Clean Data Beats Company Size for AI Readiness

Lightly edited for clarity. Solo episode: Ysi Gonzalez, Vantelira.

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

Two guests, two sides of the same market

Over the last couple of weeks, you've met two of my guests. One of them built a brand. She packs every single order in a fresh bag per order, and she once fixed a national melted-chocolate crisis in seven days flat. The other sells the platform — a supply chain AI that spots the disruption before it even hits your manufacturing line. One is the brand and one is the tool.

I've spent 20 years of my career as the person standing between those two. And today I'm going to do that job here on the show. This is Ysi, and welcome to The Forward Deployed Operator.

If you haven't heard the last two episodes, pause right here. Go listen to Mariela first, then Heather, and then come back. This episode will assume that you have heard them both.

Catch-up: Mariela Katz, Nutterie

For everyone else, let me do a quick catch-up. Mariela, co-founder of Nutterie — direct-to-consumer nuts, seeds and dried fruit. She imports direct, roasts, coats and mixes in-house, and packs every bag fresh per order. Over 250 SKUs, selling one kilo at a time. Her stories — oh my god. The chocolate that melted in May. The SMS list that she had to split in two so her packing line survives. And a founder who learned every job by doing it herself.

Catch-up: Heather Sye, Centrum AI

Then Heather, VP of go-to-market at Centrum AI — risk-informed supply chain intelligence. Her frames: the tier 2 supplier in Chile whose problem you never heard about until it actually hits your manufacturing line. 17 instances of SAP spelling one supplier 17 different ways. And the real reasons 87% of AI projects do fail. We talked about governance, and master data, and expectations.

What happens when you put an enterprise supply chain platform inside a small brand

So today's question is the one nobody has really asked on either episode. What happens when you put Heather's platform inside a brand like Mariela's? Let's run it.

Would the platform have caught the melted chocolate?

Would the platform have caught the melted chocolate, for example? And this is hypothetical — assuming the brand has the volume and wants to get a platform like this to prevent this type of crisis.

Let me reframe the crisis in the language of the tool. Heather's whole example is the tier 2 supplier: the disruption you can't see coming. But Mariela's disruption wasn't really a canal, or a coup, or a geopolitical event. Her tier 2 supplier was the calendar.

She launched the chocolate category in October. Canadian summers are brutal. If you didn't know that — everybody only knows about the cold winters here in Canada. We have the two extremes. My sister says it's zero or nothing here in Canada: brutal winters, brutal summers. So the Canadian summer was already on Mariela's schedule.

Seasonality is the single most forecastable disruption that exists. And it still got to her because — and she was very honest about this on her episode — it didn't even cross her mind. She was really busy doing everything that you do when you're running a brand.

What a risk lens would have flagged, in January

Now, run this through the risk lens. A risk-informed view of her catalogue: every SKU by temperature sensitivity, crossed with shipping mode, crossed with the season. And then it flags chocolate, parcel, June — in January — as a planning line item, not a May fire drill.

And that's Heather's entire thesis. You cannot predict the disaster, but you can absolutely pre-model your exposure to it. And even in the moment that you are about to get hit — because it could have been that, already starting the summer, they would have found out not necessarily in January, but in April or May, and they can react to it faster, before it even impacts the customer.

The tool buys you the warning; it does not buy you the response

This is where I need you to trust me, okay? The tool buys you the warning, but it does not buy you the response.

The seven-day fix — calling frozen food shippers. Walking away from the box with the one-truckload minimum. Testing the solution by leaving packed boxes in the sun outside the warehouse for a full day. I mean, that only comes from a founder who had personally done every job in their company. There's no dashboard on this earth that does that for you.

So write this down. A platform would have moved Mariela's crisis from May to potentially January, or a couple of weeks before. It would not have replaced Mariela. The warning is software; the response is the operator.

Is a three-year-old DTC brand even ready for enterprise supply chain AI?

Is a three-year-old direct-to-consumer brand even ready for enterprise-grade supply chain AI?

Heather gave us the entry exam from her own mouth. The data has to be harmonized, and most companies fail before they even start. And this is any size of organization. 17 SAPs, one supplier, 17 spellings — every analysis on top of that data quietly being wrong.

The plot twist I only realized listening back to both tapes: Nutterie accidentally passes the exam. Think about it. One platform of record. Clean SKUs. She knows her inventory is tracked at the kilo level. The docket per order, with one person already owning it. There are not 17 spellings of anything at Nutterie, because one founder built one system and did the job herself.

Why AI readiness has nothing to do with company size

Why does that matter for you listening right now? Because readiness for AI has nothing to do with company size. A three-year-old brand with clean, named data is more AI-ready than a global manufacturer with 17 ERPs.

The 87% that fail — they don't fail on model quality. They don't fail on the tool. They fail the exam that Nutterie has passed already.

