Warren B. Powell
Professor Emeritus, Princeton University
Chief Innovation Officer, Optimal Dynamics
On July 10 2025, I gave a presentation at Toyota’s North American Headquarters with the title:
The Road to Digital Automation at Toyota:
Learning How to Be an Informed Consumer
The talk was designed for a general business audience. It attracted over 300 attendees, and was warmly received by an enthusiastic and receptive audience who appreciated the message that applied to every person who was “making decisions.”
The core messages included:
- There are 7 levels of AI, where LLMs such as ChatGPT are level 4. Level 5 is traditional deterministic optimization, and level 6 covers sequential decisions, the central topic of the talk. The concepts are illustrated with business-friendly images.
- The central theme of the talk is the introduction of a step in the modeling process I call “Framing the Problem” which involves describing real-world problems (any real-world problem) in English (no math) that provide a bridge from the original problem to my universal modeling framework which can be used to model any sequential decision problem. Framing the problem starts with identifying metrics, decisions and uncertainties.
- I present the universal modeling framework which fits entirely on one powerpoint slide, but only to make the case that the metrics/ decisions/ uncertainties form the basis of a sequential decision problem.
- I give a very brief presentation of the four classes of policies for making decisions. Most important from the perspective of business is that “policies” help identify what information is needed.
- I make the point that large language models (level 4 AI) are useful, but primarily for administrative support. If you want to improve physical systems, you need to make better decisions, which means level 6 AI (sequential decisions). In automotive terms, LLMs are like windshield washers, while sequential decision problems are like engines. If you want to have an impact, you have to make better decisions.
- I close with a 5-step process for designing and implementing methods for making decisions.
I have recorded the 90-minute presentation in 5 segments. You do not need to watch all of these in sequence. Feel free to pick the topics that are most interesting to you:
Part 1 – Introduction and the 7 levels of AI – The presentation makes the distinction between machine learning (levels 2, 3 and 4), where the goal is to match a training dataset, versus levels 5 and 6 for making decisions which require an explicit model of the underlying problem, including user-specified metrics such as cost minimization.
Part 2 – Framing the problem – This is the heart of the talk, where I make introduce the need for people who can translate real applications (such as business problems) into language that captures the information needed by a modeler.
Part 3 – The universal modeling framework – I introduce the notation for the universal modeling framework on a single powerpoint slide. Although it is quite simple, it is not necessary to know the framework. But this framework is guiding the questions that need to be answered if a decision problem is going to be solved on a computer.
Part 4 – Making decisions – I give a very brief overview of the four classes of policies (with no math!). This material is being taught by only a few universities, but it is critical for solving the complex problems that arise in problem settings such as supply chain management (among others). This part also covers the importance of including uncertainty in forecasts, along with the need to model the ability to make decisions in an uncertain future.
Part 5 – The path to digital automation – I describe five steps in the path to implementing the process of making decisions. I recommend watching Part 2 before watching Part 5, but it is not essential.