Where to Start With AI When Nothing Is Documented

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

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

Episode 12: a solo episode that closes a block

Hello and welcome to The Forward Deployed Operator. I am Ysi. This is a solo episode, and it's actually a milestone one: we're celebrating our 12th episode, and it's just you and me today. It closes a block, so let me catch you up on what we have been doing. If you're new here, this is a great episode to start with.

There are two conversations before this one that you may want to go back and watch.

Recap: Antonella Gioni and a company growing faster than its systems

Two weeks ago, I had Antonella Gioni on the show. Twenty-five years in operations and sourcing, from Novartis and L'Oréal Canada, and now she is the Vice President of Commercialization and Innovation at our beloved Earth Rated.

Her mess was a company growing faster than its systems. Her team was spending the whole day going to get information: log into portals, copy the numbers, send the email, wait, chasing data. And she fixed the foundation first, before anybody in that building said the word AI out loud.

Recap: Marie-Claire Asseko and every department's own version of the truth

Then last week I had Marie-Claire Asseko, a consultant and product strategist who also coaches people through the change, which is the part most people skip.

Her mess is the one she walks into in other people's companies: every department with its own spreadsheet, its own definitions, its own version of what is actually true. And nobody notices it for years, because they're all using the same words with different definitions behind them.

The tool was never the hard part

The reason I wanted them back to back is that both of them told me the same thing in very different words: the tool was never the hard part. Antonella from inside the brand, Marie-Claire from outside, in other people's businesses. Different chairs, same sentence.

The actual problem sitting underneath both of those conversations is the one I hear every single week right now: the exec team wants AI agents by next quarter, and there is not one single accurately documented process in the building. That's what we're going to tackle today.

The peanut butter and jelly sandwich game

Before we get into it, let me tell you about a game I played with my oldest daughter. We write processes. That's the game. For her, she was coding, right?

I told her, "Write me the instructions to make you a peanut butter and jelly sandwich." And then I would do exactly what she wrote. Exactly what she wrote. We spent hours doing that game, no screens, no TV. It was actually pretty fun. Seriously, a couple of hours. It's my trick: activities that are fun and are actually teaching something.

I watched her write the instructions, and then I watched her face when she saw me following them to the exact T. She forgot to write some things that are, potentially, common sense. And her face when she realized: "Oh, shoot. I didn't tell you to take the bread out of the bag." She told me to spread the peanut butter, and I spread it on the outside of the bag. Of course. You never told me to open it.

We do that on purpose, for demonstration and education. I actually spread the peanut butter on the bag.

Hold on to that, because we're coming back to it. That game is the single most valuable thing you can do before you spend $1 on an AI tool.

The three parts of this episode

Today is going to be three parts.

First, I'm going to decode the vocabulary from those two conversations. If you haven't watched or listened to them yet, go do that, because a lot of words got used and I want to make sure they are easy to understand.

Second, the method I actually use with customers and clients. Five steps, and one of them is the sandwich.

Third, what you're going to go and build. Not "AI could transform your business." Actual things you can do today. So let's go.

Family one, the plumbing: what an ERP is

In the previous two episodes, we used several business-acumen words that get thrown around a lot, and I want to make sure we explain them together. I grouped them into three families, because they actually work together.

Family number one is the plumbing.

The ERP (I explained this one in the episode with Antonella) is the enterprise resource planning system. It is the base of everything that happens inside operations in a company: your purchase orders, your inventory, your invoices, your receipts, your vendor management, everything in one place. It's not just the software finance uses. It's the base. And if your ERP is wrong, everything downstream is going to be wrong too. And it's going to confidently get worse.

What an API is: the waiter in the restaurant

Then the API. API connections are what make it possible to have a hundred different systems giving you information, instead of you chasing it down from Shopify, from your warehouse management system, from wherever the information sits. You have access to it all in one place.

I found somebody on YouTube who explained it better than I could ever think of, so I'm going to borrow it. Think of it like a waiter in a restaurant. You give the waiter your order. That's your request. The waiter takes it back to the kitchen. That's the system. And then the waiter brings you the food. That is the response. That's an API. It's the waiter, basically.

Antonella gave the perfect example of why you should care. Her warehouse system and her ERP could not talk to each other, so her team's whole day was going to get the information. They built the connection, and the information started coming to them instead. Same people, same job. But now the day is deciding instead of chasing the data.

Stop chasing the data

That's a thing I say all the time: stop chasing the data. Stop trying to go gather it. You are not paying people to go get information. You're paying them to make decisions with it.

What a data lake is

Data lake: that's another word that got used. It's one place where all of the company's data lands in its raw form, so everybody is pulling from the same source. It's not a dashboard, by the way. A dashboard is what you build on top of it.