This is the same cheat sheet I give every scaling brand that reaches out to me asking questions. How do they get ready to scale? Get your SKUs and vendor IDs ready first, before anything else.

Mariela is living proof of what that discipline buys you. You have options. You will have options, because you can play and mingle and predict with your real data. She could adopt a tool like Heather's next quarter if she wanted to. Her competitors — with a 3PL, a distributor, and three spreadsheets that don't agree with each other — could not.

Three use cases where the tool would actually pay for itself

So where would the tool actually pay for itself at a brand like that? Let me do the scoping exercise out loud. Three use cases, okay? Each one priced against a pain that Mariela already told you about herself.

Use case one: origin risk on direct imports

She owns the whole supply chain. They are able to import direct, and then they get up until the consumer's pantry. She is her own tier one supplier. Shorter chain, better visibility — and total concentration. However, a harvest failure or a port event at origin doesn't hit her eventually, the way it hits her competitors. It hits 100% of her supply chain immediately. So the first use case scenario where I would use a tool like Heather's is on her tier one supplier, which happens to be herself.

On Heather's episode, I mentioned a brand I know that bought a full trailer of chocolate to cover the entire year. That's origin risk management by brute force. They did what they could do. The platform version models exposure per origin, per ingredient, with the buy-forward scenarios costed before the panic, not during it.

Use case two: demand spikes as scheduled disruptions

Mariela already splits her SMS sends across two days to protect the packaging line. Do you see what that is? She's doing demand scenario modeling on pure instinct. I mean, I love this woman.

The platform version of that: if we send to 225,000 subscribers at 10:00 a.m., what does tomorrow's line look like? What will the impact of that be on our manufacturing and packaging facility? Answer it before the send button, with the split calculated instead of felt. It's a scenario engine pointed at the demand side instead of at the supply side.

Use case three: negotiation data on demand

Mariela admitted that there were shipping expenses above what they should have been, potentially, for the first full year — to build the volume portfolio that finally made carriers start taking her seriously. And she changed carriers a couple of times in between. Harmonized order and freight data produces that negotiation position in one export, any day, for any carrier conversation.

Now notice what I just did. I did not pitch a single feature. I took three stories the founder already told and I priced the tool against her pain. That is the difference between a demo and a diagnosis.

Would I tell her to buy it? Not this quarter

Would I tell Mariela to go and buy Centrum AI? First, let me steal something from Heather herself. She said that the best thing a person in her seat can do is be prepared to say no when the fit isn't there — to decline a customer. So let me be the operator who says the equivalent out loud for Mariela: not necessarily this quarter. And here's why.

First, the risk register comes from Process Archaeology, not from a sales call

She would need a process investigation — archaeology. Take the shovel, unlock everything that is being done with her operations. Her own three stories are part of that risk register: the melted chocolate, the SMS spikes, the freight. We name them, we rank them, we put an owner on each one.

Second, validate the governance before there's a connection

Who acts on the alerts? Who has the authority to buy forward? What is the tool allowed to touch? What are the actual SKUs versus vendors? Is everything aligned on the platforms that are used? Heather told you exactly what happens to companies that set governance up after something breaks. Don't be that case study.

Third, and only then, connect the platform

And then third, and only after this has been completed, you connect the platform — pointed at the two or three risks that are actually her own, not the 40 in the brochure or in the offer.

AI is not a magic wand

And set expectations. In Heather's own words: AI is not a magic wand. I really love this expression. It's not a magic wand. It will not fix a failing company. It will only fix what you point it at — what is already prepared to be fixed. Tools amplify whatever process they land on. So land them on a good one.

The empty chair both guests described without meaning to

What did these two guests prove without even meaning to? I want to put two quotes side by side. So listen carefully. A founder describing an empty chair from the demand side. A vendor describing the same empty chair from the supply side. They've never met. Neither one of them was prompted. And both of them describe the seat that sits between the business and the tool. Vacant, untitled, and very decisive.

That is the seat that has a name now. It's a Forward Deployed Operator. Embedded enough to know the business the way Mariela knows her kilos. Fluent enough in the tools to know what Heather's platform can and cannot do. And disciplined enough to say not yet.

Every system you fixed became someone else's win. The next one has your name on it. An eight-week live cohort. You solve a real problem inside a real company, ship a working AI system, and leave with the case study, the build and the title: Forward Deployed Operator. Join the waitlist at fdoschool.com. That is fdoschool.com.

Why 13% of AI projects succeed

This is becoming even more real every time. Both sides of the market discovering it independently — and that just happened on this show two episodes in a row.

This brand just proves that operational knowledge is earned by doing the jobs, and that clean, named data is a side effect of doing them well. The platform proves that tools only land where that data and that process already exist. The vendor herself will tell you the tool fixes nothing that was broken before.

And between the two stands a seat that both of my guests describe without knowing it: the person who translates the business into the tool, and the tool into the business. That is the reason 13% of AI projects succeed. Make sure somebody in your company is sitting in it.

