The Reinforcement Loop

Why AI Transformation Is an Operating System Problem, Not a Technology Problem

Chuck DeVries | DeVries Technology Advisory | March 2026

The Moment We're In

We are in a great reset.

Not a correction. Not a cycle. A structural discontinuity in how organizations create value, develop talent, and make decisions.

The pace of AI capability advancement has compressed what used to be a decade of disruption into weeks. Reasoning models are 1,000x cheaper than they were 18 months ago. Agent-based software adoption is accelerating faster than any category in the history of enterprise technology. AI is no longer a technology activity confined to engineering teams. It is touching every function, every role, every decision surface in the organization.

And yet most companies are still trying to apply AI to their existing operating model, like bolting a jet engine onto a horse-drawn carriage.

That does not work, that will not work.

The organizations that will thrive in the next 3-5 years are not the ones that adopt the most AI tools. They are the ones that redesign how their organizations actually work around the capabilities AI makes possible. That redesign is what I call the Reinforcement Loop.

The Core Thesis

AI does not primarily replace work. It compresses the feedback cycle between:

Intent → Action → Feedback → Learning → Adaptation

This is the Reinforcement Loop. It is the fundamental unit of organizational performance, and AI is accelerating it by orders of magnitude.

Organizations that redesign around these faster AI-native reinforcement loops will outperform those that merely automate existing tasks. This is not AI adoption. This is operating system redesign.

The difference matters. AI adoption asks: "How do we use these tools?" Operating system redesign asks: "How does our organization actually need to work now that these capabilities exist?"

Part I: The Dirty Hands Imperative

Your References Are Off

Here is the hardest truth in this entire article.

Most senior leaders managing organizations, even those shifting to AI-driven organizations, have never personally used AI as a working tool. Not in a demo. Not in a boardroom presentation. Certainly not in any form of AI-native workflow. Chatting is not working with AI I mean as an actual, daily, hands-in-the-work operating tool. How could they? Most have been in leadership roles and out of direct contribution roles for years, or even decades, while these tools have accelerated to unforeseen capability in under three.

This creates a problem that no amount of strategy consulting can solve: their management references are wrong.

Every leader learned to manage by doing the work first, then managing others who did it. You wrote code before you managed engineers. You sold deals before you led a sales team. You analyzed data before you ran an analytics function. Your judgment about what is hard, what is easy, what takes time, and what constitutes quality was built from direct experience.

AI has fundamentally changed what is hard, what is easy, what takes time, and what constitutes quality. If you haven't experienced that change directly, you are managing from outdated references. You are, in effect, managing a fleet of cars by the principles you learned managing horses.

This is not a criticism. It's a structural observation. The technology moved faster than any leader's calendar could accommodate. But it has to be addressed, because the consequences compound.

A leader with outdated references will:

• Underestimate what a team of three people with AI can accomplish
• Overestimate the time required for knowledge-intensive work
• Misallocate budget toward headcount when the bottleneck is workflow design
• Set quarterly goals calibrated to pre-AI productivity baselines
• Fail to recognize when AI output needs human judgment versus when it doesn't

The fix is not more briefings. The fix is direct experience. Leaders need to get their hands dirty with these tools. Not once. Regularly. Not to become technical experts, but to recalibrate their intuition about what is now possible.

More than that: leaders need to push one level deeper than what the tools can comfortably do, even as that comfort zone shifts. Today's frontier capability becomes next quarter's baseline. A leader who tried ChatGPT once in 2024 and formed their mental model there is already two generations behind. The discipline is not a one-time calibration. It is a recurring practice of going just past the edge of what feels easy, because that is where the real understanding of impact lives. The context gained from that experience reframes everything: what to invest in, what to stop doing, what your team is actually capable of.

The Accountability Anchor

None of this changes two things that remain permanently human.

Someone still needs to decide what is worth doing. AI can generate options, surface data, and model scenarios at unprecedented speed, but the decision about where to invest attention, which problem matters most, which risk is acceptable, that requires judgment shaped by values, context, and accountability. No model has that.

