The Knowledge Lock | Work Is Code

You built something. It works. The architecture is sound, the tests pass, the documentation is thorough. Everything a competent engineer needs to understand what you did and why is right there in the repo.

Your investor has never seen any of it.

Not because they chose not to look. Because they can’t. No IDE, no git credentials, no context for what a commit message means. The knowledge exists. It’s trapped in a format they can’t open.

This Has Happened Before

Nobody writes machine code. The raw instruction set of a processor is incomprehensible to almost every human who has ever lived. But that didn’t stop people from telling machines what to do, because someone built a translation layer.

Assembly was the first unlock. A human-readable format on top of the machine’s native format. C abstracted assembly. Higher-level languages abstracted C. Frameworks abstracted the languages.¹ Each layer was the same move: take knowledge trapped in a format only specialists can read and make it accessible to more people. The machine’s knowledge didn’t change. The accessibility did. And each time the accessibility changed, the number of people who could participate exploded.

The entire history of software engineering is a history of breaking knowledge locks.

Now the same thing is happening to organizational knowledge. The repo, the database, the clinical system, the financial model. These are the machine code of the modern organization, unreadable to most of the people who need what’s inside them. AI is the next abstraction layer: a compiler that takes organizational knowledge and renders it in whatever format the person in front of it can actually use.

Work is following the same trajectory that code already did. That is what this site is about.

The Lock Nobody Installed

Every organization runs a permissions system that decides who gets to understand what. Nobody designed it. Nobody audits it.

The permissions are enforced by format.

The repo is readable by engineers. The Figma file is readable by designers. The financial model is readable by the person who built the spreadsheet. Not because anyone decided to restrict access. Because the tools themselves are the access control. You can’t read what you can’t open.

In software engineering, this pattern has a name: role-based access control.² Users get permissions based on their role. It’s deliberate, documented, reviewable. The Knowledge Lock is the same mechanism, except nobody built it. Format and tooling barriers do the enforcing automatically. And unlike a real permissions system, there’s no denial notification. No “request access” button. You don’t see a locked door. You don’t see anything at all. You don’t know the knowledge exists.³

Not all knowledge locks are accidental. Some are deliberate: a company restricts financial models to the finance team, a practitioner limits what a contractor can see, an engineer scopes an API key to read-only. These are access decisions. But making them well requires someone who understands the knowledge on both sides of the boundary well enough to judge what should cross and what shouldn’t. The deliberate lock depends on a person with enough context to design it. The Knowledge Lock produces the same outcome, but nobody made the decision. The format made it for them. The difference matters because the response is different. You don’t break a lock someone installed on purpose. You break the ones that installed themselves.

An entire profession grew up around the locks nobody installed.

The Business Analyst sits at the format boundary between business and engineering.⁴ The translation work is real: take what the business needs, render it so engineering can build from it; take what engineering built, render it so the business can evaluate it. Requirements documents, user stories, acceptance criteria. Three documents, one piece of knowledge, two translations.

But a good BA is more than a translator. They advise. They know enough about both domains to judge which requirements matter, what the tradeoffs are, which stakeholders need what level of detail and why. That contextual judgment is what makes the translation valuable, not the format conversion itself. AI can move knowledge between formats. Knowing what to move, for whom, and what to leave out is the part that stays human.

The BA profession is under real pressure right now, and framing it as “just translation” doesn’t help. The mechanical translation is the part AI replaces. The contextual advisory work is the part that gets more important when the cost of translation drops to zero, because more translations will happen and someone still needs to decide which ones should.

Project managers, technical writers, marketing coordinators. The same pattern repeats at every format boundary in the org chart. In each case, the role blends mechanical translation with contextual judgment. The ratio varies. The judgment is the part that endures.

