What kind of AI decides which trucks carry which goods, how a factory schedules its production lines, or how an airline crews its flights? According to Nevra Ledwon, who has spent 25 years in mathematical optimization, it's a field of AI that predates the chatbot era by decades, runs quietly behind some of the most measurable ROI in business, and still cannot be replicated end to end by generative AI, even when the AI has been trained on every operations research textbook that exists.

That gap, between what generative AI can do and what a senior optimization expert actually brings to a hard planning problem, became the center of this episode of Unscripted. Ledwon walked through what optimization actually is, why prediction and decision-making are not the same thing, and how she's using generative AI to make a historically expensive, inaccessible field usable by companies that could never afford it before.

Who is Nevra Ledwon?

Nevra Ledwon has spent her career in mathematical optimization, a specialized branch of operations research focused on planning, scheduling, routing, and resource allocation rather than prediction or language understanding. She has sold and delivered optimization software and consulting to companies across logistics, manufacturing, hospitality, and transit, including work with a major hotel brand on revenue management pricing decades ago. She now works on bringing generative AI into the front end of optimization consulting, with the goal of making a field long reserved for companies with the budget for teams of PhDs accessible to the mid-market.

What is mathematical optimization, and how is it different from predictive AI?

Ledwon defines AI loosely as computers doing work for us, which places optimization inside the same broad category as generative AI while making clear the two solve fundamentally different problems. Predictive AI, including the recommendation engines behind streaming platforms and e-commerce, tries to anticipate what will happen. Optimization takes that forecast as an input and prescribes what to actually do about it, narrowing down billions or trillions of possible plans, schedules, or routes down to one based on a defined set of objectives and constraints. Real-world examples she points to include how retailers decide which products go in which stores, how delivery companies route trucks, how warehouses sequence picking paths for workers or robots, and how airlines and hotels price seats and rooms.

Why couldn't generative AI replace her own operations research experts?

About a year ago, Ledwon set out to test whether generative AI, trained on the same textbooks her PhD-level consultants studied, could facilitate the multi-week stakeholder interview process her firm normally runs at the start of every optimization project, then convert that interview directly into a working model. It couldn't. No AI system she tested came close to reliably translating a real business problem into an optimization model that would actually run and generate a usable plan. The experience changed her mind about how much of that expertise lives outside textbook knowledge, and gave her, in her words, a new respect for what a senior operations research professional brings to a project that generative AI alone does not replicate.

What's a concrete example of optimization delivering measurable ROI?

Ledwon cites two examples with hard numbers attached. A large automotive company set a goal of reducing inbound transportation costs by 2 percent through better logistics routing and truck-loading sequencing. The optimization project delivered a 10 to 15 percent reduction in miles driven instead, along with lower fuel costs, reduced CO2 output, and less product damage from more efficient loading. A separate supply chain project for a client processing waste cheese into pet food powder produced a seven-figure reduction in cold storage costs from a project that cost roughly a hundred thousand dollars to implement, by optimizing which production facility and truck route to use for each order based on transit time and storage trade-offs.

Why is a soccer league's schedule harder to solve than it sounds?

One of the clearest illustrations of why optimization resists a simple AI shortcut came from a European soccer league that needed a season schedule for 18 teams playing each other home and away, subject to roughly 20 constraints, including limits on how often marquee teams could play in the same round and rules against three consecutive home games. The league needed a thousand compliant schedules generated for a random draw. Ledwon's team found the number of possible arrangements exceeds the number of atoms in the universe, making brute-force computation impossible. Solving it required multiple failed mathematical approaches before a developer found a way to split the problem in two, first generating home-and-away patterns for each team, then matching teams together round by round. Ledwon describes this kind of nonlinear, trial-and-error problem-solving as very hard to simply hand to a generative AI model and expect a working answer.

Does optimization make supply chains more resilient or more fragile?

