Strategic AI Transformation

Large metropolitan areas require a lot of gasoline and rely on distribution from the supply tanks to the gas stations. Distributors buy and deliver gas from the suppliers and sell to the gas stations, making only pennies on each gallon. Distributors have many variable costs besides just the gasoline they deliver, such as driver, truck operation (fuel, maintenance, repairs), and staff operations costs that must be considered in planning and implementing gasoline distribution.

Keeping costs low is crucial to making a profit. Distribution dispatchers were spending hours calculating the most cost-effective truck routes based on the complex balance of: prices at various suppliers, locations of the gas stations, the routes they needed to take, and their operational costs. Taking the costs into consideration, simply picking up gas at the cheapest supplier was not always the best plan, especially if there were long waits or distances required to get the cheaper gas.

Long lines at the cheaper supplier added hours of waiting, which increased driver costs and affected the delivery schedule. It was often more cost-effective to pay slightly more for gas at a supplier facility with no line than to have the driver wait two hours in line at a cheaper supplier. Prices often fluctuated in the middle of the delivery schedule, adding additional randomness and problems.

When exceptions occurred, and they frequently did, dispatchers had to recalculate the distribution plan quickly. These recalculations were time-consuming and fraught with costly mistakes.

This was a well-defined least-cost routing problem suited for an AI solution. Given all of the parameters and in-the-moment changes, AI would need to create new schedules within seconds.

The project started with observational research.

Larry and his team rode along with drivers, and they made office visits with dispatchers to more accurately understand the task domain and the problem from the users’ perspectives. This research drove our task analysis and subsequent task optimization, where we explored how AI could serve the users.

We then designed mockups to test an AI-First design concept with real users, who were all emphatically pleased with the assistive AI approach. We completed the project by collaboratively reviewing and turning over the design specifications to the client’s developers.

This was more than a static calculator problem since the variables and priorities were in a constant state of flux, demanding a more fluid AI analysis capability. Initial testing of the prototype proved this solution worked well, with dispatchers claiming that it saved them six hours of work each day. This also eliminated errors since the calculations were performed by machine instead of by hand.

Aircraft and pilot scheduling in the US Air Force usually takes two people about a week to complete. It typically requires immense effort to develop a complex schedule that meets all the training objectives and priorities. Last-minute rescheduling is even more difficult, resulting in scrubbed missions or missed training. Schedulers use a myriad of tools and processes to solve this problem, such as disjointed apps, puck boards (a type of whiteboard), and spreadsheets. 

Ultimately, the complex scheduling task is performed in the heads of humans. The tools they use don’t create the schedule; they only provide a way for the user to represent or share a schedule created in their head. A lack of standardization across different squadrons and units means that every scheduling process is different, resulting in a steeper learning curve when aircrews transfer to new units, and making it difficult for schedulers to assist sister squadrons.

Now imagine a day when a squadron’s flight scheduler arrives at the office and creates the next week’s flight schedule before their coffee gets cold. If one of the pilots scheduled for tonight’s inflight refueling exercise calls in sick, the scheduler marks the pilot as “unavailable,” and the system recalculates and schedules another pilot who needs night refueling practice. Total time on task: five minutes. 

Artificial Intelligence/Machine Learning to the Rescue

Led by Larry Marine, the user experience team conducted observational user research at several squadrons. They discovered that units typically have the same scheduling problems but solve them differently with custom solutions, custom apps, and custom spreadsheets. The biggest issue was that, regardless of the tools used, the process placed high demands on user cognition, thus relegating each solution to the limitations of each individual scheduler’s capabilities and processes. 

A solution to this problem relies on reducing the demands on human cognition by developing a system that balances complex needs and priorities to create a workable schedule. From the research, Larry’s UX team defined a single, enterprise-wide Smart Scheduler Paradigm that can be applied to various scheduling tasks across the Air Force, such as pilot scheduling, Airman training, and readiness forecasting/preparation.

The Smart Scheduler Paradigm is an AI/ML algorithm designed to do the work for humans, effectively reducing the cognitive burden of balancing varying assets (planes, classrooms, etc.) and personnel needs (vacation, dental appointments, etc.). 

After a system is primed with rules-based objectives, assets or resources, personnel, and adjustable parameters and priorities, the system is designed to automatically generate a suggested schedule. Users can then adjust the parameters, such as indicating a pilot is out sick or that a plane is down for maintenance, which would trigger the system to recalculate the schedule.

This Smart Scheduler Paradigm will reduce the typical pilot scheduling efforts from around 60 hours per week to roughly 30 minutes. Other scheduling tasks will likely see similar benefits.