Fireside: what exactly is a connector?

Welcome to fireside time. These are real questions I've gotten from clients I've been working with these past couple of weeks. I will ask them out loud so you will get the details behind it.

What exactly is a connector?

I was giving an AI introduction training to a group of people from a client, and it was their very first time being exposed to the AI tools that are already out there. And that was the first question in the room when I started showing them how to connect their email, Google Drive where they store the information, the calendar, all these things.

My honest answer was: a connector is like marrying the AI agent or tool to an app, so it can read and write your email, your Google Drive, your calendar, for example.

However, the real answer is what I did next with them. I did not give them a lecture. Within 10 minutes of creating the account, I had connected their inbox and asked one question to the agent: can you check the emails and tell me what is the priority today? Two orders that were needed to ship surfaced instantly — orders that they would probably have found hours later, or at the end of the day.

And that's my whole adoption playbook. Not a slideshow, not theory. Just sitting down with the people that are going to use it and showing them inside their workflows, on their data, and finding something real. That aha moment. 10 minutes in, they weren't learning AI. They were triaging their morning with it.

I believe adoption is not a training problem. It's a first-win problem. Prove the usage exists before anyone can get committed to it. Sell the outcome, not just the license.

Fireside: tokens feel like a black box — how do we not waste them?

Question number two: the tokens feel like a black box. How do we not waste them?

This is very funny, because lately I have found myself talking a lot about how to measure the amount of tokens versus the return on investment. And honestly, it depends on the account. Enterprise customers, or the ones that have team accounts and the higher tiers, will have that visibility into the tokens. Simpler plans might not.

If you don't see exactly where your tokens are going because you're not on that enterprise tier, you can't really fully predict consumption. You will see certain patterns on the days that the sessions stop earlier or later. But that is the vendor's business model — really not your problem. You just have to adapt and figure out yourself how to optimize it in a certain way.

So instead of trying to predict it, I try to engineer around it, so to speak. I taught them the two sentences I normally end my prompts or instructions with. If I ask for an analysis, or whatever approach I'm using, or I'm pinging an agent for something, there should be a layer of asking clarifying questions before doing it.

Because I remember when I started working with AI a couple of years ago, I would ask the questions and then not necessarily define all of it. And because of the way that the models have been trained, they have certain patterns that they follow, and they believe — or they assume — that things are happening in a certain way. So asking it to ask you back clarifying questions lets them stop and reflect and say, "Oh, I'm making an assumption here. These are the questions that I should be asking."

And sometimes I get a full page of questions — at a time that would probably have become an iteration and a waste of tokens, going back and forth on "this was not properly done, it should have been done the other way." So the first one turns every big task into a reviewable plan before anything gets executed. The second one forces the assumption out in the open.

Rework burns money — budget tokens the same way

In the years of my career in factories, one thing it has taught me is that the cost is never the task. It's having to redo it. All the wasted time that we get into, having to fix the problem that could have been avoided from the beginning. That's one of the first things that they show us in university, and then on the factory floors. Rework burns money. Budget tokens now exactly the way rework burns labor hours.

And AI doesn't hallucinate because it decides, "oh, I want to hallucinate." It assumes because you didn't brief it enough. Same as any new hire. I onboard AI the same way I would onboard employees, and the same way I've done it in the past: context first, checkpoints before execution. Prompt discipline is just process discipline wearing a new shirt.

LiliBloom makes personalized wooden and acrylic keepsakes out of Montreal. It is my choice for any keepsakes that I want to give to my friends, my family, my customers. First birthday photo boards, cake toppers, first day of school signs, anything in between and anything you need. Every piece is thoughtfully designed and crafted with love, care, and an eye for detail, making your celebrations and your gifts even more special. Find them on Etsy as LiliBloomShop. That's L-I-L-I-B-L-O-O-M-S-H-O-P. LiliBloomShop.

Fireside: why are we still doing this manually?

And the last question: why are we still doing this manually? Third question, and it came from the same person who told me — and I quote — "I don't like doing manual anymore," which honestly is music to my ears.

My answer: this process is still manual because we're in a transition. Once we get out of this transition, it will get automated this way and that way. So I will only automate the process after it has been stabilized. While we are in a transition, there is no need to be applying an automated workflow or an AI agent, or training whatever tool you have to train.

On that same call, I did the mirror image. I showed them an order that got stuck because it was a one-time exception, because of the transition process we were in. And I had to go and work around it, and the team learned. They learned to do it, and they learned that automating something like this would just be a waste of time. This is not a process that you're going to keep. This is not something that is going to stay as is, because it's a transition period.

Nobody remembers that it was supposed to be temporary if we automate it that way. The idea is to identify where it needs to be improved. Anyone can add an automation anyway, especially now in the AI era. The operator skill is placement: what to automate, what to leave manual on purpose, and what to actively refuse to repeat.

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.