A take on architectural strategy and AI

When Microsoft CEO Satya Nadella published his article on X, A frontier without an ecosystem is not stable, I saw the tech world focused heavily on what it meant for big companies. But as I read it through it, I thought about the impact on individual creators, freelancers, and small companies like Weinto. This article is a good warning for everyone, big or small.

Nadella talks about a distinction between Human Capital (our judgment, taste, and relationships) versus Token Capital (the proprietary AI workflows, context, and memory we build). He warns that simply renting a frontier model like GPT, Gemini, Opus is a dead end. Models are becoming commodities. True, durable intellectual property sits in the learning loop. The compounding cycle where human feedback refines AI systems, and those systems retain the organizational memory.

For freelancers or small businesses, ignoring this means falling into the "wrapper trap". In other words, wrapping a business value around someone else's model, instead of wrapping an AI model safely inside our own proprietary system.. Those who just use AI as a faster typewriter without a loop are just offloading execution and learning. From my personal experience, this is the best way to become entirely replaceable.

Then I saw a perfect connexion with my article "Curiosity and intelligence: The case for M-Shaped professionals" where I shared my thoughts on curiosity, learning and leveraging its practice.

The connexion between curiosity, intelligence and the Learning Loop

In my article, I exaplained that traditional "I-shaped" vertical experts are at risk. Under structural technological shifts, professionals often suffer from a high need for cognitive closure, retreating deeper and deeper into their silos. It is fair to say that the future belongs to M-Shaped professionals or at least the curious ones. Why? Because these profiles combine boundary spanning, metacognition, and prospective capability.

When we connect this to Nadella's learning loop, becoming multi-disciplinary is how you build Token Capital. This isn't trivial but the goal is to turn raw human curiosity, exploration, and judgement into a structured operational framework.

  • Human Capital --> (Curiosity & structured workflow) --> Systemic documentation
  • Systemic documentation --> (Prompt Chaining) --> Token Capital
  • Token Capital --> (Prompt Chaining) --> Compounded moat
  • Compounded moat --> (Feedback loop) --> Human Capital

This has to be everyone's task, not an easy one but critical, to document the vocabulary, the rules, the dependencies, the steps and the exceptions.

Documentation is already undervalued in small businesses but the gap will grow if they don't produces clean, structured knowledge that serves two purposes:

  1. Feeds the team (Human Capital): Preserve the organizational playbooks, so the organisation's memory isn't locked in individuals' heads.
  2. Feeds the AI (Token Capital): Provide the exact context, custom knowledge bases, and ways to evaluate and improve outputs for the AI to execute autonomously.

By modeling the curiosity and learning architecturally, we ensure that every manual intervention or creative pivot translates into a reusable context and an improved system. We turn raw human judgment into a the "hill climbing machine" described by Nadella.

The human factor to avoid wasting resources

Building a learning loop is exciting and, to me, a stimulating intellectual exercise on its own already. But it must serve the business and the humans involved. Otherwise, there is no point on spending resources at it. We risk what many builders focus on, going fast but blindly toward appealing outputs rather than outcomes.

In an AI-native world, it's is becoming easier and easier to fall into this AI wrapper trap where we start celebrating how many prompts we have, how many skills we copy, how many token are burnt. But there is no serious and professional automated pipelines without moving a single business metric. Even if the work done is beautiful, we're here for value, not for a hobby.

To ensure the learning loop translates into real value, we must organize the execution withing 3 levels of work: Strategic, Operational, and Tactical.

  1. Strategic work: Vision and agnostic architecture.
  2. Operational work: The Learning Loop (traces, framework like TOGAF, data, continuous improvement).
  3. Tactical work: Model-assisted Execution and experimentation.

1. Strategic work

At that level, the focus is on creating a sovereign Token Capital, the creation of intellactuall property that cannot be copy/pasted by competitors.

That means ensuring that the business's core intelligence lives within its own data and architecture, not inside external models. It means designing systems so we can swap out the engine (a model for example) without loosing the organisation's memory.

2. Operational Work

Here, the focus is on connecting long-term strategy to daily execution through repeatable workflows.

This is where the TOGAF framework can be implemented or used as an inspiration for small teams. Architect and build the infrastructure, the DBs, the MDs and any other artifacts repositories, the automated feedback loops that keeps and traces human work. Every task performed is captured to reinforce the learning environment. This operational work ensures that when a tactical task succeeds or fails, the signal is automatically kept and used to upgrade the system.

This is when metacognition is an asset. Being able to audit one's self work and trace it back to the architectural decisions that enabled or constrained it. It's the difference between being a thinker or a doer. Along the way, it becomes the difference between being a wrapper at risk of an underlying technology or being in control of the business.