I was part of a huge transformation building a data lake for a very large organization years ago, so let me tell you: it is a struggle at the beginning. You find out three different systems have the same vendor under three different names or numbers, and now somebody has to merge that. Do the taxonomy, work out that this vendor is the same vendor as that one in the other system under another name.

But once you're past that mountain, it is beautiful. The amount of improvement you can make afterwards, and the speed of it, it takes off. It's totally worth it.

What data governance actually means

The other word we use a lot, and I've used it in many, many episodes (I talk about this all the time), is data governance. This one I'm very passionate about. It's not a secret.

Governance is who owns the definition of a word, who is allowed to change it, and how often and where it gets updated. That's it. That's the whole thing.

The data needs governance. It needs to be managed by somebody. There's normally a team, or a person or two, who are actually the owners of the data. That is how a data lake gets a translator, so to speak. So when finance produces a dashboard, the supply chain team produces another one and the marketing team produces another one. Whether they're talking sales or customers, it's the same number they're talking about.

Three teams, three definitions of "client"

Marie-Claire gave the example I have now seen in literally every company I've ever walked into: three teams, three definitions of the word client. For one team, a client is anybody who ever signed anything. For another, a client is one that's over $50,000, for example. And another team has its own definition entirely.

They're all sitting in the same meeting talking about how many clients we have. They use the same word, but because they have different definitions in the back end, they're giving different numbers. Then people start wasting time clarifying: "Why does your number look like this? Why does that number look like that?"

If there's data governance behind it, no time is wasted on those conversations. You're there to actually make decisions.

You cannot automate your way out of that, regrettably. You have to go through it. You have to define it. If you try to automate it, you'll just get the wrong answer faster. But once you do have governance in place, you're set. I'm telling you, you're set. Beautiful things can happen.

Family two, the AI words: the model is the engine

The second family is the AI words we used.

The model is the engine. There are different sizes at different costs, and everybody's instinct is to reach for the biggest one for everything. Just don't. And sometimes it's not even people reaching for it. Sometimes the biggest one comes as the default and we fall into it. I say we because I also became a victim of it recently: I just forgot to change the model for the task I was actually executing that day.

Tokens are the meter

Tokens are the meter. Everything the model reads and everything it writes costs you tokens, including the time it spends thinking.

And I just told you how I learned that one. I actually posted about it on my LinkedIn. It was mid-morning, not even noon, and I had already run out of tokens. I was like, what just happened? Normally I have a hundred different things running at the same time, and I don't run out. Well, that day I opened new tabs and new chats and new Cowork flows, and I forgot to set the model I actually needed for each specific task. And that's what happened.

Break

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.

LiliBloom makes personalized wooden and acrylic keepsakes out of Montreal. It is my choice for any gifts 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.

Automation vs. workflow vs. agent

Now the next words: automation, workflow and agent. Some people use these three words like they mean the same thing, and they really don't.

An automation is the same trigger, the same steps, the same result every single time. There's no judgment. For example, a form fills in a row whenever a specific thing happens.

A workflow is a bunch of automations chained together, with a handoff from one step to the next.

An agent has a goal, it has tools, and it has permission to pick its own steps to get there, or to ask you questions so you make the decisions.

The difference comes down to one word: the agent decides, or asks you to decide.

Family three, the control words: guardrails, human in the loop, decision rights

Which is why we need the last set of words: the control words.

Guardrails are the first one. What the agent can do, what it can never do, and when it has to stop.

Human in the loop is the moment it stops and asks a person.

Decision rights say which person, not the team. A name. An actual human behind it, with a name.

The bottleneck is not where the work piles up

One more, and this is my favorite thing from both episodes. Marie-Claire's operations professor back at university told the class the bottleneck is not where the work piles up.

Everybody looks at the station with the biggest pile and goes, "Well, that's obviously where the bottleneck is." And it is not. Usually the constraint is upstream. That station is running exactly as it's supposed to. It's just being fed too much, or it has to correct something that happened before it.

So the pile you see is a symptom. And if you point AI at that symptom, congratulations, now your symptom is just faster.

Start AI where you can prove it worked

So where do you start? Antonella answered this, and I don't want anybody skipping past it, because she said it like it was nothing and it's actually a strategy.

They started AI inside operations, not because operations is an exciting place. Her reasoning was that operations is repetitive, it's high volume, and the processes are structured most of the time. It's the part where you can tell, without arguing, whether it worked or it didn't.

That is the criterion. Not "where will AI look impressive?" Where can I tell, unambiguously, whether it actually worked?

