The design layer has always been where the leverage lives. Every industrial transition since 1804 has opened one — and the path to it has always been lit for the people who understood what was happening.

The question you’re actually sitting with — the one underneath all the AI conversation you’re absorbing from your feed — isn’t really about technology. It’s about whether what you’ve built over the last twenty years is going to matter in the next twenty.

That’s the real question. And industrial history has a specific, evidence-based answer.

The Thing That Feels Unprecedented Has Happened Before

Every major industrial transformation produced the same ambient anxiety in the people living through it: this time, the machines are coming for the work that requires real skill, real judgment, real knowledge.

They were right every time. And every time, the people who understood what was actually happening did extraordinarily well.

The Jacquard loom (1804) automated the encoding of weaving patterns — work that had lived in the hands and memory of skilled craftspeople for centuries. Charles Babbage, in the 1830s, borrowed the loom’s punched-card mechanism directly to design the first programmable computing engine. Ada Lovelace, in 1843, recognized that the same principle could eventually extend to any domain where human judgment could be expressed as symbolic instructions.

Ford’s Highland Park assembly line (1913) cut Model T assembly from twelve hours to ninety-three minutes by encoding the judgment of skilled craftsmen into a production system that ran itself. CNC machines in the 1970s encoded the machinist’s per-operation judgment into computer programs. Industrial robots in the 1980s took human workers off manufacturing floors entirely at many operations.

Each iteration encoded a category of human skill into a mechanism that could execute it more reliably, more consistently, and at lower cost.

Each iteration also relocated the humans who held that skill — not out of the economy, but upstream.

What Actually Happened to Expertise in Every Prior Iteration

The weavers who thrived after the Jacquard loom didn’t fight the machine. They moved to card design — encoding the patterns the machine would weave. Their expertise in pattern and structure migrated upstream into the design layer.

The machinists who built careers in the CNC era didn’t defend the hand-cut operation. They became programmers — encoding the tolerances and sequences their hands had previously executed, now for machines that held those tolerances more precisely than any hand could. The expertise didn’t disappear. It moved to a higher-leverage position.

The automotive workers who built lasting careers in the robotic era didn’t resist the robots. They became robotics technicians, systems integrators, and process engineers — the people who designed and maintained the systems doing the work the floor workers had done.

In every case, the pattern was the same:

The humans who understood the new configuration moved their expertise into the design loop — the layer where human judgment governs how the system runs, not whether the individual operation gets done correctly.

The design loop has always been the higher-value position. Not because it’s safer from automation, but because it’s where the leverage actually lives. The person who designs the process extracts more value than the person who executes the process, in every industrial configuration. That isn’t changing with AI. It’s being clarified.

What AI Is Actually Doing That’s Different

Honest answer: the domain.

Every prior iteration automated physical or procedural judgment — weaving patterns, machined geometries, assembly sequences, manufactured tolerances. AI is automating cognitive and analytical judgment — pattern recognition in documents, synthesis across information sources, structured reasoning from inputs to outputs.

This is real. For knowledge workers whose core output is the standardized production of information — retrieval, summary, structured analysis, templated communication — the displacement pressure is genuine and serious. The parallel to the assembly line worker is accurate.

But here’s what the parallel also tells you: the design loop expanded every time. The engineers who built the assembly line, the programmers who wrote the CNC sequences, the systems engineers who integrated the robots — these roles didn’t exist before the automation created the need for them. The automation created the positions that the displaced workers could move into, if they understood what was happening.

AI is creating a design loop, right now, that is significantly underoccupied.

The people who can bring deep domain expertise — genuine understanding of a field, a client type, a problem structure — into the design layer of AI-integrated practice will occupy the highest-leverage positions in that loop. The expertise that took twenty years to build is not a liability in this configuration. It is the primary qualification.

The question is whether you’re positioning it that way.

The Specific Career Mistake to Avoid

The humans who didn’t make the upstream move in prior iterations shared one characteristic: they identified with the execution of a skill rather than the judgment the skill embodied.