3. Tactical Work

Now how to do at the task level? Think short-term daily actions, feature generation, and rapid experimentation.

This is the domain of execution, heavily offloaded to AI models. The AI handles the bulk processing, code generation and other draft of our documents. The human should act as a supervisor with his judgement, context, and domain knowledge to guide the AI. The AI delivers the output, the human still thinks outcome at his level to craft and refine the AI loop. This is where every manual edit made is treated as a mini-experiment to train the larger "operational system".

How to stay on track?

We have the "Why" of why a business of any size will be called to implement the learning loop and sovereign system around AI Token Capital while keeping Human Capital as a driver. We also thought about a framework of execution. Now come the measurement of success. It's always good to be more productive or to increase capital but to what end? How do we know it serves the business and the humans involved?

Not measuring a process leads to broken systems. Thus, not being able to measure the learning loop means running a lottery, not a system. Especially in a context of fast and highly contributive changes. Objectives and Key Results cannot be exported quarterly as lagging indicators. They must be measurable rapidly, continuously I'd say, so the loop can be adapted in real time.

Here is the operational OKR framework I use to monitor whether my system (or Weinto's) is actually serving humans and compounding capital:

1. Internal value

Help the team accept AI as a tool and not as a threat while allowing everyone to think more broadly. The goal is to make sure that the team isn't burning out but instead building their own horizontal work capabilities.

  • KR1: Reduce time spent on low-cognitive, operational tasks by x%, while maintaining the current work hours in a week.
  • KR2: 100% of critical project post-mortems and client feedback cycles are systematically ingested into the private knowledge base within x hours of project close.

2. External value

Clients don't care if we use AI or not. They care about the value we provide. If the system is working, the learning loop should compress delivery time, yes. That is where charging work by hour or by day doesn't make sense anymore. Client will give less value to time and more to whom has the best systems (AI, human or else) for their needs. Delivering faster and tailored experiences is hard, often due to a lack of time. The Token Capital can be leveraged to flip that constraint.

  • KR1: Increase the incrementatal scope acceptance rate by x%. In other words, how team can pivot on features and other requests wille keeping control over the delivery deadline.
  • KR2: Reduce the time to a working proof-of-concept by x%.
  • KR3: Increase the post delivery satisfaction score by x on AI-assisted outputs / outcomes. It is important to make sure that speed does not lower the quality and value provided.

3. Structural value

We've talked about a business being wrapper around AI or AI being a wrapper around the business. We want the latter. So we need to ensure that we're building toward that goal while remaining sovereign.

The goal is to prove that the business has real value that exists independent of people or external AI providers. In other word, is it real or just a belief?

  • KR1: Migrate a core business workflow from one foundation model vendor to another with 0% loss in terms of context, accuracy and results.
  • KR2: Decrease by x% the human input required for repetitive complex workflow over 2 quarters. The loop must grow with its users.
  • KR3: 100% of critical work and edge-cases resolutions are documented and available to the system. No knowledge loss should occur, even when a freelancer is offboarded.

Observations

In theory, that approach implies teaching systems that will need less human intervention over time. As an engineer, a system (OS, software, application, etc...) could never replace my job. It's a system like a car is a system. Yes, we need less horses. Did transportation disappear? No. It increased instead !

In that regard, the illusion that deep technical specialization or legacy processes will shield professionals from the AI is wrong. This logic is dangerously blind to the economic reality hitting us right now. This is like a horse rider specializing in breeding an elite breed because "a car will never be able to"...

In complex engineering environments, scoping a problem is hard but not every business is a large company with hundreds or thousands of employes. True. Most small businesses don't run with 50 pages technical specifications. They sell days of human labor wrapped in some creative hype, relationship-building, or worse, emotional alignment with the client. Nothing that an AI system won't be able to provide. The difference is that we're only at the beginning.

That said, I don't see firing juniors or over-relying on seniors is a sustainable strategy on the long term. Businesses serve humans, even if they're not aware of it. So human should be part of the picture. They should be trained to use and build their systems. Who will build the AI loop and adapt system to human clients' needs if only AI driving?

AI is not just a tool. It is a structural change, a new paradigm. New architectures are emerging and each business should build its own systems. And don't believe those saysing "the level of expertise is not matched", they often fall short at completing the sentense by "yet". This mindset grants only a few months, if not lost already.

Conclusion

For decades, freelancers and experts (myself included) have used the "daily rate" as a metric to bill clients. I've never really enjoyed it because it meant I was rewarded for being slow and punished for being fast. With AI, that equation gets flipped, this is good. My competitive advantage should remain my capacity to think and increase value.

Don't spend your energy trying to protect a legacy workflow or comforting yourself with the illusion that "the AI still needs me to click the button". Widen your own scope and horizon, build an ecosystem for yourself or contribute to a learning loops.