How I think about machines making decisions

The short version

• Some decisions have too many moving parts, too much uncertainty and too little time for a person to weigh well.
• In my PhD I used simulation and optimisation algorithms to make supply chain decisions, and they beat the policies the companies were using.
• That taught me to choose the tool by the problem: simulation to compare strategies, genetic algorithms and stochastic approximation to tune them.
• Today I do the same with traditional machine learning, such as decision trees, and with large language models. I choose between them by the cost of an error.
• I moved from academia into industry because I wanted real impact on complicated business problems, and consulting let me see manufacturing, healthcare and logistics.
• Machines that decide need controls: explainability, audit trails and human oversight. That is why AI governance is part of my work.

Why do I believe machines should help with complicated decisions?

I have always believed that some questions are too complicated for a person to answer unaided. A supply chain with several stages, uncertain lead times and uncertain demand is one. People are good at judgement and context; they are poor at holding thousands of interacting numbers in their head. I want to focus on problems where a better decision has a real, measurable impact.

What did my PhD show?

My PhD at the University of Plymouth, supervised by Dongping Song, Michael Roe and Sarah Tuck, studied two small manufacturers in China, one making aluminium and one making fine chemicals. I modelled their multi-stage supply chains with several uncertainties, used their own enterprise data, and tested strategies in simulation against what each company actually did. Then I used genetic algorithms to tune the strategies, including a multi-objective version that balances total cost against customer service level.

The later paper with Dongping Song, Integrated optimisation for production capacity, raw material ordering and production planning, reports what happened in the tests:

• All the improved strategies beat the company’s original approach.
• Optimised strategies cut cost by roughly 36% to 48% against the original, and planning capacity together with ordering and production cut cost by roughly 36% to 64%.
• Stochastic approximation was slightly better than a genetic algorithm in most cases.
• More shared information did not always help. How the information is used matters.
• The experiments used data from one company, so the results show what is possible rather than a guarantee.

How did that lead to LLMs and traditional AI?

The discipline carries over. Model the decision, define what “good” means, test strategies against the old way, then tune. Optimisation and simulation taught me to do that; I now apply it with other tools.

• Traditional machine learning, such as decision trees, gives rules you can read and explain. It suits decisions where someone must justify the answer.
• Deterministic algorithms suit calculations that must come out the same every time.
• Large language models suit interpreting requests and unstructured text.

My job is to assign each task to the right one, based on the consequences of a wrong answer. In my current role this hybrid design reached 99.9% accuracy in standard reporting and cut language-model spend by up to 30% through routing. See my work.

Why did I move from academia to industry?

I wanted real impact on complicated business issues. Consulting let me see many industries quickly, including manufacturing, healthcare and logistics, and showed me that the same kinds of hard decisions appear everywhere. I then moved inside companies to own the results: first at Veralto, leading automation and data transformation in global logistics, and now at MealSuite as Director of Data Analytics & AI.

What do I believe about AI governance?

If a machine is going to make or shape decisions, people need to see how, check it and override it. I set responsible-AI standards, model controls and data-governance rules so that teams can adopt AI in regulated settings, including sensitive healthcare data, and so that the results are audit-ready.