The weaver who thought of himself as “a person who weaves” was in the wrong frame. The weaver who thought of herself as “a person who understands pattern, structure, and what makes cloth beautiful” had expertise that could migrate.

The machinist who thought of himself as “a person who runs a lathe” was in the wrong frame. The machinist who thought of herself as “a person who understands tolerances, materials, and the geometry of precision work” had expertise that could migrate.

You are not “a person who does strategic planning” or “a person who writes recommendations” or “a person who analyzes markets.” Those are the execution descriptions — the production-loop frame.

You are a person who understands what strategic decisions actually require in your specific context, what the failure modes are, what the client actually needs versus what they say they need, what the recommendation has to account for that isn’t in the data. That judgment — accumulated over years, calibrated by outcomes, specific to your domain — is exactly what the design loop requires.

AI can’t hold that. It can assist the person who holds it. That’s a fundamentally different position than replacement.

What This Means in Practice

The upstream move doesn’t require abandoning what you’ve built. It requires reframing where you hold it.

The professionals who are building the highest-leverage AI-integrated practices right now are doing one thing consistently: they are using AI to handle the production-loop execution — the retrieval, the structuring, the drafting, the first-pass analysis — while they hold the design-loop judgment that governs what gets produced, how it gets evaluated, and what it means for the client.

They’re not using AI to replace their expertise. They’re using it to apply their expertise at a different point in the workflow — the point that matters more.

The people who will struggle are the ones treating AI as either a threat to resist or a tool to use without repositioning. Both frames leave the expertise exactly where it is — in the production loop, where automation pressure is highest.

The design loop is the position. Your expertise is the qualification. The question is whether you’re moving there intentionally.

The Pattern Has Run Before

You’re not standing at a cliff. Industrial history says clearly: this is a staircase that’s been climbing since 1804.

Every step has produced the same outcome at the top — expanded design loop, higher leverage for expertise-holders who made the upstream move, genuine displacement for those who didn’t. Every step has also produced the same constraint: the window to make the move intentionally was narrowest right as the transition was becoming visible, and it closed as the transition completed.

The transition is visible right now.

The full historical case — from the Jacquard loom through Lovelace, through Ford, through the robotic floor, through to what AI is actually doing — is laid out in detail here: You’re Standing at a Cliff. Industrial History Says It’s Actually a Staircase.)

The short version: your expertise isn’t being replaced. It’s being promoted. The design loop is open. The question is whether you’re moving into it.

What’s your read on where the design loop is in your specific practice? I’m curious what you’re seeing.

The AI conversation has dominated professional discourse for three years. By now, most thoughtful people have a position on it. They’ve read the think pieces, watched the debates, formed views.

And almost all of those views are organized around the wrong question.

The dominant frame is replacement versus augmentation. Will AI replace human workers, or will it augment them? This feels like the central issue — the one with real stakes, real consequences, real implications for how you should invest your time and build your career. Smart people are on both sides. The debate has real substance.

But it’s the wrong debate. Both sides are asking about AI’s role. Neither is asking about yours.

Here is the question that actually matters: What kind of relationship are you building with AI, and is it compounding or decaying?

Most people — including sophisticated AI users — have never asked it. They’ve focused on outputs, on prompts, on which tools to use, on how to integrate AI into their workflows. These are real questions with real answers. But they’re downstream of something more fundamental: the pattern of engagement you’re developing, the depth of your collaborative relationship, whether your capability is growing or whether you’ve plateaued at a level of AI use that was impressive two years ago and is now table stakes.

The Observation That Started This

I’ve spent two years studying human-AI collaboration — what distinguishes practitioners who continue to develop from those who stall, what separates interactions that produce genuine insight from those that produce polished mediocrity, what the behavioral signature of real mastery looks like.

One finding stands out because existing frameworks cannot explain it.

