Why AI Projects Fail Inside Supply Chain Teams

Lightly edited for clarity. Speakers: Ysi Gonzalez (Vantelira) and Heather Sye (Centrum AI).

This episode is powered by Vantelira. We go inside a scaling operation, fix the process, and build the system that keeps it running.

Introduction

Ysi: Hello and welcome to The Forward Deployed Operator. I'm Ysi, and I'm here with Heather Sye. She is the VP of Go-to-Market at Centrum AI. Centrum AI is a deterministic artificial intelligence platform designed to optimize and secure global supply chains.

Let me tell you, I've been the whole of my career on the other side of these tools, needing them to be available for us. So I have many, many questions about how it works and how it helps the supply chain. Heather, welcome.

Heather: Thank you, Ysi. Wonderful to be here, and I'm excited. Let's jump into it.

From banking and finance into AI

Ysi: Amazing. So do you want to tell us who you are, where you come from, how you ended up working at an AI platform that's supply chain driven? What is this?

Heather: Yeah, for sure. So I'm Heather, I'm the Vice President of Go-to-Market, which essentially is a tech term for basically anything customer-facing — from marketing to business development, sales, customer success, anything along the customer journey. I'm leading that, so I'm very much customer-facing.

Previous to Centrum AI — actually the entire executive team — we've worked together now for the better part of a decade in similar startups, in artificial intelligence, in automation. So we really understand technology and implementation, and what works, what doesn't, and what creates success.

And previous to tech I was actually in banking and finance. So a little bit of a pivot, but still customer-facing my entire career.

Ysi: It's a good journey. And I love when you said implementation, because that's when things can get a little bit messy. If you have the experience already, that's the way to go.

What a risk-informed supply chain platform actually does

Ysi: Do you want to share a little bit about what Centrum does, how it works, how it can help others? And you mentioned something about a Chile example in our previous conversation, if you remember — you can share that with the audience.

Heather: Sure. So Centrum AI is a risk-informed supply chain intelligence platform. What does that mean? We look at how we manage our supply chain from a risk-informed lens: looking at the impact of disruptions, looking at how we better protect our supply chains, becoming more strategic in how we manage our inventory, how we look at and interact with our suppliers — all from a risk lens.

The Chile example: the disruption you never see

Heather: What this means in practice is, let's say in Chile you have a tier-2 supplier located in southern Chile, and there's a geopolitical event that prevents folks from getting to the manufacturing plant. You may not see this in the news. You probably wouldn't. You wouldn't hear about it in an email. But over the course of time, the impact is that you don't get your supplies on time. There could be a delay, and you'll never know about it until it hits your line.

So being a little bit more prepared, and knowing in advance the impact of these disruptions on your supply chain, can mean that you're coming at your planning from a more strategic position versus a defensive position.

Do companies even know who their tier-2 suppliers are?

Ysi: You mentioned tier-2 supplier — and for younger brands that might not necessarily be there yet: when you go to the enterprise level, you define your supply chain from tier one, tier two, tier three, and that's how you define how important they are to the actual production and manufacturing lines.

Given the experience you've had implementing this, are these companies already there, saying yes, we know who our tier-one suppliers are? Or is there a little bit of groundwork that you or the team has to do behind Centrum before even thinking of plugging the app in?

Heather: Yeah. I'm constantly surprised. There are a lot of unknowns in our supply chains, and there are various levels of visibility across the board. Some folks come to me and they're like, "Honestly, Heather, I have no idea even the costs that I'm incurring — never mind what my tier-2 suppliers are doing. That's way out of the realm of my visibility." But then you have other companies that either have good control of it, or are trying to improve their visibility levels. There are all different gradients across the board.

Connecting the systems you already have

Heather: But in order for us to understand better, we essentially connect with those disparate systems that you already have — unlocking layers of data that your analysts may not have access to, or don't have the time to process, or that's locked in different areas they don't have a connection to, like your WMS or your CRM.

All of this information, if you collate it together and harmonize it and see what's actually in a warehouse versus what's being said in your ERP, can make a total difference for understanding the more at-risk SKUs, or the more at-risk suppliers, and what their historical behavior has been like.