Someone still needs to own the result. When AI-generated output fails, and it will fail, accountability cannot be diffused into "the model did it." The human in the loop is not a compliance checkbox. It is the difference between an organization that learns from failure and one that automates it.

The Reinforcement Loop model does not diminish leadership. It raises the bar. Leaders who engage directly with the tools, recalibrate their references, and maintain clear accountability will multiply the capability of everyone around them. The Multipliers research already showed that the best leaders unlock more than 100% of their team's capacity. AI extends the upper bound of that multiplier dramatically, but only for leaders who understand what they're multiplying.

Part II: The New Operating Model

AI as Operating System

The shift that most organizations haven't made is conceptual. They are still thinking about AI as a set of tools to be deployed. The more accurate frame is that AI is becoming the operating system through which work flows.

Consider how work actually moves through a modern AI-augmented workflow:

Spec → Plan → Tasks → Implement

Each of these stages is now AI-participatory. Specifications can be drafted, compared to prior art, and stress-tested against edge cases in hours rather than weeks. Plans can be decomposed into task structures that account for dependencies, resource constraints, and risk profiles. Tasks can be executed with AI augmentation that compresses what used to take a team of five a sprint into what takes one person an afternoon. Implementation can be validated, documented, and iterated on within the same working session.

This is not hypothetical. This is happening right now. Google's Vibe design tool lets a non-designer describe an interface in natural language and produce a working prototype. Replit's agent-based development environment lets a single person build and deploy a functional application in hours. Agentic coding tools are writing, testing, and debugging production software with a human providing direction rather than keystrokes. The best programming language is increasingly English, or whatever language the business thinks in. The idea-to-delivery cycle is collapsing, and with it, every assumption about how many people it takes to do meaningful work.

Three forces drive this:

1. AI removes friction work. The administrative, coordinative, and connective tissue that used to consume 40-60% of knowledge worker time, formatting documents, searching for information, synthesizing across sources, translating between audiences, scheduling, summarizing... is being compressed toward zero.
2. AI removes knowledge bottlenecks. The institutional knowledge that used to live in one person's head, accessible only through that person's availability, is increasingly surface-able through AI systems trained on organizational data.
3. Business processes themselves are beginning to self-improve. This is the force most organizations have not yet internalized. Agents and automated workflows don't just execute; they observe, iterate, and adapt. The business process is no longer a static artifact that gets redesigned every few years during a transformation initiative. It is becoming a living, evolving part of the business ecosystem, continuously optimizing itself based on outcomes. The recursive self-improvement pattern that AI researchers have studied at the model level is now manifesting at the business process level, and it changes the nature of operational management from periodic redesign to continuous stewardship.

Together, these three forces don't just make existing work faster. They make more things possible. The strategic question for every organization is: now that your capacity to act has expanded, what will you do with it?

The Data Reality

There is a prerequisite that underpins all three of these forces, and it is the one most organizations are least prepared for: data.

Data is the oil that drives the engine of AI. Every promise of AI-native workflows, self-improving processes, and compressed feedback cycles depends on data that is accessible, coherent, and trustworthy. And in most organizations, it is none of those things.

Here is what actually exists in most enterprises: people and processes that hold the data together while serving customers. The knowledge of what makes data useful, how fields relate to each other, what the values actually mean, where the gaps are, lives inside people's heads. It is the proverbial tribal knowledge. Few companies understand their own semantic data frameworks or data ontologies at a level where advanced systems can simply operate. AI is powerful, but it is not magic. It cannot reason over data it cannot find, cannot trust data that contradicts itself, and cannot build relationships across data that was never designed to connect.

This does not mean there is no hope. It means the work is real and it requires investment. Better and better AI systems will make data curation, relation mapping, and quality management progressively easier. But the destination will not be reached without people driving the bus. Someone has to define what the data means, where the authoritative sources live, and how the pieces fit together. That work will be hard, and the current state is often worse than senior leaders believe.

There is a silver lining here, and it connects directly to the Friction Paradox. The work of curating, connecting, and modernizing an organization's data is precisely the kind of purposeful difficulty that builds deep understanding of the business. It is a smart target for early AI implementations: use the tools to move and improve the data, reinforcing the systems as they grow, turning historical process and bureaucracy into new AI-native and accelerated capability. The journey from messy data to coherent data is not just a technical prerequisite. It is a learning path, and one of the highest-leverage places to start.

