Reading time: 7 minutes | Issue 30 | Book a Call

Happy Tuesday.

Mark here.

Satya Nadella dropped a post on Saturday that got 60 million views. He introduced a concept he calls "token capital" and argued every company needs to build it alongside human capital.

His direction is right. His framing skips the part that makes it work.

Nadella described a world where companies build AI systems that compound by combining human judgment with model capability. He called it a "hill climbing machine." He didn't describe what happens when you try to build that machine inside a company that can't draw its own workflow map.

Inside the Issue

Nadella wrote: "Every company is going to have to build what I think of as human capital and token capital."

He described token capital as a firm's AI capability. The systems and models it builds and owns. He argued human capital becomes more valuable as token capital grows. He called for companies to build agentic systems, private evals, learning loops.

All correct.

But token capital is not capital by itself. It becomes capital when it's embedded in a system that converts human judgment into reusable, compounding workflows. The scaffolding around the model is the asset.

Designing those loops requires process design. And what is process? The boring thing your COO keeps bringing up while everyone else wants to move fast.

What does the process need to be compliant? Governance. Another boring word people hate to hear.

Without process, governance, and compliance? You get AI slop. Which is what most companies are shipping right now.

We see this in our engagements. A company buys the best model. Points it at their workflows. Gets output that works but doesn't compound. No structured feedback loop. No governance layer defining what good looks like. No process architecture the model can learn from. The model runs on vibes. Fast vibes, but vibes.

This is where AI value lives:

Multiplication, not addition. If any multiplier is zero, the whole thing is zero.

Your scaffolding is zero? Your frontier model gives you suggestions you never implement. Your feedback loops are zero? Your scaffolding never improves. Human capital is zero? Nobody sets the goals, validates the output, or catches when the model is wrong with conviction.

Nadella said "human capital does not become less valuable as token capital grows." True. But he undersold it. Human capital is a multiplier in the equation. Pull it out and the whole thing collapses. Your people become more important in an AI-driven company.

Who is more successful: the smartest person in the room, or the most disciplined and persistent one? Same applies to AI. A smaller model with proper scaffolding runs your business. The frontier model without scaffolding gives you coaching advice and you get nothing done.

You saw that AI Buildout Supply Chain infographic floating around X this week. Twelve layers from critical minerals to the application layer. A map of what you need to build AI.

That's the infrastructure stack. What's missing is the value stack: the organizational layers that determine whether any of that infrastructure delivers business results.

This is how we think about it:

Seven layers. The bottom three are the foundation: process design, governance, knowledge architecture. The middle three form the compounding loop: human judgment, feedback loops, scaffolding. The model sits on top. Thin by design.

Most companies start at Layer 7. They buy the model. They skip Layers 1 through 3. They end up with AI that produces output but doesn't improve, doesn't compound, and never becomes institutional knowledge. They get velocity without direction.

The supply chain image shows what you need to build AI. This stack shows what you need to get value from AI. Different problem. Different investment.

LLM ops is a new abstraction layer. But it doesn't make everything else obsolete. It demands thought patterns and modes of operation that companies stopped investing in years ago.

Running agentic workflows requires your team to think at the system design level. You need people who can map a process before automating it, who understand that governance is the quality signal that makes output trustworthy.

The data backs this up. Info-Tech's "Future of IT 2026" survey found 58% of organizations have integrated AI into enterprise-wide strategies. Up from 26% last year. But only 30% feel prepared to operationalize. That 28-point gap is made of the layers most companies skipped: process maturity, data infrastructure, governance, feedback loops.

78% of executives say AI is advancing too fast for their training efforts. 82% of companies in early AI maturity haven't implemented a talent strategy.

Organizational readiness is the bottleneck.

01 Nadella's 60M-view post is a strategy document, not an informative post.

When the CEO of Microsoft publishes a 1,200-word essay introducing "token capital" as a concept, he's setting the narrative for enterprise AI purchasing. Two weeks after Microsoft launched 7 in-house MAI models and restructured the OpenAI deal. The message is consistent: the model is temporary, your organizational capability is permanent. His direction is right. He skipped the scaffolding part.

Sources: BeInCrypto, ForkLog

02 Apple opened Siri to Claude and Gemini. The model layer is now a commodity input.

At WWDC, Apple rebuilt Siri and built an Extensions framework in iOS 27 that lets users swap between ChatGPT, Claude, and Gemini inside Siri. Xcode 27 brings coding agents from Anthropic, Google, and OpenAI into the same workflow. Apple made models interchangeable. Apple is proving Nadella's implication: if you can swap the model, it's not your competitive advantage. Your advantage lives in the system around it.

Sources: PopSci, TechCrunch

03 58% of enterprises have AI strategy. 30% can operationalize. The gap is scaffolding.

Info-Tech's "Future of IT 2026" survey shows adoption doubled year-over-year but readiness didn't keep pace. 78% of executives say AI moves faster than their training. 82% of early-maturity companies have no talent strategy. The bottleneck is process maturity, data infrastructure, governance, and human judgment integration. The same layers in the value stack that most companies skip.

Sources: IBM, TechRepublic, EIN Presswire

04 The US government shut down Anthropic's most powerful AI. The vulnerability was three words: "fix this code."

Anthropic launched Fable 5 on June 9. By June 12, the US government issued an export control directive forcing Anthropic to disable it worldwide. The trigger: researchers bypassed Fable 5's safety guardrails using multi-step jailbreak techniques, producing working exploit code and synthesis instructions the model was built to refuse. Anthropic disputes the severity, says the vulnerability exists in other public models too. Doesn't matter. The government pulled the model because the governance layer failed. Three days from launch to shutdown. The most capable model Anthropic ever built, and it got killed not by a competitor but by missing scaffolding. The safety architecture around the model was the asset. When it broke, everything broke.

Sources: Fortune, TechCrunch, Time

The thread connecting all four: Microsoft's CEO tells enterprises to build learning loops. Apple proves the model is interchangeable. The enterprise survey confirms companies lack the organizational foundation to execute. And the US government just showed what happens when your scaffolding fails at the highest level.

Before your team builds its next agentic workflow, ask one question:

Can you draw the process map the agent will operate inside?

If you can't, you have a scaffolding problem.

We build that first on every engagement. We call it Step Zero. Scan the workflows. Map what's current. Define the feedback loops. Build the structured context that turns a model into a compounding asset instead of a fast suggestion engine.

It takes two weeks. It prevents months of AI slop.

Two slots open this month.

Until next Tuesday,

— Mark Ajzenstadt

Founder, Limestone Digital