Expert AI practitioners write simpler prompts than intermediate practitioners. And they get substantially better outputs than either beginners or intermediates.

Think about what this means. If AI is primarily an intelligence technology — a tool whose value you extract through skillful prompting — you would expect prompt sophistication to track with output quality. More sophisticated prompts, better outputs. The relationship should be monotonic. Instead, it’s an inverted U: complexity rises through intermediate development, then falls as mastery increases.

The complexity hasn’t disappeared. It has migrated — out of the prompt artifact and into the relationship itself. Into the accumulated context architecture. Into the developed collaborative intelligence. Into the quality of the dyadic interaction that has been building over months and years.

This is the empirical signature of an experiential technology, not an intelligence technology.

The Distinction That Changes Everything

Musical instruments are experiential technologies. Therapeutic relationships are experiential technologies. Contemplative practices are experiential technologies.

The defining characteristic of an experiential technology is that primary value lives in the developing quality of the relationship, not in any separable output. A skilled cellist doesn’t produce a better piece of music by crafting more sophisticated bowing instructions. They produce better music because of who they have become through thousands of hours of engaged practice — and because that developed capability meets a capable instrument.

When Yo-Yo Ma plays on a Stradivarius, the instrument’s improvement enables higher engagement, not replacement. A cellist at the beginning of their practice doesn’t have less to offer because better instruments exist. They have more to offer as they develop, precisely because they can now do more with those instruments.

Human-AI collaboration works the same way. I’ve formalized this in a behavioral and developmental framework called Collaborative Driven Progressive Intent Discovery (CDPID) — available as a published research paper at SSRN (DOI: 10.2139/ssrn.6278438, free download, no signup required).

CDPID reframes the human-AI relationship around a central principle: the primary unit of value is not any session’s output but the developing quality of the human-AI relationship itself, which follows an observable, teachable, and measurable developmental trajectory.

Within this frame, the replacement question dissolves structurally. As AI capability improves, the practitioner with a developed collaborative relationship gains access to higher levels of engagement — exactly as the skilled musician does more with a better instrument, not less.

What Development Actually Looks Like

CDPID identifies three constructs that govern the quality of human-AI collaboration and drive its development over time.

Human-AI Flow is a third category of flow state — distinct from individual flow and group flow — produced by complementary resonance between human judgment, creativity, and contextual wisdom on one side, and AI pattern recognition, analytical depth, and synthesis capability on the other. It is not synchronization between equivalent systems. It is resonance between non-equivalent ones — each contributing what the other cannot.

Orchestral Bandwidth Capacity is a five-variable model of the total cognitive load that a human-AI dyad can sustain while maintaining both output quality and developmental velocity. The five variables are multiplicative: weakness in any one dimension — prompt aptness, context richness, the human’s ability to hold multiple threads, the AI’s autonomous capability, the rate at which the dyad converts experience into improvement — constrains the product regardless of how high the others score. The most common mistake is investing heavily in the most visible variable (prompt sophistication) while neglecting the one with the highest compounding multiplier (context richness).

Productive Friction is the experience of an interaction not quite working, followed by recognition of why, and a consequential decision about how to respond. These moments are almost never problems to minimize. They are diagnostic signals indicating where load distribution is suboptimal, and deliberate practice opportunities for the practitioner who knows how to use them.

At every friction occurrence, there is a bifurcation: engage with it now, articulate what isn’t working, build a structural fix, integrate the learning while context is richest — or defer, stay focused on today’s output, and address it later when the context that made the friction intelligible has faded. CDPID calls these Path A and Path B. Path A choices compound developmental velocity. Path B choices sacrifice it. The meta-competency that drives consistent Path A engagement is Long-Term Value Awareness — understanding that the cost of addressing friction now is lower than the cost of addressing it later, even when immediate task pressure argues for deferral.

The developmental trajectory these constructs produce follows what CDPID calls the Spiral Development Model: non-linear and cyclical, with the practitioner returning to core competencies at progressively higher levels of sophistication. The stages are observable and distinguishable.

