Episode 11 — How AI Is Changing Supply Chains

Introducing Edwin de Boer

Sunil Baliga: This week on Token Drop, we have a very special guest: Edwin de Boer, an expert in the supply chain world. Sajjad and I knew Edwin from a previous company here in the Valley — he's since moved to Europe. We thought a conversation about supply chain would be a great fit for the podcast. Standard disclaimer: anything discussed today reflects the panelists' personal opinions, not necessarily those of their companies.

Supply chain, to me, is one of the most critical functions in any company — look at Tim Cook's background at Apple, for example. It matters because it has both a revenue impact and a cost impact: fall short on supply and you lose revenue; carry too much inventory and you're eating unnecessary cost. That dual exposure is exactly why AI in supply chain is such a high-stakes area — a mistake at a large company with significant revenue can be very costly in either direction. Edwin, what's your perspective on AI and supply chain — where do you see the real challenges and opportunities?

What do we actually mean by "AI" in supply chain?

Edwin de Boer: Thanks for having me, Sunil. I've spent a significant part of my career in supply chain, both in high-tech electronics and in med-tech healthcare. Supply chain has become something of a board-level topic in recent years — largely because of the disruptions we all lived through with COVID, and beyond that, some very visible supply chain shocks. Everyone remembers the container ship stuck sideways in the Suez Canal. That single event triggered a wave of urgent questions: are my goods on that ship? Are they stuck behind it? If the ship now has to reroute around the Cape, how much longer will my goods take to arrive?

Those questions all came back to one thing: what data visibility do we actually have into our supply chain, and how quickly can we understand the real consequences of a disruption? That's where I see AI's critical role — the ability to look at data, understand its consequences, generate real insight, and share that insight across the company effectively.

I'd add one caveat up front, though: "AI" is such a generic term. As we get closer to AI's true value, I think we'll naturally get more specific about what we actually mean by it. A lot of what used to be called "digitization" now gets rebranded as AI, but I don't think that's really the AI value we're talking about here. There are many legitimate definitions of AI, and all of them add value somewhere in the supply chain — but the most recent development, agentic AI, is where I believe the real value is going to come from over time. Whether it's Red Sea disruptions, scenario planning, or the trade impact of globally applied duties and tariffs, there's enormous opportunity to use AI to better understand where your costs sit and how they affect revenue.

Where agentic AI fits: plan, source, make, deliver

Sajjad Khazipura: That's great context — we've heard people call 2026 "the year of supply chain AI." You mentioned agentic AI and volatility — and that volatility hasn't gone away since COVID; if anything, we're seeing more of it now. That must be a real challenge. Can you speak to how you're thinking about applying agentic AI specifically to these challenges, especially in a regulated industry like MedTech, which adds another layer of complexity?

Edwin: The opportunity is broad, but I think we're still at an early stage — and that's part of what makes this hard to talk about precisely. When people discuss AI in a corporate or supply chain context, there's a lot of debate: are we talking about real, adopted use cases that have meaningfully changed organizational productivity? Or are we talking about small incremental steps versus genuinely game-changing transformations that make us rethink how we design the supply chain and reallocate resources?

The clearest example is probably procurement. If you simplify supply chain into four stages — plan, source, make, deliver — you plan what you want to sell, source the components you need, make the product, and deliver it to its destination, whether B2B or B2C.

In the plan stage, there's real opportunity to optimize through scenario-driving and better forecasting, though that's less about agentic AI specifically. In the source stage — procurement — there's a genuinely interesting agentic AI opportunity: driving down component costs, managing procurement for a subset of your supply needs, and potentially using multiple AI agents to actually negotiate prices and volumes directly. I believe a meaningful share of procurement — whether in a regulated environment like MedTech or in high-tech electronics — will eventually involve agentic AI on both sides of the negotiation: your agent and your supplier's agent working out a better outcome together, freeing up valuable human resources from transactional work so they can focus on higher-value decisions across the range of scenarios you're facing. I see the same kind of opportunity extending into make and deliver as well.

