The New Capital Inside a Company
I used to think about AI mostly as a capability layer.
You add it to a workflow. You make a process faster. You summarize something. You generate something. You automate a task that used to be manual.
That is still useful, but it now feels too small.
The better way to think about AI is as a new kind of capital inside a company. Not capital in the financial sense, but in the productive sense. A company has human capital. It has brand capital. It has operational knowledge. It has relationships, processes, data, judgment, taste, and habits that have accumulated over time.
AI gives companies a new question:
How much of that can compound?
Token Capital
Every company is going to have some version of token capital.
By that I mean the reusable AI capability that comes from the company's work: prompts, workflows, agent traces, evals, fine-tuned behavior, internal context, models, customer interactions, support patterns, product decisions, and the small pieces of judgment that get embedded into systems over time.
The interesting part is that token capital is not just "we use AI."
That is like saying a company has a finance strategy because it has a bank account.
The real question is whether the company is turning its daily work into reusable intelligence. Are the conversations, edits, decisions, exceptions, and corrections becoming part of a learning loop? Or are they disappearing into chat windows and vendor systems with no durable advantage created?
This matters because the companies that learn fastest will not only be the companies with the most data. They will be the companies that know how to convert work into improving systems.
Tacit Knowledge Is the Real Treasure
Most companies talk about data as if data is the whole asset.
I think that misses the more important thing: tacit knowledge.
Tacit knowledge is the stuff that is hard to write down. It is how a good clinician knows a note feels wrong. It is how a support person knows a customer is anxious even when the ticket sounds simple. It is how a product person knows a feature is technically correct but emotionally confusing. It is how a team has learned, through years of mistakes, which details actually matter.
That knowledge usually lives in people.
It also lives in artifacts: emails, tickets, Slack threads, call notes, product specs, customer interviews, code reviews, onboarding documents, sales objections, support macros, implementation details. But the artifacts are not the knowledge by themselves. The knowledge is in the pattern of judgment behind them.
AI can start to capture those patterns.
That is powerful, but it is also dangerous. If the most valuable knowledge in a company is being expressed through AI tools, then the question becomes: where does that knowledge go?
If it trains systems the company controls, it compounds.
If it leaks into systems the company does not control, the company may be giving away the thing that made it special.
The AI Supply Chain
This is why I think every company needs to understand its AI supply chain.
Not in a vague procurement way. In a concrete way.
What models are being used? What data is being sent? What traces are being created? Who owns the outputs? What feedback loops exist? What is being learned from users? What becomes reusable IP? What leaves the company forever?
In healthcare, this feels especially important. A clinical workflow is full of sensitive context, professional judgment, patient nuance, and specialty-specific practice patterns. If an AI system helps a clinician document faster, that is great. But the long-term value is not only the note. It is the learning loop around how clinicians edit, correct, trust, reject, and adapt the output.
That loop is part of the product.
It is also part of the company's future intelligence.
The Firm Becomes a Learning System
The companies I am most interested in are going to look less like static org charts and more like learning systems.
People will still matter deeply. Probably more than ever. But their work will increasingly create traces that agents can learn from, systems can reuse, and teams can inspect.
The firm becomes a loop:
- Humans do work.
- AI helps with the work.
- The company captures the trace.
- The trace improves the system.
- The improved system helps humans do better work.
That loop is where the advantage lives.
The trap is thinking the advantage is simply having access to the best model. Models will matter, but access to a model is not the same as owning a capability.
The capability comes from the combination of model, context, workflow, feedback, evals, data rights, and human judgment.
This Changes Product Strategy
For product teams, this changes how I think about building AI features.
The product cannot only answer: did the user get the output?
It also has to answer:
- Did the system learn from the user's correction?
- Did we preserve the right context?
- Did we avoid collecting what we should not collect?
- Did we make the workflow better next time?
- Did we create something reusable for the customer or only for ourselves?
- Did the customer keep control of their own knowledge?
That last question matters a lot.
If I am building tools for clinicians, operators, or knowledge workers, I do not want to strip-mine their expertise. I want the product to help them compound it. The best AI products will make customers feel more capable, not more replaceable.
The New Balance Sheet
I do not think companies will literally put token capital on a balance sheet anytime soon.
But strategically, they should know what it is.
What proprietary intelligence did we create this week?
What did our users teach the system?
What did our team learn that can now be reused?
What tacit knowledge became explicit enough to compound?
What did we accidentally give away?
Those are going to become executive-level questions.
Because AI is not only a tool you use. It is a way a company remembers, learns, and improves.
And the companies that understand that early will not just move faster. They will become harder to copy.