Exploring the Potential of AI in Process Improvement

MoreSteam's Founder and CEO, Bill Hathaway, gives his thoughts on how AI fits into process improvement, following a recent article published by the Harvard Business Review.

If you are old enough, you may remember the dawn of the internet age, when it seemed like we went from morning to mid-afternoon without even stopping for lunch. There was so much dramatic change, so fast. We went from idling to 60 miles per hour in a heartbeat. The minute the internet was invented, it was seemingly connected from everywhere to everything, and new internet service providers were giving away Netscape for free. I remember thinking – why? - how did that business make sense? New websites, email, e-commerce, search engines, and millions of experiments…the birth of the internet age was like the Cambrian explosion 530 million years ago when a huge variety of new animals suddenly appeared.

In the last year, with the rapid development and adoption of AI through large language models (LLMs), I can’t help but compare the feeling of disruption to the internet explosion, and there may be an even faster rate of evolutionary change, again with millions upon millions of ongoing experiments to learn what might be possible.

One of the promising areas of application for AI and LLMs is process improvement, a messy and labor-intensive undertaking involving two complex elements: humans and data. Recently, Harvard Business Review published an article on “How AI Fits Into Lean Six Sigma.” It’s a good start at identifying some of the possible applications. I believe there are other fruitful avenues for exploration (and it’s all exploration at this point because so many new possibilities are emerging).

Brainstorming Use for Generative AI in OpEx

Here are some of the areas of AI/Machine Learning applications that MoreSteam is actively investigating:

• Large language models such as ChatGPT are really good at analyzing textual information. They can be used to analyze Voice of the Customer (VOC) data from surveys, interviews, or support tickets. We’ve been using ChatGPT to summarize and group text into logical categories, accelerating the messy process of creating affinity diagrams. The same approach can be used to organize brainstorming data.
• If you’ve ever reviewed a project charter, you’ve probably noticed weaknesses in the Problem Statement and/or Business Case. Given the proper prompting and context, LMMs can be used to aid Problem Statement and Business Case development by suggesting improvements for greater clarity, specificity, or even strategic alignment. MoreSteam has already implemented this AI-driven functionality into the TRACtion project management system and the DMAIC project simulations used for training.
• On the subject of training, LLMs open the door to constructing much more realistic open-ended questions, providing a dynamically crafted custom response akin to what students would hear from a teacher, allowing us to move away from multiple-choice questions. That’s a much more realistic learning exchange since the real world is seldom, if ever, a multiple-choice experience.
• Exploratory data analysis is now possible with Chat GPT and other large language models. You can upload data and ask for a summary, including descriptive statistics, or use it to help decide what visualization tools to use (Bar chart v pareto). It can also convert what you want to explore into code to run on the dataset, to name a few examples.
• There are potential concerns about data security, prompting some organizations to restrict such practices, or run large language models inside their firewalls.
• One common struggle is determining which tool to use to answer a question of interest. LLMs can be trained to assist and help with interpretations and conclusions. Error-proofing the interpretation of statistical outputs greatly benefits those who do not frequently undertake statistical analysis. At MoreSteam, we’re expanding this capability within the EngineRoom data analysis application.

Note: You might wonder, as others have, how we develop critical thinking skills if software starts doing more and more of the critical thinking. It’s an important unresolved question.

• An emerging application of machine learning is root cause analysis of the faults in complex systems using classification or regression trees. This type of analysis previously required either a data scientist, expensive software, or both, but it has recently become more accessible. EngineRoom added this machine learning application earlier in 2023.

• Of course, finding the root cause of a problem does not equal solving that problem. Something in the process or product must be corrected, changed, or revised. Brainstorming of potential solutions can be turbocharged by large language models right now. Try typing this prompt into Chat GPT or another LLM: “What is the best way to design a retail check-out process to minimize queue time?”. You’ll probably be surprised by the quality of the response. There’s a lot of potential to tap into best practices and even generate guidance on resolving design trade-offs, a la TRIZ. If you ask ChatGPT, “How could I design an automobile to be fast and powerful and also have great fuel economy?” you will probably get a very comprehensive and coherent answer (with the usual caveat that there can always be some hallucinating). We’re exploring various ways to accelerate ideation using AI. Large language models are so fast and virtually free, so why not use them to help generate and sort/summarize ideas? AI probably won’t come up with THE idea directly, but it can help stimulate the creative connections that are at the heart of all great design. More to come on this…
• When implementing improvements, there are normally (always?) potential risks to mitigate. Examining potential failure modes is another messy human activity fueled by data and experience. Perhaps LLMs could play a role in imagining “what could go wrong if X happens?”.
• One step further would be to consider potential unintended consequences. As with solution brainstorming, AI probably won’t generate the definitive answer but could help spawn other ideas and connections. We’ll be investigating this.

That’s a few ideas. You will be able to come up with more. We encourage you to run experiments and share what you learn. One of the important experimental questions is the open issue of displacing critical thinking mentioned earlier; we all have to figure out how to design AI interactions to aid critical thinking rather than replace it.