Beyond agentic AI specifically, there's also a lot of opportunity around large language models more broadly — how we discuss and determine the best outcome given a set of inputs. A lot of people already use tools like Copilot or ChatGPT within a corporate environment to speed up transactional work. Overall, I think agentic AI specifically is going to make our transactional work far more efficient and let us engage with our broader ecosystem more effectively.

The accuracy problem: why "just run it again" doesn't work in supply chain

Sunil: I used to work in semiconductors, where I was in marketing and worked closely with supply chain teams on forecasting. Every conversation with those folks was defined by precision — if you asked how much inventory was on hand, "approximately" wasn't an acceptable answer. You needed to be exact.

I was talking recently with a friend who runs an AI startup selling into supply chain, and I asked him about accuracy. He told me that if someone using his software gets an answer they're unsure about, they just run the question again. That struck me as fundamentally inconsistent with how I know supply chain people to operate — because a hallucination is a probabilistic event. If you get an uncertain answer and simply re-run the query, how do you know which run was the hallucination? Maybe the first answer was wrong; maybe the second one is. That seemed to me like one of the most critical issues in applying AI to supply chain: the actual accuracy of the answers.

Edwin: I completely agree — trust is going to be fundamental to how AI gets adopted in a corporate environment, and the accuracy of outcomes matters enormously. People sometimes laugh it off — "I asked a question and got a dumb answer, so I guess I can't fully trust AI" — and everyone has a personal anecdote like that. But in a corporate environment, how do you actually leverage AI if you can't trust the outcome of what you're asking it? That's a fundamental challenge, and the more we can build solutions that reduce hallucinations and genuinely embed a real level of trust into the responses, the more likely we are to see real organizational adoption.

Otherwise, people revert to the old way of working the moment they conclude, "this doesn't make sense, this isn't what I need." I think AI organizations broadly need to think through: is there a way to represent an evolving confidence level, a percentage of accuracy, some quantified sense of risk attached to a given response? If you're doing scenario modeling in a planning environment — best case, worst case, and what that means for your inventory decisions — and I've got five satellite warehouses and need to decide how much of a manufacturing run goes to the main warehouse versus the satellites, I can never make that an accurate decision if I can't trust the scenario planning output. Trust has to be built up over time, but it starts with acknowledging that the output might not be 100% accurate and attaching some real range or risk level to it rather than presenting a single confident number.

Sunil: Going in with eyes wide open.

Edwin: Exactly.

Two-way trust: why supply chain visibility has to go deep

Sajjad: Follow-up on that — trust is a two-way street. You can trust the machine, but only to the extent you've given the machine enough visibility in the first place; the system needs the same degree of visibility you might or might not have actually provided it. Can you speak to the layers of visibility here? You've got your Tier 1 suppliers, but those Tier 1 suppliers are often sourcing from Tier 2 suppliers, who may be sourcing raw materials from Tier 3 providers — there's a deep chain on the supply side. On the delivery side, you're moving from warehouses to retail and wholesale outlets and beyond. Sometimes even that level of visibility isn't enough — you need to know which specific truck, train, or ship your goods are on, where that ship actually is, and what weather or geopolitical events are unfolding during transit. What's your view of that landscape?

Edwin: It's an interesting question, because I think adoption increases significantly once you can actually quantify the confidence or trustworthiness of an output. Take your delivery example — ocean containers crossing the sea. There's genuinely strong AI technology today that can show you the real GPS location of your freight on a container ship somewhere in the Atlantic, and use that data to predict how long it will actually take to reach its destination.

