There's AI the model, and there's AI the infrastructure

A COO of a 150-person organization told us where they are with AI. They wrote an AI policy early. They surveyed their staff. They gave a license to everyone who asked for one, and most of them asked.

Nine months later, in his words: "We’ve handed the tools out." Then the question: "These tools are amazing. We see what they can do. They’re really kind of breathtaking. But how do we actually use it, beyond a chat where you ask a question and it writes an answer?"

Where most companies are

This is the state we see most often.

People use AI on their own licenses, or inside the third-party tools that all want to add some AI. It’s very scattered.

Some people rise to it and become active thinkers. They love it, and say it’s the greatest thing they’ve ever had. Some feel threatened. Most are dabbling: reword this, clean up that email, and the work is done.

Why it stalls

You still have the same problems.

You have to onboard people. You have to train people. When people leave, the knowledge goes with them.

Your champions make real gains. But unless they’re extremely organized at sharing and teaching, those gains don’t live anywhere the rest of the company can use. So even the champions only compound themselves. They can’t lift their team.

You have the skills and you have the people. What’s missing is the structure underneath, the one that says: the best version of this task can now be done by all five people who run it.

How we came to be

My background is growth, marketing, product and sales.

In my last role, as COO, I ran marketing, operations and HR, managed the board and every outside counterpart. We had to grow the company. We had a very big remit, multiple entities, multiple businesses, and no ability to hire.

We knew AI was here. We just couldn’t figure out how to run it.

So we asked: what would it take to have intelligence in the company that does the tasks? That took us on a crazy journey of almost 18 months, trying everything. Pierre Merzeau, now our CTO, was head of engineering at that company, and he kept being my guide. I made myself a proclaimed chief AI officer without knowing much about AI at the time, and went down the complete rabbit hole figuring out how to make an organization run on AI.

The aha came about 12 months in. There’s AI the model, and there’s AI the infrastructure, and they are two fundamentally different things.

Give people licenses and they can dabble with AI, but nothing else changes. Pairing every person with the best agent, trained for their function and every subfunction in the department, is a fundamentally different endeavor. Once that works, the company is fundamentally different.

It’s the most energizing thing I’ve worked on in my career, and it had to become our business. Since January we’ve onboarded about a dozen clients, very small, medium and multi-billion-dollar, and every time it’s exactly the same principle.

Three things, inside one

When we say infrastructure, we mean three things that live inside one bigger thing.

• Your people’s expertise. Right now it’s in people’s heads, and rarely in an SOP. The only time it really comes out is when you hire someone and hand them the real information they need. An agent has to be taught the same way: a learning phase, a calibration phase, a loop, until it works at par.
• Your data and systems. If the agent has arms and legs, it needs to reach into the systems to do the work.
• The model. Suddenly it’s a third of the importance, whereas by default, it’s 100% in most places. The model is only useful if it has the data and the expertise.

All three live inside one operating system. The technical word is a harness, and it’s popping up everywhere: every company is now trying to build the ability to have AI inside.

Pierre explains what the harness does better than I do:

"The role of the harness is to make sure that you get the same output every time you ask, and that your team members get the same output if they ask the same thing as you. And there are agents that try to disprove the result, because AI has a tendency to lie, a tendency to please. They check it without using the same information, and then it gives you the source: this number comes from this system, you can check it. Once you have that, you can really start to trust the output." — Pierre Merzeau, CTO

The job changes

Think about a building full of people doing calculations by hand. Some were very fast, some were slow. Then I give you a calculator. If you keep calculating by hand, it’s still nice and still good, but it’s no longer valued. And the people next to you who use the calculator are now much faster than you.

Once AI is organized the right way, on the right fundamentals, the nature of work changes.

People become the drivers. They define, approve and improve. They are not doing.

Everything you type on your computer is doing: publishing a web page, writing a blog post, drafting a board deck, building your P&L. Your job is to define the expertise the task needs and what the best version of it looks like, until the only thing left to say is "run the task." It runs before your eyes, or overnight, ready for your review in the morning.