The Two-in-a-Box Model

The organizational unit that makes this work is not a traditional reporting structure. It's a partnership: business domain expertise and technical AI capability, working together, managing a portfolio of agents and automated workflows.

This resembles the "two-in-a-box" model that already works in some product organizations, a business leader and a technical leader jointly accountable for outcomes. In an AI-native operating model, this pairing becomes the fundamental management unit. The business leader provides domain judgment, customer context, and strategic prioritization. The technical leader provides system design, AI orchestration, and capability architecture.

Neither can function without the other. The business leader who doesn't understand AI capabilities will set the wrong priorities. The technical leader who doesn't understand the business domain will build the wrong systems. Both need the exposure and experience of the other to drive business success.

The Air Traffic Control View

As AI automation scales, the human role shifts from executor to supervisor. The right mental model is not a manager overseeing a team of people. It's closer to an air traffic controller: monitoring multiple automated systems, intervening when conditions change, escalating when patterns diverge from expectations, and making judgment calls that no automated system is equipped to make.

This requires a different kind of attention. Not the deep focus of doing the work yourself. Not the periodic check-ins of traditional management. A continuous, pattern-aware supervisory attention, what you might call operational vigilance.

But here's the problem: we don't have the interface layer for this yet. A whole new class of tools is desperately needed, one that allows humans to scale their attention across an expanding set of AI capabilities, volumes of information, and speeds of execution. The tools need to elevate what matters, suppress what doesn't, and separate signal from noise in real time. Without that layer, the "human in the loop" is not a supervisor. It's a designated blame absorber with no real ability to intervene meaningfully. Nobody signed up for that job, and no organization should design for it.

Organizations will need to build for this deliberately. Alert systems that surface the right signals. Dashboards that show system health without drowning the operator in noise. Escalation paths that are fast enough to matter. The organizations that get this right will operate at a speed and quality level that seems impossible from the outside. The ones that don't will discover that unsupervised automation creates problems faster than it solves them.

Part III: The Career Lattice

From T-Shaped to Something New

For two decades, we've talked about T-shaped skills: deep expertise in one domain, broad familiarity across adjacent domains. That model is breaking.

When AI can provide "deep enough" expertise in any domain within minutes, the competitive advantage of deep human specialization narrows. Not to zero; domain knowledge remains fundamental, and the radiology example is instructive: AI can read the scan, but reading the scan was only ever one part of what a radiologist does. The judgment, the patient context, the integration with treatment planning, that remains human. This is why for many years now "radiologists won't be needed" but they continue to be. A piece of the job is not the job.

But the shape of valuable human capability is changing. The T is becoming something more like a sideways E, or an F, or a Z, or shapes we don't have names for yet because the roles they describe don't fully exist yet. What they share is this: multiple areas of working-depth competency, connected by an adaptive core that can learn new domains fast enough to stay relevant as AI capability expands.

The career lattice replaces the career ladder. Lateral movement, learning adjacent domains, building new competencies, applying existing judgment in unfamiliar contexts becomes more valuable than vertical promotion into deeper specialization.

The Friction Paradox

Here is one of the most important design tensions in the entire framework.

AI removes friction from work. That's a feature. But human learning requires friction. The struggle of doing something difficult, the productive failure of getting it wrong, the slow accumulation of pattern recognition through direct experience, these are the mechanisms by which humans develop capability.

If we remove all friction from work, we accelerate output but starve development.

The Reinforcement Loop model has to account for this deliberately. Organizations need to design friction into their talent development pathways. Not artificial busywork. Purposeful difficulty: stretch assignments, cross-domain rotations, projects where AI assistance is deliberately constrained so that the human builds the muscle.

This is counterintuitive in a world optimizing for speed. It requires leaders who understand that long-term organizational capability depends on short-term development investment. The career lattice requires friction to enable learning, even as the operating model requires reducing friction to enable performance.

