will develop into our digital assistants, helping us navigate the complexities of the trendy world. They’ll make our lives easier and more efficient.” Inspiring and completely unbiased statement from someone who already invested billions on this recent technology.
The hype is real for AI agents, and billions are pouring in to construct models that may make us more productive and more creative. Hard to disagree after I happily enjoy my morning coffee while Cursor is coding my unit tests. Yet, asking people in my network how they use AI of their day-to-day, their answers often mention anecdotal use cases, anywhere from “I take advantage of it to inform bedtime stories to my son” (I assume that may not even be a use case when you had more imagination) to “I take advantage of it to optimize my schedule” (Motion AI, please stop targeting me for the love of god).
As a Data Scientist, my mind goes forwards and backwards between two conclusions. The FOMO a part of me that doesn’t need to be late to the Robot revolution party, and the cynical one which thinks that there remains to be a protracted strategy to go before artificial intelligence actually becomes intelligent. To search out out which side of my schizophrenic personality I should bet on, I’m going to make use of an easy yet powerful framework: reviewing all of the projects I actually have worked on because the starting of my profession and assessing how 2025 state-of-the-art AI models could have helped.
Today, we return to 2018. I’m a candid summer intern at one of the disruptive startups in America: Rent the Runway.
What the Project was about
The Rent the Runway success center in Secaucus, NJ, was once the most important dry cleansing facility in the US.
Within the Summer 2018, as an Operations Analyst intern, I used to be given a fairly hard problem to take into consideration: on a regular basis, the success center was receiving 1000’s of units back from throughout the country. All of the items needed to be first inspected, then would undergo an intensive cleansing process, before being dried or receiving some special treatments. This may very well be:
- Spotting if the garment was stained throughout the rental
- Pressing if it was too wrinkled and needed to be ironed
- Repairing if it had been damaged
Most of those tasks were done manually by different departments, and required specialized employees to be available as soon as the primary batch of units were reaching their department. With the ability to predict days ahead what volume of units would must be processed (and when) was crucial for the success center planning squad, with the intention to be certain that that each operations team could be staffed appropriately.
The complexity of the flow made it even trickier. It was not only about predicting the inbound volume, but in addition assessing what a part of this inbound volume would require special treatments, where and when bottlenecks could appear, and understanding how the work done at one department would impact the opposite departments.
The 2018 Solution
At this point you might wonder: given the complexity and the stakes of the project, why was it within the hands of a young inexperienced intern? To be fair, during my 10-week summer internship, I only scratched the surface and wrote an insanely complicated Pyomo script that was later refined by a more senior Data Scientist, who spent two years on this project alone.
But as you possibly can imagine, the answer was this huge optimization model taking as an input the inbound volume forecast for daily of the week, the common UPH (units per hour, i.e the variety of units that could be processed in an hour) at each department, and a few assumptions on the proportions of units that may require specific treatments. The most important constraints were on the timing and regularity of the shifts, and the variety of full time contracts. The model would then output an optimized labor planning for the week.
How AI could have helped
Let’s re-clarify things first: you is not going to see words like “AI-enthusiast” or “LLM believer” in my LinkedIn bio. I’m pretty skeptical that AI will magically solve all our problems, but I’m excited about seeing if with today’s technology, one other approach could be possible.
Because our approach was, you might say, pretty old-fashioned, and required months and months of refinements and testing.
The most important limit is the static aspect of the answer. If something unexpected happens throughout the week (e.g a snow storm that paralyzes the logistics in some parts of the country, delaying among the inbound volume), a whole lot of assumptions of the model must be modified, and its results have gotten obsolete.
This can be a solution that requires data scientists to go deep into the weeds, as a substitute of counting on an out-of-the-box framework, to depend on a whole lot of assumptions and to spend time maintaining and updating these assumptions.
Could AI provide you with a very different approach? No.
For this particular problem, you clearly need an optimization model, and I’m yet to examine an LLM with the ability to handle a model with such complexity. One could propose a framework with an AI agent acting as a General Manager, and counting on sub-agents to handle the planning of every department. But that framework would still require agents to have tools that allow them to unravel a posh optimization model, and the sub-agents would want to speak because the situation of 1 department can affect all of the others.
Could AI significantly enhance the “human-generated” solution? Possible.
It’s at this point pretty obvious to me that LLMs wouldn’t make the issue trivial, but they might help improve the answer in multiple areas:
- To start with, they might help with reporting and decision making. The output of the optimization model might need a business sense, but making a choice out of it is likely to be hard for somebody with no strong understanding of linear programming. An LLM could help interpret the outcomes and suggest concrete business decisions.
- Secondly, an LLM could help react faster to certain unexpected situations. It could for instance summarize information on events that might have an effect on the Operations, equivalent to bad weather in some parts of the country or other issues with suppliers, and as such, recommend when to rerun the planning model. That’s assuming it has access to good quality data about these external events.
- Finally, it is feasible AI could have also helped with making real time adjustments to the planning. As an illustration, it is usually predictable based on the garment characteristics whether or not they would require special care (e.g a cotton shirt will all the time must be ironed manually). Having a VLM scanning every garment on the receiving station could help downstream departments understand how much volume they need to expect hours upfront.
Could AI enable Data Scientists to take care of and update the model? Yes!
It is actually hard to disclaim that with tools like Copilot or Cursor coding and maintaining this model would have been easier. I might not have blindly asked Claude to code every constraint of the Linear Program from scratch, but with AI code editors being smarter than ever, modifying and testing specific constraints (and catching human errors!) could be easier.
My conclusion is that an LLM in 2018 wouldn’t have trivialized the project, even though it could have enhanced the ultimate solution. Nevertheless it shouldn’t be not possible to imagine that a number of years (months?) from now, agents with enhanced reasoning capabilities shall be sophisticated enough to start out cracking a majority of these problems. Within the meantime, while AI could speed up model iterations and adjustments, the human judgment on the core stays irreplaceable. This serves as a useful reminder that being a Data Scientist isn’t nearly solving mathematical or computer science problems—it’s about designing practical solutions that meet evolving, often ambiguous and never so well defined real-world constraints.