The first thing you're doing is not transforming your business. The first thing you're doing is learning as a company. And you cannot learn anything from a project where six people disagree about whether it succeeded.

Steps one to three: definitions, decision rights and ownership

So where do you start?

One: the definitions. What's a client? What's an opportunity? What does "done" really mean?

Two: the decision rights. When this thing gets stuck between A and B, who does it ask? What does it do? Name the person who is in charge of it.

Three: the ownership. Who owns the process? Who owns the edits? Who owns the data? We talked about it in these last two episodes: if you start with operations, who has the decision rights, and who's going to do each one of the jobs.

Step four: point AI at your process to document it

Then you start the fourth thing: documenting your process.

One trick I like to show clients when I'm working with them is that you point your AI to do the job with you. Find the process that's causing the bottleneck, the actual constraint that's making things pile up at another stage, and that's where you go. You point the AI at it to help you document exactly what you're doing.

This is why we talked about the sandwich at the beginning. You need to write down literally everything you're doing.

I talked about this with Marie-Claire in the previous episode. As humans, we have reflexes. There are decisions that come very naturally to us. She said it, and I've experienced it myself: somebody can call me and I can walk them through a system: click here, click there, then you're going to see a button in this shape, in this colour, because I've used the tool so many times that I recognize the buttons and have them in memory. I could be doing anything, you call me, and I'll be able to walk you through it.

Those things normally get lost when you write the instructions for a process. When you write an SOP as a human, there are nuances that are invisible to you. But when you point the AI at the process and have it observe you executing it on your computer, for example, it will find all those nuances and write them down.

So that's the fourth step I use with clients and that I advise: let's sit down and you walk me through the process. There will probably be a recording of the session, or a Cowork agent watching every step of the way as you click here and click there, or the recording gets passed to an agent that does the analysis to understand exactly what the actions are.

Step five: decide where a human has to decide

The fifth step is to decide, based on this process, where the decisions need to be made. This is where we need to ask the questions before proceeding to execution. We want to make sure there is always a human in the loop making the decisions.

There are basic decisions that, yes, can be handed to an agent. But a decision like "we need to increase this forecast by 30% because, as a company, we've decided to push for growth." The agent is not going to know that unless you feed it to it somehow. So that's the decision point that needs to be written in when you're writing an SOP, the instructions for the agent to go and do the forecasting process.

And that's the last step of where you should start with AI in the business.

Where to start tomorrow: choose the right tool first

I want to close this episode with a specific example of how I've implemented this for customers, so you have an idea of where you could start tomorrow if you haven't started developing AI within your business yet.

At the enterprise level, as we discussed with Marie-Claire, risk management and risk assessment are already done. If you're working at an SMB level, it's very important that you choose the right tools for your business, the ones that save your information and protect it.

After you decide which tool you're going to use, sit down and choose the process that takes you the longest and involves a lot of data gathering, for example. You could pick that one. Or you could pick a process that doesn't bring any value but is necessary.

Example: validating allergens and specs across every system

In the food industry, there's a lot of documentation that has to happen. There are a lot of checkpoints for allergens, for certifications, things like that, and that is a very good example to point your AI at.

Once you've defined the allergens and the specs of your product, you can point AI at validating whether that information is correct across all of your systems.

The way to do it: turn on Claude Cowork, point it at the website you're going to be using, then go through the process of validating each allergen written in your spec, and validating it in every site, app or piece of software where it needs to be updated.

Turn the checklist into your first agent

Then ask Claude to write you a checklist, or a documented instruction for another AI agent to actually execute. And that's how you build your first agent, if you haven't so far.

That's my recommendation for today. Use whatever tool you're using. My preferred one is Claude, and I keep saying it: she's Claudia. She is my best friend. She's my assistant. I bring her in with me when I'm doing a process for the first time, or when I want a process to be automated in the future or to have a workflow around it. I connect her to whatever I'm doing and say, "Okay, this is what we're doing. Write me the instructions, write me the checklists." Then I hand that off to the next person or agent who is going to use it. Works wonders.

There's also a lot of reporting you can do easily with AI, and I hope this helps.

Start with what's repeatable, and don't wait for permission

It's really a matter of starting with where your operations are: what is repeatable, and what can you document. And think outside the box. If there's anything you haven't thought of yet, ask AI and just start using it.

Don't wait for permission. Don't wait to be perfect at it. Just ask the questions and start using it as if it were your personal executive assistant, and you're going to see where that can take you.

New to the idea of a forward deployed operator? Start with what a forward deployed operator is.

This has been The Forward Deployed Operator. Thank you so much for listening. Until the next one.

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.