The Skills That Matter

When AI handles the knowledge retrieval, the pattern matching, the content generation, and the analytical heavy lifting, what's left for humans?

Four things: courage, curiosity, creativity, and grit.

Courage to make decisions with incomplete information and own the result. Curiosity to explore domains you don't yet understand and ask questions that haven't been asked. Creativity to see connections that no training data contains. And grit to persist through the inevitable failures that come with operating at the frontier of what's possible.

These are not soft skills. They are the hard skills of an AI-native economy. They cannot be trained by a model. They can only be developed through experience, and they become more valuable as AI commoditizes everything else.

Part IV: The Permission Architecture

Governance That Enables

Every organization that has tried to deploy AI at scale has run into the governance question: how do you move fast without breaking things that matter?

Most governance frameworks answer this by slowing everything down. They create review boards, approval chains, and compliance checkpoints that ensure nothing risky happens, and, as a side effect, ensure that nothing much happens at all.

The Reinforcement Loop requires a different approach: a Permission Architecture that creates graduated lanes for different levels of risk.

Permission to Learn: Open but bounded. Anyone in the organization can experiment with AI tools on non-sensitive data. The goal is literacy, not production output. Failures are expected and encouraged. Documentation is lightweight. The boundary is: no production data, no customer-facing output, no regulated processes. Everything else is fair game.

Permission to Build: Declared and standardized. Teams building AI-enabled workflows declare their data sources, patterns, and tools. The organization provides approved architectures, vetted model endpoints, and standardized integration patterns. This accelerates development by replacing ad-hoc decisions with proven components. The boundary is: use the sanctioned building blocks, document your choices, and submit for review before promotion to production.

Permission to Operate: Validated and auditable. Production AI systems undergo security review, output validation, bias assessment, and compliance verification. Monitoring is continuous. Audit logs feed traceability. The boundary is: you can move fast, but the system must be explainable, monitorable, and reversible.

Audit and Traceability: Bring the auditors along. Don't build the system and then figure out how to explain it. Involve compliance, legal, and risk teams from the Permission to Build stage, heck, make them first class learners in the Permission to Learn stage. Feed validation tracing into continuous monitoring. Make the audit trail a feature of the system, not an afterthought.

This architecture lets organizations move at three different speeds simultaneously: fast for learning, measured for building, rigorous for operating. Most organizations today are stuck at one speed (usually slow) because they haven't differentiated the risk levels.

Culture, Not Dictum

One critical caveat: this permission structure cannot be imposed as a top-down policy and expected to work. Safety, experimentation, and learning need to be cultivated as culture, not stated as leadership dictum.

The difference is lived experience. When a team sees that their failed experiment was treated as useful learning rather than a career risk, they experiment again. When they see that their successful automation was recognized and shared, they build more. Culture is not what you say you value. It's what happens when things go wrong and what gets celebrated when things go right.

The organizations that will lead in AI adoption are the ones where psychological safety is genuine, not performative. That requires leaders who model the behavior, including the willingness to fail publicly, to say "I don't know," and to demonstrate that learning is valued above certainty.

Part V: The Economic Reframing

First Principles Still Apply

Amidst all the disruption, one thing hasn't changed: value creation still follows first principles. Someone has a problem. Someone provides a solution. The solution is worth more than the cost to deliver it. Everything else is mechanism.

What AI changes is the mechanism. The cost to deliver solutions is dropping dramatically. The speed of delivery is increasing. The range of problems that can be addressed is expanding. But the human judgment about which problems matter, which solutions are trustworthy, and which outcomes are worth pursuing, that remains the core of value creation.

Every company that exists started by filling a customer's need. That is still the fundamental reason any company should exist. Not to build a better spreadsheet, a sharper report, or a slicker interface. Companies exist because they solve a problem for a customer. In this moment, that has to be the first point of focus: be willing to drop the bureaucracy that once gave consistency in exchange for new AI-native speed and capability to serve that customer need. The organizations that cling to their processes for the sake of process will be outrun by those who remember what the process was supposed to accomplish.

