That familiar supply chain puzzle is getting a new kind of help from artificial intelligence.

New research co-authored by Parshan Pakiman, assistant professor of operations management in the University at Buffalo School of Management, explores how large language models, the technology behind tools such as ChatGPT, can work together with data and commercial optimization technologies to design better inventory policies.

The idea is simple, even if the math behind it isn’t: Let AI suggest new ways to make ordering decisions, then make those ideas prove themselves. 

Importantly, the AI is not being asked to decide how many units a company should order tomorrow. Instead, it is being asked to design the rule itself: a decision rule that determines how much to order based on the company’s current inventory position.

Pakiman’s research combines data-driven optimization, reinforcement learning and advances in AI to develop new methods for tackling emerging sequential decision-making problems in inventory control, healthcare management and revenue management.

The researchers created a framework that assigns distinct roles to AI, optimization and data.

The LLM proposes the structure of a new ordering rule: what inventory information it should pay attention to and how that information should influence an order. An off-the-shelf optimizer then adjusts the parameters, or numerical “knobs,” inside that rule using demand data. Then, demand simulation tests how well the resulting rule performs. The strongest ones are fed back to the LLM, providing guidance that helps it propose better rules in the next round.

It’s a bit like an idea lab with a very strict scorekeeper, but the scorekeeper also helps determine which ideas the lab explores next.

In computational experiments across 30 lost-sales inventory scenarios, the average cost reduction relative to an optimized base-stock policy (a standard inventory replenishment rule) grew from 17.5% after the first generation of search to 30% after 10 generations.

The results also suggest that optimization does more than fine-tune AI-generated ideas. By determining which rules perform well and advance to the next round, it creates a self-correcting feedback loop that guides the LLM toward better ideas.

The resulting rules were not mysterious black boxes. Many used recognizable inventory-management ideas such as limiting unusually large orders, putting different weights on inventory scheduled to arrive at different times and using thresholds to determine when and how aggressively to replenish. 

Even more striking the discovered rules worked well beyond the settings in which they were found, which is important because real-world environments often differ from those used to develop a model. The researchers tested three LLM-proposed policy classes across more than 10,000 new inventory environments, keeping each rule’s basic structure but retuning its parameters for the new setting. Each reduced average costs by more than 21% relative to the optimized base-stock benchmark.

Thus, AI was not simply discovering one-off tricks that worked only in the environments where they were found. It was identifying decision rules that could remain useful as important operating conditions changed. 

That points to a practical role for generative AI in operations: Rather than asking an LLM to make each decision, companies could use AI to broaden the set of decision rules, while commercial optimization and simulation technologies do the hard work of testing what holds up.

Here, generative AI does what it is especially good at: exploring a large space of possible ideas. Optimization and historical data provide the discipline needed to determine which of those ideas deserve to survive.

For the School of Management, Pakiman’s research reflects a broader approach to making AI useful where business happens: pair the technology with domain expertise, data, and rigorous methods for evaluating its output.

The school’s Center for AI Business Innovation brings together faculty research, student talent, and industry needs through AI education, student consulting, research support, and business collaboration. The focus is practical: take a fast-moving technology out of the abstract and put it to work on problems organizations face. 

That same thinking is showing up in executive education. The new Managing with AI and AI Mini MBA programs help managers apply AI to everyday work and to decisions across operations, people, accounting, marketing, finance, and strategy. 

The school was also recently recognized as an exemplar for integrating AI into global business education. The common thread is clear: AI becomes more useful when it works alongside business judgment, data, and rigorous testing.

Pakiman’s research brings that idea down to the shelf level, quite literally. Start with a messy business problem. Let AI widen the field of possible answers. Use optimization and data to determine what works. Then refine and test again. 

That’s where promising technology starts to become a business solution.