Would this have changed COVID?

Ysi: Do you think — and this is totally unprompted, I'm sorry to throw you this question, I hope somebody else has asked you this before — do you think it would have changed something in the supply chain world if Centrum AI had existed six years ago, when COVID was rumbling? If I'm not mistaken it was November, December, the news started to pop up here and there about China and the virus. If we were to roll the tape back and say Centrum AI was there and connected to your ERP and all of your systems, do you think that would have changed how supply chains were impacted?

Heather: I mean, I'd like to just say yes, we would have predicted a global pandemic. But I don't think anybody could have predicted the chaos that ensued six years ago.

Disruption used to be yearly. Now it's weekly

Heather: But it's all about your reaction to these disruptions. Before six years ago, there would be a catastrophic disruption a few times a year that would stop the lines, or chaos would ensue globally. But now it's weekly. Somebody sends out a tweet and all of a sudden there's an entire disruption in our networks. A canal is blocked here, or there's an earthquake there.

So it's not necessarily about knowing what type of disaster is going to happen, but being able to react, and have scenarios modeled out quickly, and have a team deployed quickly to respond to these challenges.

Ysi: I believe it would have changed how we all reacted to it. There was a lot of disruption because people didn't believe it was going to really have an impact. And I believe there was a lot of missing information — companies not aware of their tier one, tier two, tier three, and where they were located, and what their lead time was, and all these things. Definitely, that's a lot of what happened.

Running tariff scenarios by hand

Ysi: And this is why I said at the beginning, in the intro, that I wish I'd had a tool like this before, because I've had to run these scenarios manually. There were the tariffs that recently happened, and they were announced with time, right? Considerable time. But when you're running things manually and running your Excel scenarios to figure out which markets are going to get the impact, it delays decision and reaction time, which ends up losing money for the company or risking customer satisfaction.

Heather: Yeah. By the time you've replanned everything, we've already moved on to the next disruption.

Ysi: Yes. Yes.

Heather: Whereas now I should be in a meeting with the relevant stakeholders and I should be able to say, okay, let me run a scenario plan on the spot and have an answer for you right now.

Ysi: You know how many times I was asked to do that in the middle of meetings, and I'd be there with my 36,000-SKU spreadsheet. I mean, yes, I can — but I'll get back to you in a couple of days.

I work a lot with trying to not have to do things again a hundred times. So if there was a disruption and somebody asked me for a model, I'd try to set up my reporting and my modeling in a way that I'd have to feed very little information the next time I had to run it. But in these cases, as you mentioned, we're getting a different disruption every week, every month. You can't predict manually what's going to happen.

Over 350 connectors

Ysi: So again, I'm very excited about this tool, and I want to know even more details. How many connectors does it have already?

Heather: We're constantly adding more, so it's over 350 now. Your standard SAPs of the world, your WMS, but then there are some more niche ERPs out there that we've connected with directly for customers. There's the ability to write back to Slack and any sort of communication device.

Is this only for enterprises?

Ysi: And is that because right now Centrum AI is more enterprise-level focused? Is there anything for, for example, smaller brands dealing with chocolate disruptions right now? I have very close experience of having to buy a whole trailer of chocolate for the supply of the year, because we don't know what's going to happen. Is there anything for those smaller businesses that maybe cannot afford an enterprise solution — something Centrum already offers, or something on the lookout?

Heather: No, actually, my favorite customers to work with are that mid-market manufacturing space, where they do have to be agile. They don't have 18 months to work with a Deloitte to plan through a various course of POCs.

I like companies that just have a few people to get through, that can move fast. They have a real need and a real personal desire to get these problems solved. Those are my favorites — that midsize regional manufacturer.

Who cleans the data, you or the customer?

Ysi: In your experience going to market with Centrum AI, have you found there's a lot of harmonizing of the data that has to be done prior? And is that something your team does, or the customer has to figure it out?