The person stays accountable. You can’t say ChatGPT told you that, or Claude made you do it. You’re the thinker, and that’s why you were hired. Your hands are on the wheel. That doesn’t mean you do the task; it means the best way to do it now runs every time, and your job is to ask what else it could do.

That’s a harder job, by the way. You burn more calories. And you’ll see two kinds of people: those burning more calories every day, and those who have stopped thinking and take whatever they’re given. That’s why change management is part of the work. You’re teaching people a new way to work, and that is not trivial.

Underneath all of it, every task lives in a shared record, private where it needs to be. Every output improves the next input. Improve 1% here and 1% there, every day, and it compounds. That’s what AI should do: not only reach parity with a person, but improve the task every single day.

Start with one function

You don’t have to think about the whole company. Start with one function, wherever your biggest leverage is: the biggest cost or the biggest upside.

And forget tokens. Tokens are not a unit of adoption. Output per employee, for the function they’re accountable for, is the real measure: time saved, how complete the work is, dollars saved, the people you didn’t have to hire because hiring was the only way to do more.

Then you build one agent per role. A function might have one agent; it typically has five to eight.

Each task gets taken apart and rebuilt:

• What is the goal of the task, where does its output live today, in what system, on what cadence, and what is the true problem?
• What expertise does it need? We draw a hard line between inputs and knowledge. A document is an input, a conversation is an input. Knowledge is what we’ve processed into the best information on how to do this.
• What are the steps? Read this knowledge, do this, verify it.
• Which systems does the agent need, the same ones a person used to connect by hand?
• Which agent disproves it? A second agent comes in and assumes everything it was given is wrong. The output reaches the person only once it’s confirmed with evidence and sources.
• Does it run on a schedule? Every Monday morning, every three hours, at the end of every month.

The moment of approval is objective. It’s when the head of the function says: this agent does each task with the same precision my people do, the way they do it.

And then something changes for the team. Today five people do the same job, and each has their own recipe. Capture the absolute best way, and each of them runs it through their agent, customized to their voice and their needs where it should be. The agent always follows the strategy. Every person is doing the best version of the best performer in the function.

Patrick sees the same thing from the brand side: "This structure also helps a tremendous amount with brand management. You get to continually manage the brand across all of your functions, making sure the articulation is exactly the way you’ve always wanted it, the way you built it, and the way it needs to flow through all the different functions. That’s always trying when you’re managing a brand across a larger organization." — Patrick Adams, CMO and President

"The level of effort must be massive"

A CMO said this to us, right after saying the approach made complete sense: "I see the exact reason organizations like ours aren’t really moving forward. The level of effort to get to what you just walked us through is, I assume, massive."

Doing this on your own, yes, it’s massive. That’s true. Someone has to go through this pain.

The infrastructure. It took us about 18 months to build the harness. What is private lives here, what is public lives there, and before an output is shared, it’s verified, because otherwise it’s not reliable. If it’s not reliable, people lose trust, and if they lose trust, they’re not going to use the system.

The method. You have to encode the judgment in a way the agent can do it, and that’s not trivial. You need a method that says what to capture about each task, so that within three to four weeks, not six months, you have the true decision tree, the principles, the measure of success, all the way up to the person who was the foremost expert on the task saying: I trust it, it’s doing the work like me.

Your people. Training individuals to work on AI is not normal. It’s not natural. We’re all learning to do it.

The person. Somebody has to take the job of understanding, capturing, translating, mapping and doing the change management. It cannot be done by the people who are paid day-to-day to do the task. There’s no way to expect them to do that.

You also don’t migrate anything. As Pierre puts it: "We don’t ask any of our clients to change systems, cut access or move things. The old way of working and the new way stay side by side. And what we see, naturally, is that once people start using the new way, they forget to do the old way, because the new way is much easier, much faster and more complete."

It’s a fallacy to think you can get people licenses, and then all of a sudden they’re going to rethink how to rewire everything, and have the time to think it through. So we help companies grow by moving them to run on agents their people direct.

There’s a way to hire. Whether you’re a startup or a very big company, you need a hiring model, an onboarding model and a culture that gets people working the way the organization needs.

AI is the same. You run it the way you’d think about growth through people. Now it’s growth through people plus intelligence, connected together, leveling up both the company and the person.