholders focus too often on the technological fireworks — the “rocket science” of predictive models, not the rocket launch. By following the six steps of BizML—value, target, performance, fuel, algorithm, and launch—organizations can bridge the gap between technical capabilities and business objectives, unlocking the full potential of AI to drive innovation, efficiency, and competitive advantage.
Value: Establish the deployment goal
Define the business value proposition: how ML will improve operations in order to improve them.
Target: Establish the prediction goal
Define what the ML model will predict for each business case.
Performance: Determine benchmarks to track
Establish the salient benchmarks to analyze during both model training and model deployment and determine what performance level must be achieved for the product to be considered a success.
Fuel: Prepare the data
Use the model to render predictive scores and then act on those scores to improve business operations.
Algorithm: Train the model
Generate a predictive model from the data
Launch: Deploy the model
Use the model to render predictive scores and then act on those scores to improve business operations.
Upkeep: Maintain the model
Monitor and periodically refresh the model as an ongoing process.
Monitoring and maintaining the model that’s deployed is a lot like launching a rocket. Once Nasa launches a rocket, its mission control doesn’t go to bed. They work round the clock to keep their astronauts alive. It’s no different for a training model that would require constant monitoring for its performance that will eventually degrade. Which means you have to refresh it and train a new model over more recent training data.
All the above steps require deep collaboration with business stakeholders and do not occur linearly – backtracking happens. “But here’s where the faucet and the hose have to connect,” Siegel reminds us. Business and data science teams have to work together to operationalize the predictions. Only when the model gets deployed is when you actually improve business. Because you don’t capture value until you act on it. “Deploy it to make a difference. Operations don’t improve unless they change” he says. “We need to use those predictions, integrate them, put them into operations, operationalize it. The part that’s so often not achieved,” Siegel emphasizes.
So, in a nutshell, the data comes in from the left, and the core method, “the rocket science,” the machine learning part, the predictive modeling, learns from it. It ascertains patterns or rules from the data and generates a model. But the true magic happens only when what’s taken from the data is deployed in its use or, in other words, when the AI and machine learning rubber hits the road.
The Road to AI Deployment Starts Here
Integrating AI into your business strategy is no longer an option but a competitive necessity. Success in the future belongs to those who deploy AI, aligning it with their strategic goals to create more agile, efficient, and forward-thinking solutions. With over 25 years of experience in Microsoft solutions, we’re here to help you future-proof your organization through successful AI integration.
Ready to get started with AI?