Heather: We actually do that. Typically the standard process would be, you have to clean your room before an AI company can come in, or else it's bad data in, bad data out. We actually wanted to solve that problem first.

So whatever state your data is in — you have somebody fat-fingering the numbers, there's an extra zero, or there's a period versus a comma, company names spelled incorrectly all the time — our first step is actually harmonizing and making sense of and cleansing that data automatically. We do that with artificial intelligence. It takes less than a week to clean your master data, and then that's where the fun starts.

A non-healthy obsession with master data

Ysi: You mentioned master data and I get very excited.

Heather: I too get excited about master data.

Ysi: It's a non-healthy obsession with master data. It follows me, and I guess you have the same experience being in supply chain. You're talking about companies having different names — I've seen it at any level. It doesn't matter if it's a two-person company or a ten-thousand-person company, you will find these duplicates. It's SKUs, it's the data. At your level it's with the supplier, but it also happens with product. It's just one of my favorite topics.

Seventeen instances of SAP, one supplier

Heather: Well, it kind of happened as a product on its own, because this is such a challenge in supply chain — particularly if you're a company that goes through a lot of acquisitions. You might have 17 different instances of SAP, and each of them spells a company name differently. So I have 17 different instances of a company, or a supplier, that behaves the exact same way.

Well, our AI can detect the chain of events that happens to a supplier and isolate it down to that one, and say, "Hey, these 17 companies, I think they're the same company." And then harmonize it, and there you go.

Get SKUs, get vendor IDs

Ysi: I love it. I get a lot of brands reaching out to get a quick cheat sheet of how to better set themselves up for success when they're starting to scale. And the one thing I bring to the table is: get SKUs, get vendor IDs. If you have that, you will be able to do anything.

And it's interesting to hear you say you can do that analysis very quickly with AI and fix it, because I believe that's what is so life-changing about having these machines so powerful to do this analysis for us. However, you still have to look and confirm — because of those 17 companies, there might be one that isn't the same. And you are the kind of person with the authority to tell the AI: that one no, the other ones yes.

Where adoption actually breaks

Ysi: In that sense, can you share an experience of where you see this adoption breaking? Let's say companies do get Centrum AI, and then get the data cleaned up, and then the planning team, or the procurement team, is not using the app. Do you normally follow that as part of your business model, and have you found patterns around why this happens?

Heather: Yeah, absolutely. There are a whole lot of reasons and theories out there about why AI fails. It's this super exciting hook for news articles: 87% of AI projects fail, and it doesn't work, and everything's hallucinating.

Governance: who owns what, who can access what

Heather: But yeah, there are some commonalities. First off is governance — data and AI governance. Having an understanding of the boundaries and the rules and how you operationalize AI within an organization is very important. Who's in charge of what, who has access to what, and what you do with that data, and understanding the tool or the vendor that you're working with and what they do with that data. That all comes under the umbrella of AI governance.

A lot of organizations don't quite have that yet, or they sort of wait for something to break before they put in a governance team. Obviously I would recommend thinking about that before you go into a big implementation that could make changes on your behalf.

Expectations: AI is not a magic wand

Heather: But then also expectations. AI is not a magic wand. It's not a panacea. It's not going to solve a failing company that has bad systems in place and make you a successful company just by implementing AI. That's not how things work, unfortunately.

What it does do is help the folks who work in your organization work more effectively. It takes a lot of the grunt work away, so you can focus on more strategic initiatives. It's not just going to solve a bad supplier and make everything go away. So having those expectations in line with your vendor — what the tool can do and what it can't do — I think is important.

When a customer isn't a fit, say no

Ysi: Have you found a company so far where you felt, oh, they bought Centrum AI too early for the stage they're at?

Heather: I mean, I hope not. We do work with organizations large and small. I have even a very, very small company in Germany that I'm working with that's not super advanced in their transformation journey. And that's fine.

Where we do — and this comes down to me as a person in sales — is that I have to make sure that it's a good fit for a customer, and be prepared to say no and decline a new customer if we're not going to be a great fit for them. I think that's the biggest gap: if I'm working with automotive manufacturers, I'm not moving to another industry I'm not very familiar with. That's the biggest challenge.

