UpTrajectory Review
Sankalp Arora's Inc. piece argues that warehouse operators have spent two years treating large language models and classical optimization as rival religions, when the teams actually cutting cost per pick run both on the same data. LLMs parse unstructured inputs—supplier emails, shift notes, weather, SKU descriptions—while optimization engines do the math on slotting, batching, and routing. The article's core claim is simple: the value lies in the handoff between the two, not in either alone. Arora, a PhD, writes from the optimization side of the house, which shows; he gives the math engine top billing and casts LLMs as the helpful translator. That framing flatters operations researchers, but it also understates how messy the translator's job is. Still, the thesis lands: a WMS that can read a vague note from a receiving clerk and re-plan the afternoon's pick waves without a human re-keying anything is a real operational gain, not a demo.
For a small distributor running 5,000 to 50,000 square feet, the stakes are concrete. Labor is 50 to 70 percent of warehouse cost, and the labor market stays tight. If an LLM can absorb the daily chaos—late trucks, damaged pallets, a customer's 'can we add 200 units to yesterday's order' email—and feed a clean, structured problem to an optimizer, you cut the planner's swivel-chair time and reduce the 'we'll just work overtime' reflex. The catch is integration debt. Most SMB warehouses run a legacy WMS, a spreadsheet, and a whiteboard. Bolting an LLM plus an optimizer onto that stack is a project, not a plugin. Arora's piece assumes you have data pipelines and API access; many operators do not. Budget for the plumbing, or the AI stays a slide deck.
What is genuinely new here is not the pairing itself—supply-chain vendors have demoed 'AI plus OR' for a year—but the insistence that the LLM's job is problem formulation, not decision-making. That distinction matters. Let the LLM turn natural language into constraints; let the solver guarantee feasibility. We agree with that division of labor. We are skeptical of the article's optimism about data quality. Warehouses are full of dirty master data: phantom SKUs, wrong dimensions, bin locations that were correct in 2019. An optimizer fed bad dimensions will produce elegant, wrong answers faster than any human. Arora waves at this; he does not dwell on it. The under-reported story is the data-cleanup sprint that must precede any of this.
The second-order effects cut unevenly. Large 3PLs with in-house data teams will compound their advantage; they can build this stack and offer 'AI-optimized fulfillment' as a premium line, squeezing smaller competitors. WMS vendors will respond by bundling LLM features, which raises software costs and renews pressure to upgrade from perpetual licenses to subscriptions. Workers will feel it too: better optimization means fewer 'deadhead' walks but also tighter rate standards. If your pickers already suspect the system is watching, adding an LLM that reads their shift notes will not build trust. Managers should plan the change conversation, not just the API contract. There is also a talent gap: the person who understands both prompting and mixed-integer programming is rare and expensive.
Watch three things over the next two quarters. First, whether your WMS vendor ships a credible LLM-to-optimizer connector or just a chatbot that answers 'where is SKU 1234.' Second, whether the ROI case in your own building survives contact with your data; run a two-week audit of SKU dimensions and bin accuracy before you sign anything. Third, whether labor reacts to tighter standards with attrition; measure pick rate and turnover together, not separately. If you want to test the concept cheaply, pick one narrow problem—say, re-slotting the top 200 movers—and prototype the LLM-to-solver handoff on that. Do not start with a full-building re-plan. The teams pulling ahead, as Arora says, use both tools together; the teams that fall behind buy two tools and use neither.
“LLMs and classical optimization have been treated as separate tools. The teams pulling ahead use them together.” — Inc. Magazine
Takeaway: Audit your SKU and bin data before you buy any AI-optimization stack; the solver will only be as good as the dimensions you feed it.
Excerpt from the original — Inc. Magazine
LLMs and classical optimization have been treated as separate tools. The teams pulling ahead use them together.