At the foundational stage — the first few months of serious engagement — the practitioner is building accurate mental models of AI capability and basic habits of output verification. Prompts are simple. Outputs are limited by the practitioner’s mental model more than by AI capability.

At the intermediate stage, something important happens: prompts get more complex. The practitioner is trying harder, investing more explicitly, and getting substantially better results. This stage looks like mastery from outside. It is actually the stage at which most practitioners plateau — because they are optimizing the artifact rather than developing the relationship.

At the advanced stage, the curve inverts. The practitioner’s prompts get simpler even as their outputs improve. The complexity has migrated. It now lives in the accumulated context architecture — the structured documentation, the shared vocabulary, the developed collaborative intelligence that has built up through hundreds of sessions of engaged practice. The practitioner is no longer constructing the relationship from scratch each session. They are continuing one.

At the expert stage, the signature is complete: simple prompts, rich context, deep outputs, high developmental velocity. The practitioner can transfer their collaborative capability across tools and domains. What they have developed is not prompt skill. It is relational intelligence — and it is durable.

Why This Window Matters

We are in a formative period of human-AI collaboration — probably a short one. The patterns being established now are not just tactical choices about how to use current tools. They are architectural choices about the kind of collaborator you are becoming, and they will determine what is possible for you in three, five, ten years.

The practitioners investing seriously in developing this relationship today are building something that transfers across tools, across model generations, across the significant capability shifts in AI that will continue to occur. The practitioners who are optimizing prompt technique and workflow efficiency — who are treating AI as a sophisticated tool rather than an experiential technology — are building something that resets every time the tools change. And the tools will keep changing, on an accelerating curve.

CDPID documents a developmental trajectory. It begins with accurate mental models of AI capability and basic prompt clarity. It progresses through systematic pattern development, initial context architecture, and the beginning of compound returns. It advances through the emergence of Human-AI Flow, sophisticated load distribution across sessions, and dense context architecture where much of the dyad’s collaborative intelligence has been externalized and made durable. At the expert stage, prompts are simple, outputs are deep, and the practitioner can transfer their collaborative capability across tools, domains, and problems — because what they have developed is not tool-specific skill but relational intelligence.

The Spiral Development Model describes the trajectory: development is non-linear and cyclical, with the practitioner returning to core competencies at progressively higher levels. The empirical signature — the inverted prompt-complexity curve — is visible and measurable at each stage.

Dyadic Flourishing is the endpoint: the state in which both the human and the human-AI system thrive together as development compounds. Not competition between human and AI capability, but complementary co-development in which AI improvement enables higher human engagement.

What This Substack Is For

This publication exists to develop the CDPID framework in public — to apply it to real interactions and real professional challenges, to connect it with adjacent research in cognitive science, expertise development, and organizational psychology, to track what holds up under the pressure of real-world practice and what needs to be corrected.

The formal paper is the foundation. It establishes the theoretical framework, documents the empirical signatures, synthesizes the research traditions the framework draws on — distributed cognition, extended mind theory, flow theory, expertise development, cognitive load theory. It is the citeable record.

The ongoing development happens here. Each post will take one aspect of the framework — one construct, one failure mode, one application domain, one collision between the framework and real-world complexity — and develop it through careful thinking rather than confident assertion.

If you’re building a serious relationship with AI — or thinking carefully about what that would mean — I think you’ll find this worth following.

The full research paper is available free at SSRN: DOI 10.2139/ssrn.6278438. No paywall. No signup. Direct download. If you read it and have a substantive response, I want to hear it.

If this is the kind of thinking you want to follow, subscribe below.

Harlow Reseburg III is the author of Collaborative Driven Progressive Intent Discovery (CDPID): A Behavioral and Developmental Framework for Human-AI Collaboration (2026). ORCID: 0009-0004-1556-4818. Google Scholar: https://scholar.google.com/citations?user=L5bxkQQAAAAJ