Historically, our supply chain processes relied on standard lead times — four weeks from Rotterdam to New York, six weeks from Singapore to somewhere in the UK — fixed estimates set at the start of the transportation cycle with essentially no real confidence attached to them. But once you're building up historical data and can pinpoint your container's actual GPS location day by day, you can meaningfully increase the accuracy of your prediction — flagging, for instance, that you're not going to make the ETA you set a few weeks ago when you left the port of origin, and giving a much more accurate revised estimate based on years of accumulated historical data. That lets you recalculate the ETA and improve downstream planning — customs clearance, delivery to the distribution warehouse, or in project-based cases, making sure a piece of medical equipment reaches a hospital on a strict timeline it's depending on. Those are examples where you get real quantification behind your AI output, which builds genuine comfort and trust.

Where there's more latitude or ambiguity in a decision, building that same level of trust is harder. Predictive maintenance in a factory is another good example: if you're collecting data on when an asset was installed, when it was last maintained, and what issues came up during that maintenance, you can genuinely predict when preventative maintenance is needed before something breaks down. That's also quantifiable — and your confidence in it grows the more you see preventative maintenance actually happening exactly when it's needed. So there's a real distinction between AI solutions you can trust immediately versus ones that need to earn some rapport over time.

Sunil: You have to earn that trust — just like with a person.

Edwin: Exactly right.

The real-world deployment problem: data quality and digital twins

Sajjad: This connects to something I find really compelling about scenario planning — if you use technology efficiently, Monte Carlo methods and simulation-based modeling let you run through an enormous number of scenarios and estimate risk reasonably accurately, then use your supply chain visibility to see how that risk cascades end-to-end. A lot of people are trying to build exactly this. My concern is that it's one thing to design these systems in a lab, but the real world looks very different — every company has its own proprietary supply chain, its own way of managing vendors and materials. That requires a lot of "last mile" integration, customization, and adaptation to actually make these systems work — the AI industry sometimes calls this the need for deployed engineers. Have you seen that play out?

Sam Pooni: And related to that, in the same context — how do digital twins factor in as a kind of dynamic projection layer? Supply chain itself increasingly looks like a graph problem, with knowledge graphs involved. How does ERP data feed the graph, the graph feed a live digital twin of the network, and the twin feed agentic simulation before anything actually executes?

Edwin: Honestly, that's going to be a real challenge for a lot of companies, and it's an area where, collectively, we still need to mature. I can guarantee that data quality is a persistent issue at every large corporation I've worked in or been close to — making sure data is uniform, correctly modeled, and complete enough and accurate enough is a constant struggle. A lot of our systems need that baseline data quality to give us the right kind of response in the first place.

That actually opens up an additional opportunity for agentic AI: helping corporations clean up and improve existing data quality, and making sure new data entries are captured correctly going forward so we can rely on agentic AI more confidently over time. It's a two-way street — you want to clean up what you already have, but you also need to make sure you're not injecting new incorrect data into a system that's already running. Everyone knows "garbage in, garbage out" — and I think agentic AI can help both with cleaning up past data quality issues and with driving outcomes we can actually have confidence in, based on how accurate the underlying inputs really are.

Large corporations struggle with this constantly — mergers and acquisitions force you to merge different business processes and data models, and consistency in applying the same logic across those models is hard to maintain. It also comes down to how well-structured your data model was from the start. Take country of origin as an example: it's a data element every component needs, and it's critical for international trade — but it's often not something you set up carefully when you first create a product master, because it depends on your supplier and how they determine origin. You might even have multiple valid suppliers, and therefore multiple countries of origin, for the same product — is your data model actually built to handle that? When you're talking about duties and tariffs, knowing precisely whether a component's country of origin is China, Europe, or Israel drives a materially different tariff outcome. Getting that accurate — and using AI to both clean up historical data and ensure the right data gets captured going forward — is critical.