Three Locks

Code to non-technical stakeholder. A CFO needs to evaluate whether a software build is worth funding. The answer is in the codebase: architecture decisions, test coverage, reusable components, production readiness. She can’t read any of it. She’s making a six-figure decision from a demo and a slide deck. If the demo is good, she approves. If the demo is misleading, she approves anyway. The quality of the underlying work is invisible to the person paying for it.

Legal and contractual documents to founders. A founder receives a threatening legal letter demanding he abandon a trademark. The knowledge to evaluate that threat correctly lives in a federal trademark database, case law, and filing history. He can’t access any of it without hiring a specialist. The letter is designed to exploit this gap. It works because the knowledge that would defuse it is locked in a format the recipient can’t open.⁵

Domain knowledge to the product. A nurse mentions during a facility walkthrough that unsigned call logs are a legal liability. That observation is worth more than most formal feature requests. But it was delivered verbally, in passing, during a demo. It exists nowhere in the product spec, the backlog, or the requirements doc. If nobody translates it into a format the engineering team reads, it disappears. The person with the most relevant knowledge contributed it in the only format available to her, and the system didn’t capture it.

Each of these is the same mechanism: knowledge exists but can’t reach the person who needs it because the format is wrong. The cost isn’t the translation itself. The cost is what happens when translation doesn’t occur: a funding decision made on incomplete information, a legal threat left uncontested, a clinical insight lost.

The Economics of the Lock

Coase argued in 1937 that organizations exist because internal coordination is cheaper than market transactions.⁶ The Knowledge Lock is a specific class of that internal coordination cost: the cost of moving knowledge between formats within the same organization.

When that cost is high, it shapes everything. Org charts crystallize around format boundaries. Departments form to ferry information between containers. Information hierarchies emerge: engineers get the detailed version, managers get the summary, executives get the slide deck, customers get the marketing copy. Each step from the source loses fidelity. Each translation introduces the translator’s interpretation, priorities, and blind spots.

This isn’t a conspiracy. It’s an economic inevitability. When every translation costs time and effort, you ration translations. Rationing creates tiers. Tiers become culture. Eventually nobody questions why the board only sees polished presentations while engineering sees raw data. It’s just how things work.

The people locked out the most are the ones furthest from the native format. The CEO is further from the repo than the engineering manager. The patient is further from the clinical data model than the nurse. Distance from the source determines who loses access first when translation resources are scarce.

When AI reduces the translation cost toward zero, it doesn’t just save time. It changes the economics that shaped organizational structure in the first place.

The Tools Exist, the Problem Is Deeper

AI eliminated the translation cost. Not reduced it. Eliminated it. With AI as the translation layer, the time to move knowledge between formats dropped from hours to seconds. The skill required dropped from “person who understands both domains” to “person who can describe what they need.”

Claude Code, Cowork, Codex, Copilot, Cursor.⁷ Connect them to a repo, a database, a document store, and they can do the translation that used to require a dedicated human. An architecture decision becomes a board-ready summary. A clinical data model becomes a plain-language protocol. A sprint’s worth of commits becomes an investor update. On demand.

While writing this article, I watched the lock operate on the AI tool helping me draft it. The tool had access to a detailed knowledge base with the full portfolio of examples I could draw from. It didn’t look. It worked from the few examples already in the conversation, unaware of what it was missing. No error. No denial. Just an invisible absence shaping every recommendation it made.

But “just connect the AI to your systems” skips two problems that the technology can’t solve.

The locked-out don’t know they’re locked out. The CEO doesn’t walk into a standup and say “I wish I could query the engineering knowledge base.” They probably don’t know it exists. You can’t adopt a solution to a problem you’ve never seen. The door was always invisible.

The hesitation is rational. Connecting an AI tool to production databases, patient records, financial systems, and source code means trusting that tool with the most sensitive things your organization has. Hallucinated medical guidance. Leaked source code. Confidential financials summarized for the wrong audience.⁸ An AI that translates freely but translates wrong creates a different problem than the Knowledge Lock. A real one.

