How to build AI capital

In An agent in a folder we made a simple point: when AI gets a workspace with files, instructions, examples and output, it stops being a chat and starts becoming a self-improving work environment. This guide takes the next step.

Satya Nadella published the essay "A frontier without an ecosystem is not stable", about what makes companies durable when AI does more and more of the work. He separates human capital, the knowledge and judgement of people, from what he calls token capital: the AI capability and learning infrastructure a company builds and controls itself. His core point, as we read it, is that the value does not sit in picking the best model, but in building a learning loop on top of the models where the two kinds of capital reinforce each other.

That puts words on something we have been calling AI capital. If a folder is where the AI works, AI capital is what accumulates when the work, the feedback and the judgement stay in the system. Not as a chat log somebody has to search for later, but as files, skills, examples and rules you own, improve and reuse.

The big win is not asking AI better questions every time. It is building AI capital that gets better with every iteration.

The story of Arne

Every company is full of human capital. There is craft in how people solve tasks, built up over years of trial and error. People know where mistakes are easy to make, which decisions must come early, what needs an extra check. Most of it has never been written down.

One of our clients did something clever about this. A key person was approaching retirement, and was himself curious about how AI could preserve some of the tacit knowledge he carried. They sat him down in front of Claude with a simple instruction:

I retire in three months. Interview me for two hours to capture as much as possible of my knowledge, so I can share it with my colleagues.

It worked so well that he did several interview sessions before his last day. Afterwards the company built the agent Arne, which employees can ask questions. They still call the real Arne when it matters, but for quick everyday questions the agent has become genuinely valuable.

This does not make people less important. It makes the best people more important, because their judgement is what gets captured. The difference is that the judgement no longer disappears when the task is done.

From folder to capital

An agent in a folder gives the AI a workspace. The value comes when the folder gets better from being used. When feedback is saved back instead of staying in a conversation. When good examples are added. When a checklist improves. When a skill captures the way you solve a task.

Without this, AI use is consumption: you use the model, get an answer, and move on. With it, AI use is investment: every round leaves something behind that makes the next round better.

The learning loop

The learning loop is the cycle that turns human capital into AI capital. It starts with the person who knows the task best using the AI on real work, not in a workshop on the side. First drafts, questions, corrections, explanations of the judgement calls.

The first round feels heavy. That is the investment. Then the expert gives feedback, short and concrete: "This is too generic for a client who already knows us." "Risk should be explained before price here." "This kind of report always starts with the deviations, not the summary."

What happens after the feedback is the whole game. If it stays in the chat, it is nearly lost. If it goes into the folder, into an instruction file, an example or a skill, it has become part of the system.

Skills are the practical package

In the previous guide, CLAUDE.md was the central file: it explains how the agent works in one workspace. Skills are the next level. A skill is in practice a folder with a SKILL.md, references, templates and optionally scripts, packaged as a portable unit. Where the agent in a folder gives AI a workspace for one task, the skill makes the way of working portable.

A proposal skill can describe how you assess client, price, scope and risk. A reporting skill can define which deviations get surfaced first. A tone of voice skill helps the AI write like the company, not like a generic model.

One more reason we recommend skills: every major AI vendor is adopting the concept. If a different tool is better in six months, the company brings its skills along instead of starting over.

Where the compounding comes from

AI capital grows because the learning stays. Without the loop, every task starts near zero. With it, every round contributes: a rule gets clearer, an example gets added, a script does a check that used to be manual. Not because the AI becomes perfect, but because you stop throwing learning away.

Which tasks should become AI capital

Not everything needs a skill or a folder. For simple one-off tasks, chat is enough. For rule-based repetitive tasks, plain automation may be better. AI capital pays off most where the task repeats often and requires judgement.

That is usually where the company holds the most hidden knowledge. Unsure where to start? Ask the AI to interview you about your company, your role and how you work.

From personal experiment to company standard

Do not build AI capital centrally first. That gets heavy and lands far from the actual work. Start close to the tasks: let individuals build small skills and folders for their own work. When something works, the team adopts and adjusts it. When a skill matters to many, the company owns it as a standard.

Give every important skill an owner. Sales owns the proposal skill, finance owns the reporting skill, HR owns onboarding. Then this becomes part of the work, not a separate AI project.

Three steps to get started

1. Choose where AI capital should be built first. Look for areas with much experience and little structured context.
2. Create awareness and give teams room to test. Leadership does not need to design skills. The job is to help people understand what AI capital is and how to start capturing working methods in practice.
3. Build a sharing culture around skills. Good working methods should be easy to find, improve and reuse. Over time the best skills become shared standards.

The first goal is not a perfect structure. The first goal is to get the learning out of people's heads, out of the chat, and into the workspace. A folder gives agentic AI a place to work. A skill captures how you work and makes it portable. A learning loop makes both better over time. That is when AI use starts building value you actually own.