The Semantic Anchor: You Can Own All Your Data and Still Lose What It Means
Your organization can own every document, database, and piece of intellectual property it has ever created and still lose something far more difficult to reconstruct: namely, what all of it means. If you need a specific AI model or vendor platform to recover that meaning, you may own the information without controlling the knowledge. You have lost your semantic anchor, that plain-language understanding of what your information means, kept independent of whatever machine is processing it.
Anyone who has survived a major system migration knows this problem. The old system is overdue for replacement. Then the migration team discovers an obscure field or unwritten business rule nobody can explain, until someone finds the veteran employee who remembers the regulatory dispute or operational failure that created the exception. What looked like technical debt turns out to be institutional memory.
AI makes this problem deceptively easy to recreate. Give a capable model enough documents, prompt instructions, and feedback, and it becomes remarkably fluent in how your organization works. That creates a quiet strategic risk: the AI becomes better at applying organizational knowledge while the organization becomes less capable of stating independently what the AI has learned. The strategic question becomes how much of that knowledge you can recover when the machine changes.
When the System Knows More Than the Organization
Most enterprise AI deployments start innocently. A system prompt acquires sixty instructions. A vector database indexes internal documents. A vendor platform records custom rules.
Over time, the line between configuring a system and burying institutional memory inside it vanishes. Some prompt instructions represent actual policy; others preserve old exceptions, patch model weaknesses, or recall past mistakes. When no one can separate them, the prompt isn’t just an instruction set. It has become a dark repository of business meaning.
The technical mechanisms differ (RAG, vector databases, prompt libraries, or fine-tuned weights, to name a few), but the governance question remains identical: If we removed this model or vendor tomorrow, could we still explain the important distinctions we taught it to make?
Models are temporary containers. Vendors shift pricing, architectures, and interfaces. Today’s elaborate prompt workaround becomes obsolete when tomorrow’s model acquires the capability natively. Rapid technical improvement isn’t a failure of AI. The failure is allowing durable organizational knowledge to become inseparable from machinery we know will change.
Meaning Is Not Information
Information is not the same thing as meaning.
- A database contains information. Knowing why two apparently identical accounts are handled differently requires meaning.
- A contract contains information. Knowing why a company has never enforced a specific indemnity clause requires history, precedent, and judgment.
Lawyers understand this distinction instinctively. Possessing the statute isn’t the same as understanding the case law interpreting it. The words matter, but so do the definitions, practices, and exceptions that govern how those words operate in the real world.
AI is exceptionally good at smoothing over organizational messiness and reconciling sloppy terminology, inferring patterns, and synthesizing messy documents into clean answers. Often, that is all that we want. The danger arises when useful synthesis quietly erases a distinction the organization paid dearly to learn.
Institutional knowledge has intentional bumps in it. The experienced underwriter recognizes the standard-looking file that carries hidden risk. The lawyer knows why a boilerplate clause stopped being boilerplate after a lawsuit. Some exceptions are obsolete habits; others mark where general rules collided with reality. A prudent organization knows which is which before the machine smooths the distinction away.
Keeping Custody of Meaning
Semantic stewardship means keeping organizational definitions, rules, and context independently understandable in plain language, even as the underlying models change. The goal isn’t to document every piece of institutional folklore. Instead, you want to keep critical business meaning from becoming captive to a single vendor or implementation.
This requires separating the reasoning engine from the context it uses. Models should be replaceable while definitions, risk thresholds, and exceptions remain portable and recoverable. An organization must be able to explain its core distinctions without asking the current model to demonstrate them.
There is an underappreciated AI paradox. It is a curious kind of ownership when you own all the data but need someone else’s machine to tell you what it means.
The Reconstitution Problem
You discover whether you preserved organizational meaning the moment you try to change systems. Imagine an organization that spent three years refining AI workflows on a vendor platform. The system works well. Then the vendor is acquired, or management consolidates platforms. The company exports its documents and data. The contract guarantees ownership of its IP. But those protections fail to answer the harder questions:
- Why does the system classify these accounts differently?
- Where did this risk threshold come from?
- Is this prompt instruction a vital business rule or a patch for a model retired two years ago?
If answers can only be found by reverse-engineering the old system, you don’t have a technical migration. You have an archaeology problem. And archaeology is an expensive way to rediscover what you used to know.
This is why data portability is necessary but insufficient. A clean export preserves every byte while losing the institutional context that gave those bytes practical value. That is knowledge lock-in far more so than it is vendor lock-in.
The Diagnostic: The Semantic Independence Question
You don’t need a massive knowledge-management initiative to locate this risk. A single diagnostic question uncovers it:
What knowledge, context, or business rule would become inaccessible or unintelligible if we replaced our models or vendors tomorrow?
If the answer points to a system prompt, check if the reasoning exists elsewhere. If it points to a fine-tuned model, ask if those distinctions are documented. If it points to a single employee, ask whether the organization owns the knowledge or merely has temporary custody of the person who does.
Reconstitution time is the ultimate metric. A prompt experiment recreated in an afternoon carries minimal exposure. A workflow whose core logic takes six months to rediscover deserves immediate attention.
Models Are Replaceable. Meaning Isn’t.
Cheap, abundant reasoning capability should make organizations less attached to specific AI models, not more. The durable enterprise asset isn’t the model. It is the accumulated context that tells the model what matters: what your terms mean, which distinctions count, and where the real exceptions lie.
Models will change. Vendors will change. Architectures will change. Organizations should capture those improvements without migrating their identity every time they upgrade their technology. Models are replaceable. Your organization’s meaning shouldn’t be.
Prudent AI is not a prescription for moving slowly. It governs consequential AI commitments under uncertainty while preserving the capacity to learn, adapt, and recover. Move fastest where experiments are cheap, failures are reversible, and learning is valuable. As AI systems become more deeply embedded and harder to unwind, demand stronger evidence and preserve what you would need to change direction, including the independent custody of what your organization knows and what its information means.
Accelerate reversible learning. Pace irreversible commitment.
Dennis Kennedy – CC BY 4.0 license
[Originally posted on DennisKennedy.Blog (https://www.denniskennedy.com/blog/)]
DennisKennedy.com is the home of the Kennedy Idea Propulsion Laboratory
DennisKennedy.Blog is part of the LexBlog network.
Recent Comments