So the lock is breaking, but unevenly. The people with the technical sophistication to connect these tools are the people who were never locked out in the first place. The people most locked out are the furthest from being able to set up the solution. And the organizations in between are weighing a legitimate question: is the cost of the lock worse than the cost of a bad unlock?

For most of them, right now, the answer is “we don’t know.” That’s honest.

What Stays Locked

Breaking the Knowledge Lock doesn’t solve everything.

Knowledge that was never captured can’t be translated into any format. If the architect made a critical decision in a hallway conversation and never wrote it down, no AI helps. The lock is broken, but the room is empty.

Judgment doesn’t translate. AI can render knowledge into any format, but it can’t decide what knowledge matters for a given audience. Knowing what the board needs to hear, what the patient needs to understand, what the partner needs to see. That editorial instinct remains human.

And trust doesn’t come from access alone. Giving someone information in a format they can read doesn’t mean they’ll believe it. Trust comes from relationships, track records, and context that no format conversion provides.

But these are better problems. “We need to build trust” is a better problem than “they can’t see what we built.” “We need editorial judgment” is a better problem than “we can’t afford to translate.” The Knowledge Lock made these real problems invisible by locking people out before they could get to the trust and judgment questions.

The Unlock

People aren’t confused because they’re incapable. They’re confused because the knowledge that would help them is locked in formats they can’t open.

The Knowledge Lock has been the invisible architecture of every organization. It shaped who got to understand and who had to guess. It created professions dedicated to carrying information across format boundaries. It built hierarchies of understanding that nobody designed and nobody reviewed.

That lock is breaking. Not because people learned to read repos. Because the format barrier stopped mattering.

The constraint was never the knowledge. It was never the people.

It was always the format.

The work moves from "can we render this" to "should we render this, for whom, and at what level of detail."

Further reading: AI as Translation Layer — the mechanism of the unlock, the practitioner stack, the liability gap, and the math organizations are running.

Sources

Dijkstra, E.W. "The Humble Programmer." Communications of the ACM, 15(10), 1972. Turing Award lecture tracing abstraction layers as the central challenge of computing. doi.org/10.1145/355604.361591 ↑
Ferraiolo, D.F. & Kuhn, D.R. "Role-Based Access Controls." 15th National Computer Security Conference, 1992. Adopted as NIST Standard 359. Defines permissions as a function of organizational role rather than individual identity. csrc.nist.gov ↑
Akerlof, G.A. "The Market for 'Lemons': Quality Uncertainty and the Market Mechanism." The Quarterly Journal of Economics, 84(3), 1970. The Knowledge Lock creates information asymmetry not through concealment but through format inaccessibility. doi.org/10.2307/1879431 ↑
International Institute of Business Analysis. BABOK Guide, v3, 2015. Defines business analysis as "the practice of enabling change in an enterprise by defining needs and recommending solutions that deliver value to stakeholders." Codifies the format-bridging function across six knowledge areas. iiba.org/babok ↑
Lemley, M.A. & McKenna, M. "Irrelevant Confusion." Stanford Law Review, 62(2), 2010. On how trademark enforcement mechanisms create costs disproportionate to the underlying rights. SSRN 1407793 ↑
Coase, R.H. "The Nature of the Firm." Economica, 4(16), 1937. Organizations exist because internal coordination costs are cheaper than market transactions. The Knowledge Lock represents a specific class of internal transaction cost that AI is driving toward zero. doi.org/10.1111/j.1468-0335.1937.tb00002.x ↑
As of early 2026, agentic coding tools include Anthropic's Claude Code and Claude Cowork, OpenAI's Codex, GitHub Copilot (Microsoft), and Cursor (Anysphere). Each connects to organizational knowledge stores and can generate natural-language outputs from technical source material. ↑
OWASP. Top 10 for Large Language Model Applications, v2.0, 2025. Identifies "Sensitive Information Disclosure" as a primary risk category for LLMs connected to organizational data. owasp.org ↑