Ledwon argues optimization makes supply chains more resilient, provided resilience is clearly defined. One major lever is refresh frequency: many companies only rebuild their plans weekly because the process depends on manual human effort, whereas the ability to refresh plans multiple times a day allows a business to adjust in near real time as conditions change. She also frames resilience as a function of planning across three time horizons simultaneously: strategic decisions like warehouse location and inventory stocking levels, tactical weekly adjustments, and near-real-time fulfillment decisions made as orders come in. Skipping the strategic or tactical layers, she says, is often why a company cannot fulfill an urgent order without excessive shipping costs.

How is generative AI actually being used in optimization, if not to replace the experts?

Rather than replacing operations research consultants, Ledwon's company uses generative AI to compress the assessment phase of a project, the multi-week process of interviewing stakeholders to understand what a good plan or schedule needs to account for. Generative AI's broad exposure to prior use cases allows it to ask the right dynamic follow-up questions based on a client's industry and prior answers, whether the client is knitting socks, fermenting yogurt, or growing cells for a biotech application, each of which introduces different constraints an interviewer needs to know to ask about. The goal is to eliminate weeks of assessment time while still handing the operations research modeler exactly the information needed to build the model, and she sees future potential for AI to take on adjacent tasks like building solution checkers, a part of the process she says operations research experts generally dislike doing themselves.

What happens to operations research experts if AI keeps improving?

Ledwon does not expect generative AI to fully automate optimization consulting. Her best-case estimate is that AI might eventually automate roughly 80 percent of 80 percent of problems, which nets out to around 64 percent, still leaving a substantial share of genuinely hard, creative problems for human experts. She also points to a layer of the process that AI cannot replace regardless of capability gains: understanding what stakeholders actually care about. She cites a German transit agency's effort to build driver and pilot schedules around employee satisfaction and reduced turnover, work that starts with human interviews about what makes a job feel sustainable, not a mathematical formula.

What do most people get wrong about AI in operations?

Ledwon names two recurring mistakes. First, many companies chase diminishing returns trying to perfect their predictions instead of accepting that predictions will never be perfect and building the ability to adjust quickly when reality diverges from forecast, what she describes as being prescriptive rather than only predictive. Second, she argues too much attention goes to which algorithm to use, when the harder and more consequential work is translating a messy, real-world business problem into a solvable set of equations in the first place. Algorithm choice, in her view, is often the less important decision.

What is Nevra Ledwon paying attention to that most people aren't?

She is watching for the point at which generative AI stops answering only the narrow question it's asked and starts challenging the premise behind it, for example, questioning whether a product should even continue to be stocked rather than simply optimizing its logistics. She ties this to the limits of current context windows and wonders whether structures like knowledge graphs might eventually let AI bring genuine systems-level thinking to a business problem instead of staying confined to the frame of the original question.

FAQ

What is the difference between predictive AI and optimization? Predictive AI forecasts what is likely to happen. Optimization takes that forecast as an input and prescribes the best plan, schedule, or route to follow given a set of goals and constraints, narrowing enormous numbers of possible options down to one workable solution.

Can generative AI replace operations research experts? Not currently, according to Nevra Ledwon. She tested whether generative AI trained on operations research textbooks could translate a real business problem into a working optimization model and found no AI system came close, even though the same knowledge exists in the textbooks her human consultants studied.

What kind of ROI does mathematical optimization typically deliver? Examples cited include a 10 to 15 percent reduction in delivery miles driven for an automotive company against a 2 percent goal, and a seven-figure reduction in cold storage costs from a supply chain optimization project that cost roughly a hundred thousand dollars to implement.

Does better optimization make supply chains more resilient or more fragile? More resilient, according to Ledwon, primarily by enabling more frequent plan refreshes and by planning across strategic, tactical, and near-real-time time horizons simultaneously rather than reacting only in the moment.

How is generative AI being used in optimization consulting if it can't build the models itself? It's used to compress the stakeholder interview and assessment phase of a project, asking the right dynamic follow-up questions based on the client's specific use case, which can eliminate weeks of upfront work before the actual optimization model is built by a human expert.

Listen to the full conversation with Nevra Ledwon, and catch every episode, on the Unscripted with Jeff Pedowitz page.