Knowledge is a commodity. Judgment is the moat.

A few months ago, I shared a prompt with a coworker. It’s one I use almost daily, and it reliably produces excellent work for me. For them, it produced something generic. Same AI model, same words, same task, noticeably different results.

The difference wasn’t skill. It was history. My AI assistant carries months of my corrections, my preferences, my projects, my way of framing problems. Theirs was starting cold. The prompt was never doing the work. The accumulated context was.

That small moment changed how I think about where AI is heading, and what most organizations are missing while they focus on everything else.

Why the same words now produce different results

Two years ago, AI tools were stateless. Every conversation started from zero, so the prompt was everything, and building a library of good prompts made sense.

That era is ending. Today’s assistants draw on conversation history, saved preferences, writing style, recurring projects, and connected data. The visible prompt is only one part of the instruction the system actually receives. The effective prompt now looks more like this:

visible request + conversation history + memory + preferences + available data + tools + current circumstances

Every major vendor is moving in this direction. OpenAI describes ChatGPT as growing more useful as it builds a longer-term understanding of the user; Google says Gemini can draw on past chats to tailor its responses. Practitioners now call the underlying discipline “context engineering,” a recognition that what surrounds a request matters more than the request itself.

The implication for organizations gets far less attention. Two employees with identical licenses no longer have identical capability. One has months of accumulated corrections, examples, and project history; the other starts from an empty context. On the license dashboard they look the same. In practice, one has built a far more capable working partner.

Prompt libraries still matter. They are how people get started, and they spread good patterns fast. But they are scaffolding, not the asset. A prompt library preserves the words someone typed; it rarely preserves the accumulated context that made those words effective. What compounds aren’t the words. It’s the history behind them. Training must evolve the same way: prompting remains useful, but context literacy, knowing what the system remembers, what it can access, and what must be stated explicitly, is becoming the more durable skill.

And when that history lives only in individual accounts, organizations are quietly creating islands of personalized intelligence. When a person changes roles or leaves, the documents and prompts stay. The reasoning context that made them valuable walks out the door. That is a new form of institutional knowledge risk.

What thirty years of digitization missed

Businesses have spent three decades writing knowledge down: documents, wikis, CRMs, methodologies, training materials. AI has now read effectively all of it, which is why these tools arrive knowledgeable on day one. Knowledge was never the scarce resource; most firms are drowning in it.

But watch an experienced partner in a client meeting. They know which risk to raise now and which to hold for a later conversation. They know when a technically correct answer is the wrong move. An experienced auditor senses something is off in a set of numbers before she can explain why. A seasoned plant manager hears a change in a production line and recognizes a problem before any dashboard crosses a threshold.

None of that lives in any system, not because it’s secret, but because it was never written down, and most of it can’t be fully articulated. The philosopher Michael Polanyi named the problem in 1966: we know more than we can tell.

Knowledge got digitized. Judgment never did.

Enterprise systems record the final state: the discount was approved, the issue was escalated, the forecast was revised. They rarely record the reasoning that connected the evidence to the action. We built systems of record for what happened, and almost none for why.

This reframes the common worry that AI “can’t reason like an expert.” Today’s models reason well when the problem is checkable. What they lack is narrower and more interesting: the tacit, experience-born judgment of your specific people in your specific market. That isn’t a model problem. It’s a data problem, the training data for your firm’s judgment doesn’t exist anywhere. The AI labs understand this; they are paying thousands of professionals to record how they think through problems, step by step, because that reasoning is the scarcest data in the world. The race to capture judgment has already started. It just hasn’t reached most businesses yet.

The decision graph

So what would that data look like if a firm built it deliberately?

Not more documents. A record of decisions: the situation, the options considered, the call that was made, the reasoning behind it, and what happened next. Connect those records over time, this decision echoed that earlier one, this exception became a precedent, and you get a graph of how your organization decides. A decision graph.

Where a knowledge graph connects facts, entities, and policies, a decision graph connects the elements of judgment: the situation faced, the outcome sought, the signals that mattered, the constraints that applied, the options considered and rejected, the call that was made, who approved it, and what happened afterward.

Consider a customer pricing exception. The knowledge layer holds the standard discount policy, contract history, margin requirements, and approval limits. The decision layer shows that a similar customer received an exception after a service failure, that finance approved it under specific conditions, and that the decision improved retention without damaging long-term profitability. The outcome is the crucial part. Capturing decisions without results preserves bad precedent as efficiently as good judgment. Rationale alone is just opinion with a timestamp.

Personal AI memory is a small-scale preview of all this. My assistant got better because it accumulated a version of this record for one person, incidentally, as a byproduct of use. Almost no organization captures it deliberately, at firm scale, on purpose.

Here is why that matters competitively. Your documented knowledge is becoming a commodity. Every competitor’s AI has read the same internet, the same regulations, the same best practices. Your decision history is the one dataset that exists nowhere else and cannot be bought. In an era when knowledge is abundant, the durable asset is a record of judgment.

What to do about it now

The good news: this doesn’t require new technology. It requires a habit, plus a little governance.

• Start where judgment matters most. Highly repetitive, rules-based work doesn’t need rich decision memory. Exception-heavy work does: contract review, pricing, forecasting, underwriting, compliance, quality management, customer escalation.
• Record decisions where they already happen. Deal reviews, engagement scoping, pricing exceptions, hiring calls. One paragraph is enough: the options considered, the choice, the reasoning, the expected outcome. Software teams have done this for years with architecture decision records; the practice translates.
• Close the loop. Revisit decisions and note what actually happened. Rationale plus outcome is what turns a log into judgment.
• Let AI sit in on the deciding, not just the documenting. When an assistant drafts the decision memo or summarizes the debate, capture becomes nearly free, a byproduct of work rather than an extra task. Capturing context while the work happens beats reconstructing it months later.
• Separate personal memory from organizational memory, and govern both. A person’s preferences and exploratory conversations shouldn’t automatically become company knowledge. And remembered context is not automatically true: policies expire, experts disagree, outcomes prove old decisions wrong. Institutional memory needs provenance, dates, ownership, and the ability to retire what the organization once believed.

Whose judgment will the machines carry?

The first era of business AI was about what machines know. That race is ending in a tie. Everyone’s AI knows roughly everything.

The next era is about whose judgment the machines carry. That can’t be downloaded or licensed. It gets built from history, one recorded decision at a time.

Imagine a new employee whose AI assistant surfaces more than policies and files: relevant precedents, why similar situations were handled differently, which exceptions required approval, and how previous choices turned out. Imagine an agent that can say not only “this complies with policy” but “here are three comparable situations, the decisions made, the circumstances that justified them, and the results that followed.” That is the step from information retrieval to organizational intelligence.

Knowledge made AI informed. Memory is making it personal. A governed record of judgment can make it truly organizational.

Your knowledge made AI useful. Your decisions will make it yours.