The Largest Advisory Opportunity in a Generation

Every company, in every industry, is now facing the same question: how do we actually do this? Not "should we use AI?" That question is settled. But "how do we redesign our organization, our workflows, our talent model, our governance, and our economics around AI capabilities?"

That is not a technology question. It is a systems question, a leadership question, and an organizational design question. And very few organizations have anyone internally equipped to answer it, because the people who understand AI don't understand the business deeply enough, and the people who understand the business haven't yet experienced AI deeply enough to know what's possible.

This gap, between AI capability and organizational readiness, is the advisory opportunity. It spans every sector: healthcare, financial services, manufacturing, government, education, professional services. It will persist for years, because the technology will keep advancing faster than organizations can adapt.

The organizations that figure it out first will have an asymmetric advantage. The advisors who help them figure it out will be operating at the highest leverage point in the market.

The Design Tensions

The Reinforcement Loop is not a prescriptive framework. It is a set of design principles that must be adapted to each organization's context. What makes that adaptation complex, and interesting, is that several core tensions run through the model. These tensions are not problems to be solved. They are spectrums to be navigated, and where an organization sits on each spectrum will determine which elements of the framework matter most.

Tension 1: Friction vs. Speed

AI removes friction from execution, but human development requires friction for learning. Organizations must design for both simultaneously; fast for production, deliberately difficult for development.

Tension 2: Depth vs. Breadth

Deep domain expertise remains valuable, but broad adaptive capability is becoming more valuable. The right balance depends on the organization's competitive position, industry dynamics, and talent model.

Tension 3: Autonomy vs. Accountability

AI democratizes access to knowledge and capability, enabling more people to act independently. But someone still needs to own the outcome. Organizations must distribute agency while maintaining clear accountability lines.

Tension 4: Abundance vs. Scarcity

AI is creating capability abundance in domains that were historically scarce. Our economic models, management practices, and incentive structures were all designed to manage scarcity. The transition to abundance-based thinking is not obvious and will not happen automatically.

Each of these tensions creates a diagnostic axis. Where does this organization currently sit? Where does it need to move? What's blocking the movement? These are the questions that make the Reinforcement Loop actionable in practice.

How to Start

For leaders reading this and wondering what to do Monday morning, three starting points:

Get in the tools. Have your leadership team spend one week using AI tools in their actual daily work, not in a sandbox, not in a training session. Real work, real problems, real deadlines. Then debrief together. The conversation that follows will be more valuable than any strategy offsite, because it will be grounded in direct experience rather than abstraction.

Map your highest-friction workflows. Identify the three processes in your organization that consume the most human coordination time relative to the value they produce. Pick one and redesign it AI-first. Not "add AI to the existing process." Redesign from the outcome backward. What would this workflow look like if you were building it today with no legacy constraints?

Establish your Permission to Learn boundaries and announce them. Tell your organization explicitly: here is where you can experiment, here is what's off-limits, and here is what we expect you to learn. Most organizations have neither given permission nor set boundaries, which means people are either experimenting in secret or not experimenting at all. Both are worse than a declared, bounded space for learning.

These are not the full framework. They are the first loop of the Reinforcement Loop: act, observe what happens, learn from it, adapt.

What Comes Next

The future is not shaped by certainty. It is shaped by the choices we make today.

The organizations that engage with AI as an operating system redesign, not just a tool deployment, will set the pace for their industries. The leaders who get their hands dirty with the tools, who recalibrate their references, who build permission structures that enable rather than constrain, they will build the organizations that define the next era.

The career lattice will reward courage over credentials. The permission architecture will separate the organizations that move from the ones that talk about moving. The design tensions will reveal where each organization needs to focus.

This is not a moment to wait for clarity. The clarity comes from doing the work.

The Reinforcement Loop is a working framework developed by Chuck DeVries, drawing on over two decades of enterprise technology leadership, AI transformation at scale, and direct experience bridging strategy and execution in regulated industries. His work in AI spans from graduate research in the late 1990s through leading large-scale enterprise AI programs recognized by Gartner. The framework is informed by real advisory engagements, Gartner AI Executive Advisory Board participation, and ongoing conversations with PE firms, consulting practices, and technology companies navigating the same transition.