www.iqsource.ai Your Company's Tacit Knowledge Belongs in a Model It Controls
Ricardo Argüello — June 27, 2026
CEO & Founder
General summary
Satya Nadella argued that AI models are the new database market: companies should embed their accumulated tacit knowledge in weights they control. If that's right, the first step isn't building the model, it's documenting the knowledge that will go into it.
- Nadella compared the model market to the database market: there should be as many models as firms in the world, each containing the firm's accumulated operating knowledge.
- Tacit knowledge, the decision criteria that live in people and never got written down, is what generates durable competitive advantage. That knowledge doesn't exist in any generic model.
- The gap: you can't embed in weights what you haven't documented first. Most companies can't describe how they actually make their best decisions.
- Embedding knowledge in weights carries its own risks: it becomes opaque, non-portable, hard to audit. The discipline of prior documentation is what keeps the advantage yours, not the model vendor's.
- AI Maestro's first phase is exactly that prior documentation work: mapping real processes and decision criteria before any model gets trained or deployed.
Imagine your company has spent fifteen years learning how to evaluate whether a supplier is trustworthy. That judgment lives in three people who've been around long enough to develop it. When one of them leaves, part of the judgment leaves with them. A model trained on that documented knowledge doesn't go anywhere. That's what Nadella is describing: the model as the infrastructure where organizational knowledge lives and scales.
AI-generated summary
Satya Nadella said something this week that, coming from Microsoft’s CEO, is worth taking seriously even if it’s not yet standard practice anywhere.
Via Karl Mehta: “To me, a model is like the database market. A firm should be able to take the tacit knowledge it has and embed it inside weights in a model that they control.” And his answer to how many models there should be: “As many models as firms in the world.”
That might sound ambitious for any company outside the enterprises with their own AI research teams. But the logic behind it is more immediate and more concrete than it appears — and it starts with a problem almost no company has solved: documenting the knowledge it already has before trying to embed it in anything.
The Database Analogy Is More Literal Than It Sounds
Nadella’s comparison to the database market isn’t metaphorical. In the 1980s and 1990s, the competitive differentiator wasn’t access to a database system — it was what data you had in it. Everyone could buy Oracle or SQL Server. The advantage came from the data you’d accumulated about your customers, your operations, your suppliers.
The thesis for models is the same: everyone will be able to access capable language models. The differentiator won’t be the base model. It will be the specific knowledge your company has embedded in it.
That tacit knowledge — the decision criteria that live in your people, the operating patterns that took years to develop, the lessons learned that never made it into any manual — is what generates sustained competitive advantage. It doesn’t exist in any generic model because nobody outside your organization has it. It’s in your organization, mostly in the heads of your most experienced people.
The problem is that knowledge living in heads is fragile. When the person leaves, the knowledge leaves with them. Having it live in a model the company controls solves that, among other things.
The Problem Before Fine-Tuning
Here’s the part that almost nobody mentions when discussing proprietary models: you can’t embed in weights what you haven’t documented first.
Most companies, when asked how they make their key decisions, give answers that are the official policy, not the actual process. The official policy says suppliers get evaluated on three standard criteria. The actual process is that two people who’ve been around for twelve years know exactly which signals matter and which are noise, and none of that ever got written down.
If you try to fine-tune a model on historical decision data without understanding that tacit criterion, what you capture is average behavior — which can be considerably different from the judgment of the best people. The model learns to replicate what everyone did, including the mistakes. It doesn’t learn what makes the good decisions good.
The additional risk that emerged in the comments on Mehta’s post is real: knowledge embedded in weights is opaque. You can’t audit exactly what it learned. You can’t update a specific rule without retraining. And if you decide to change model architectures, you can’t take what you built with you in any portable form. The argument from the harness being the moat applies here: the base model is interchangeable, but what you built on top needs to be portable for the advantage to be yours rather than the vendor’s.
What AI Maestro Does With This
The first two phases of AI Maestro are diagnosis and process mapping. We don’t start by building anything because before you build, you need to know what knowledge exists and what state it’s in.
The process map we produce isn’t a corporate flowchart. It’s documentation of what actually happens: how sales decisions get made in practice, not in the manual; what signals the most experienced operations people use to detect problems before they escalate; what criteria the team applies when deciding which prospects to pursue.
That documentation work is the prerequisite for any conversation about proprietary models. The 14 stages of AI adoption that Alex Lieberman mapped start with diagnosis and process mapping for exactly this reason: if you don’t describe your real operation before you start building, what you build replicates the average, not the judgment that makes your company better than the competition.
Not every company needs a proprietary model today. But every company that wants one tomorrow needs to start documenting what it knows now.
Start by documenting what your company already knowsFrequently Asked Questions
Because the differentiating value doesn't come from accessing the same model everyone else uses. It comes from embedding a firm's own accumulated tacit knowledge — decision criteria, operating patterns, learned judgment — in a model that firm controls. That model becomes a database of organizational knowledge that generates sustained competitive advantage.
Tacit knowledge is the judgment that lives in people and hasn't been written down: how to evaluate whether a supplier is reliable, why one prospect gets prioritized over another, what signals indicate a deal will close. It can't be embedded in model weights without being documented first, and most companies haven't done that documentation work.
The main risks are opacity and portability: you can't audit a specific weight, can't update a single rule without retraining, and can't switch model architectures while keeping what you built. Knowledge trapped in weights is non-portable. Prior documentation solves this by keeping the knowledge layer separate from the inference layer.
The first step is documentation, not engineering: mapping real processes, identifying where the most valuable tacit knowledge lives, and capturing it in a form that can be used to train or guide a model. Without that prior mapping, the model captures average behavior, not the differentiating judgment that makes the company's best decisions good.
Related Articles
Cognitive Delegation, Not Cognitive Surrender
Paul Bakaus, backed by a16z, names the distinction most enterprise AI discussions miss. Delegation: you use AI to get where you decided to go faster. Surrender: you let AI decide where to go. One serves you. The other doesn't.
There are 14 stages of AI adoption. You skip to 11.
Alex Lieberman mapped 14 stages of AI adoption after 14 months with executives. Most companies skip the first ten and start by building. That is where they stall.