Physical AI vs. knowledge AI

Sunil: This connects to something we discussed a few weeks ago — physical AI versus what a guest of ours, who runs an edge AI chip company, called "knowledge AI." That conversation came up right around the Cerebras announcement here in the Valley, which everyone was talking about for a few days. He introduced a distinction I hadn't heard before: physical AI versus knowledge AI. Listening to you talk about preventative maintenance — that's very much physical AI, sitting right at the machine. Then there's knowledge AI back at headquarters, synthesizing all of that information. It seems like we're going to see a lot more interaction between physical AI and knowledge AI going forward.

Edwin: I think that framing is useful — physical AI is probably easier for people to feel comfortable trusting, given how directly its outputs map to observable outcomes. The knowledge AI side is where the harder question lives: how trustworthy is that layer for corporations to actually rely on for real decisions, given all the assumptions baked into it? That's exactly why there needs to be real transparency in the assumptions underlying an agentic AI or knowledge AI response — so people can build genuine confidence in the output and rely on it for future decisions.

Sajjad: In existing environments, you typically have an entire data organization curating data, cleansing it, running data engineering tasks, and keeping pipelines running. Even then, you get failures, exceptions, and quality issues — but at least there's a person to talk to who knows how to resolve them. Now, at machine speed, with agents doing this work much faster, errors also propagate much faster. That makes the intensity of quality focus at each step of the pipeline that much more critical.

Edwin: Right — I think the transparency concern is less about the agentic layer itself and more about needing genuine trust and confidence in how the results were actually determined. Without that level of confidence, you won't get the adoption you're hoping for.

Sam: It's interesting, too — an agent can only act coherently if it can see coherently, and no single system of record gives it that complete view on its own. Something like a knowledge graph can provide that grounded, verifiable basis. Which raises another question: once you're running simulations and "what-if" scenarios, how does governance, regulation, and rule-setting actually fit in? With a managed system, a human can step in and say "this isn't right." But with fully agentic systems, do you need to explicitly define operating bounds up front? How do you approach that?

Sajjad: Especially around reporting — I'd imagine there's significant compliance reporting involved, which means you need much more detailed records of exactly how something was produced and how each step stayed in compliance with regulation.

Guardrails, regulatory volatility, and tariff uncertainty

Edwin: There are really two areas here. One is what I'd call guardrails. With scenario modeling, you can drive optimization through Monte Carlo methods, evaluate confidence levels, and determine the optimal path from the output — but today's environment carries so much uncertainty that it's genuinely hard to be precise about which variables to even put into the model. You want your AI to run through thousands, maybe millions, of scenarios and surface the most likely outcome or the best opportunity — but to do that effectively, you need real guardrails: constraints that let the model iterate freely within defined boundaries rather than trying to "boil the ocean." That guardrail concept is coming up more and more as a way to stay focused on outcomes that actually help you make a better decision.

Regulation is a related but distinct challenge — what we've seen is a huge amount of fluctuation there. On duty and tariffs specifically: we started with IEEPA-based tariffs, and a few months ago those were ruled by a court as not the legally correct mechanism for imposing duties. We shifted to a different tariff construct, and those percentages have kept fluctuating over the past eighteen months depending on the state of geopolitical relations country by country. So a lot of scenario planning comes down to a genuinely open question: what's the most likely outcome of discussions that are still actively unfolding? Guardrails help set reasonable boundaries, but there's real unpredictability that everyone in this space acknowledges — it's genuinely hard to know whether to model a 10%, 15%, or 145% tariff on a given category.

With the AI models we've been building, we can get to a likely range of outcomes — never a single deterministic number, always a spread — but the real benefit is speed: we can rerun the model quickly and produce a revised output within a couple of hours of any regulatory change.

Wrap-up

Sunil: We're just about out of time — Edwin, thank you so much. This was genuinely informative and educational; I learned a lot, and I really enjoyed the conversation. I'm sure our listeners will too. If you can stick around for just a second after we stop recording, we'd appreciate it.

Edwin: Not a problem.

Sajjad: Thank you.

Sunil: Thank you.

Edwin: Thanks for having me.