Automotive, aerospace, and expensive inventory

Ysi: Do you have specific industries that you're focusing on right now?

Heather: Right now we're quite successful in automotive, in aerospace. The commonality would be a lot of SKUs, a lot of inventory being held in warehouses, expensive inventory. But they're also being hit by a lot of disruptions. Unfortunately for the auto sector, tariffs are tough, raw materials are a constant challenge. So we've done a lot of really cool things with automotive.

Ysi: I asked that question with the intention of hearing you say our space. And I want to know, what about the electrical components companies out there? Asking for a friend.

Heather: Asking for a friend. Oh, electrical components. I don't have a customer in electrical components, but if they have a lot of SKUs, a lot of inventory on hand, some suppliers they'd like more insights into, I'd love to talk to them.

What has to be true before you trust the output

Ysi: So what would you say is true about a company's data and their processes, for people to trust the recommendations that the tool is making? And not only Centrum AI — any AI-driven tool.

Heather: Ooh, great question. A few checks you can have — and this is everything from using ChatGPT on your home computer and running your personal financials, all the way down to more robust, company-dependent AI use cases — is having the ability to track and trace any sort of calculations that happen.

We've all heard the cases, like: what's 2 plus 2, and ChatGPT says 18. With more robust calculations, you need something to audit that. Where did you find that information? Walk me through the calculations. In Centrum you can just hover over any sort of number and it'll go through the calculation for you, so you don't get that embarrassing moment in front of your CFO where you gave them a wrong number.

What's behind the model, and what the agent can do

Heather: But also, when you're using any sort of AI: what's behind the model? Where is it accessing its data? If we're using agents, what are the controls around what the agent does and what it doesn't do? Constantly monitoring where it looks. Using AI is not something where I turn it on and I go and watch Real Housewives. It's a very active part of your operational model.

A 45-minute lesson in data governance

Ysi: Have you seen the memes out there? A couple of months ago there was a whole thread of people posting videos of themselves doing the dishes while Claude was working, and you could see the code on the screen next to the sink. Like, yes, the AI is working while I'm doing this — but I'm constantly just keeping an eye on it. That's the reality behind it.

I have a little bit of a story on that, because just yesterday I spent about 45 minutes. It's embarrassing to say, because after the 45 minutes I was like, what was I thinking? I know what I need to do from the beginning in terms of knowing the calculation behind it, and knowing the source of the data and everything. But it was 8 p.m., relaxed, just having a quick check-in with my AI tool, and I'm like, can you check what the gap would be between these two files? Specifically I was looking at timesheets and vendors providing certain services.

And it went on and on and on, and it told me it had found a gap — that some vendors were not providing the services they were scheduled for. And I was like, oh my god, it was shocking, the analysis it gave me. And then you start asking and poking and drilling down. And at the end I was like, oh my god. I never told it that the list of vendors had certain categories that it needed to consider — all of them — because the names of the vendors were different. Some of them had an M at the beginning and some of them didn't have it. It assumed that the ones with the M weren't standardized vendors, so it wasn't going to touch them. But we never agreed on that from the beginning.

And that goes back to the source of your data, your data governance, and having it all set up. Even to me it happens, and it can happen.

Heather: Yeah. I think everybody has a story where they ask for even a marketing material, or "write this email," and they ship it without double-checking. And I mean, I've done it. It's an eventuality. But you learn, and then you double-check what your AI says.

The buyer and the user are not the same person

Ysi: You said on our previous call — and I really appreciate that view from your side — that the people who get the AI and the people running the business are rarely the same person. I wanted to explore that a little bit. What do you mean by that?

Heather: Yeah. So you have various key players in any sort of implementation or AI purchase, and they're not all the same, and they don't all have the same interests. So when I engage with customers, I want to understand their interests, their motivations, their behaviors, and what's important to them. Because at the end of the day, everyone's a human, and they have KPIs that they want to achieve. They have certain tolerance and certain history with AI being successful or not.

Whose head is on the line

Heather: So oftentimes the end user — maybe it's an analyst, maybe it's a manager, or any other type of profile — and then the buyer is going to be somebody completely different, and their head is on the line if this doesn't work, if the analysts aren't using the tool that they paid hundreds of thousands of dollars for. Their head's on the line. So that's their interest in making sure an implementation is successful, making sure the outcomes are correct.

Whereas the analysts might just want to run the numbers quickly, and they may or may not understand AI or how it can benefit them, and then they want to go home and go back to their family. So you really have to find the incentives for each individual in that chain.

The connector who makes the implementation work

Ysi: I was asked that question recently, and it's exactly that. It's finding that interest. What's their end game? What's their purpose, from each side? Would you say you have a secret sauce to close that gap, or do some things work here and some things work over there?

Heather: I'm always learning, and always surprised. But the most successful and the smoothest one that I went through recently — they actually had a point of contact within their organization who played a project manager role. So they're not the end user, they're not the buyer. What they are is the connector between all of the stakeholders.

They know who to go through throughout the organization. And so they were my point of contact for making sure everything that I needed to get done was successful. They almost translated everything from their organization into how we make our implementation a success. And that was hands down the easiest, smoothest way for both sides to be happy.

Ysi: And that is what I call a Forward Deployed Operator. You got it right there.

Heather: They're coming on board as our marketing person now. And you know, that wasn't even their title.

Ysi: Not yet. It's super interesting, because I found my way into being that translator many, many times — and teaching my interns and employees at previous roles. It's those people who can become the translator between what the capabilities of the software out there are and the tools that can implement it. They plug it in, and it is a seamless transition for both sides, as you just mentioned.

Heather: Hugely, hugely valuable.

What Heather does with AI herself

Ysi: It is very valuable. And in that sense, I want to close there, because it's just been amazing having you. We do have a tradition on the show, where the person before you — solutions oriented, let's say — left a question for the person in this seat. And the question they left for you is: tell me what you do with AI yourself. Not Centrum — yourself, Heather. Beyond prompting a chat. And describe it as much as you can.

Heather: Absolutely. I love this question. It's a great question.

So Centrum is what we call an AI-native company. So I literally use AI in everything I do. And not just in using Centrum — I use Claude as my go-to. I use it with my sales calls. I record, I transcribe, I analyze my performance. I do summaries, write-backs, follow-ups.

The agent she built as a sales coach

Heather: One really cool agent that I built actually understands sales methodology from a variety of different perspectives, and then understands my methodologies for sales, and then it analyzes it and gives me feedback as a sales coach. So every time I have a meeting, I download the meeting into my sales coach and it spits back sometimes harsh criticism of what I can do better the next time. And then it triggers a whole chain of different agents to perform on my behalf.

Ysi: I love it. I share the love for Claude. Actually, I've named mine Claudia.

Heather: Oh, I love that.

Ysi: I even prompt her like that — Claudia — and she's like, "Oh, hi, Ysi." It's just the little touch.

The question she leaves for the next guest

Ysi: And then you have to leave a question for the next guest, which I won't reveal who it is, but I will tell you that it's a person who sits between the people that work with AI and the implementation of it. It's kind of in the realm of forward deployed engineers. So there's going to be a matchup between us soon. What would be the question that you can leave for them?

Heather: You know, I would love to hear — and I don't think we talk about these things enough — I would love to hear about a failure. Something like an implementation that went sideways, went all wrong from the beginning. So I would love a story.

Ysi: Heather, thank you so much for that question. I am so excited to ask it to the person that's going to be in that seat, and you will understand why I get so excited about it.

Thank you so much for sharing with us today. It's been amazing to know what Centrum AI can do, and your understanding on data governance and master data — our shared love from now on.

Heather: Absolutely. Thank you. Thank you for having me, I've loved it. It was a great chat, and we'll chat soon.

Ysi: We'll see you next time on The Forward Deployed Operator.

That's it for this one. Follow the show so the next episode finds you, and the newsletter goes out every Tuesday — the link is below. See you next time on The Forward Deployed Operator.