We Have Our Electricity. It's Time to Redesign Our Factories.

A Framework We Didn't Write — But Recognize

Last week, George Sivulka, CEO of Hebbia, published a piece through a16z titled "Institutional AI vs Individual AI: Where did the productivity go?" His core argument: AI has made every individual dramatically more productive. No company became dramatically more valuable as a result.

He draws a parallel to the electrification of textile mills in the 1890s. Factories installed electric motors to replace steam engines. For thirty years, output stayed flat. The technology was superior, but the organizational structure hadn't changed. It wasn't until the 1920s, when factories redesigned their entire floor layout around electric unit drives, that electrification actually delivered returns.

The lesson: technology alone doesn't drive transformation. Institutional redesign does.

When our team read this, we didn't learn something new. We recognized something we've been living. We sat down, mapped his seven pillars against our architecture, and realized we are ahead on five, aligned on one, and behind on one. What follows is the honest result of that mapping.

The Gap We've Been Working In

At 8Hats Lab, we work with founders and executives who are technically sophisticated — many of them building AI products — but operationally conventional. They use Claude or ChatGPT daily. They can explain transformer architectures. And yet their company still runs on the same workflows, decision patterns, and coordination habits they had before AI existed.

We call this the Chasm — the gap between AI Proficiency Level 2 (templates and prompts) and Level 4 (the system works for you). Sivulka describes the same gap, just at the institutional scale: the distance between "we have AI tools" and "our organization operates differently because of AI."

These are not different problems. They are the same problem at different zoom levels. The founder who hasn't crossed the Chasm personally cannot redesign their organization around AI. Individual transformation is the prerequisite for institutional transformation.

Mapping Seven Pillars to Our Architecture

Sivulka outlines seven pillars of institutional intelligence. Here is how they map to what we've been building, from both the founder's perspective and the architect's.

Coordination

Sivulka calls for an entire new industry of "Agentic Management" — defining agent roles, communication protocols between agents and between agents and humans, and measuring the value each agent creates. Without this, he writes, you get "thousands of agents rowing in different directions."

We have been building exactly this layer. Our organizational model includes a formal agent registry with a four-category taxonomy (Productive, Enforcement, Coordination, Operational), access matrices that define which agents can read and write to which layers of the organization model, and a multi-phase orchestration protocol that routes tasks from discovery through execution, review, and human sign-off. Every agent declares its inputs and outputs. Every artifact has a verification workflow.

At the coaching level, this starts with the founder. The 8Hats approach builds the founder's systematic AI workflows first, then designs the path for the team. The Digital Team — four specialized agents built around the founder — creates a coordination layer where agents share context, follow consistent standards, and produce compatible outputs.

Without coordination at both levels, you get what Sivulka describes: individual productivity gains that cancel out at the organizational level.

Signal

As AI output becomes ubiquitous, distinguishing quality from noise becomes critical. Sivulka frames this as the key challenge of the next decade: "finding signal in a mountain of exponentially increasing slop." He makes a useful distinction between deterministic agents (auditable checkpoints, predictable behavior) and nondeterministic ones (creative but unpredictable).

We built a concrete implementation of this spectrum: a reliability-creativity axis with four presets, ranging from fully deterministic (scripted behavior with minimal AI generation) through two hybrid modes to fully creative (open-ended AI generation). Each mode has explicit verification requirements. Deterministic outputs get spot-checked. Creative outputs go through structured review with pass/flag/fail criteria.

At the individual level, this maps to the Verification and Control pillar — one of the five dimensions we measure in the Express Assessment. Most founders we assess are either over-trusting AI output (accepting "looks good" without verification) or over-controlling it (reviewing every character, burning hours). Neither approach scales. The fix is calibrated trust — knowing when to verify deeply and when a spot-check is sufficient.

Bias

Sivulka argues that RLHF-trained models are too agreeable. They reinforce user beliefs rather than challenging them. His strongest line: "Organizations rarely fall from a lack of confidence. They fall because nobody can say no."

Our system is designed to say no. We have multiple layers of automated review that can block outputs, require justification for overrides, and flag substantive disagreements. Our coaching agent is deliberately designed to push back — Phase 2 introduces reflective questions and trust calibration exercises, Phase 3 uses Socratic method. The agent does not just help. It challenges your reasoning.

But Sivulka raises a point that pushed our thinking further. The problem is not only catching errors in finished artifacts. It is catching the moment when an AI agrees with a bad idea during generation. Our review layers work post-factum. They catch errors after the artifact exists. The conformity bias Sivulka describes happens earlier — when the AI accommodates your direction instead of pushing back on the direction itself.

We are now adding a conformity signal to our agent framework: the agent marks not only its confidence level, but also the degree to which its response was accommodating versus conviction-based. "I agreed, but this was more accommodation than genuine conviction." This turns the bias problem from a post-hoc check into a real-time signal.

Edge

Sivulka argues that general-purpose AI capabilities become commoditized. Lasting advantage comes from purpose-built, domain-specific systems. Even a superintelligence, he writes, would choose specialized tools.

This validates our entire architecture. Our organizational model is itself the edge — a formal ontology of concepts, an eleven-layer model of the organization, domain-specific review agents trained on specific standards and quality criteria. An agent operating within this structure produces outputs that are an order of magnitude more accurate than a generic ChatGPT conversation, because it has accumulated context that cannot be reproduced by a prompt.

The same principle applies at the individual level. The Digital Team is not a product you install. It is a set of agents configured around you — your communication style, your priorities, your verification standards. Every founder's context is different. A generic AI tool gives generic improvement. A system built around your specific context, habits, and goals creates compounding advantage that grows over time.

Domain knowledge is the real moat. Tools can be replicated. Accumulated context cannot.

Outcomes

Sivulka makes a critical distinction: most AI products sell cost-cutting (save time, reduce headcount). CEOs care about revenue growth and transformation. "Pure software is rapidly becoming uninvestable," he writes. Value accumulates in the solution layer — outcomes — not the application layer.

We are honest about this: much of our current framing has been on the "save time" side. We have talked about saving 4 hours per week, operating leaner, doing more with fewer hires. These are real results, but they are the wrong frame.

Sivulka pushed us to articulate what we have been seeing but not saying clearly enough. When a founder crosses the Chasm, they don't just save time — they see opportunities they couldn't see before. They make decisions they couldn't make before because the synthesis wasn't available. They scale into markets they couldn't reach before because their operational capacity expanded. The output of AI-native leadership is not efficiency. It is expanded capability.

We are working on a deeper mechanism for this: a process where individual learning by AI agents gets reconciled back into the organization's model of the world. The agent doesn't just learn something — it changes how the organization thinks. That is the shift from "faster" to "more capable" that Sivulka is describing.

Enablement

Sivulka highlights that organizational inertia persists even among senior leaders. He points to Palantir as the model — a company that succeeds not through better software, but through process engineering combined with change management. A bank, he notes, rejected an AI lab that didn't know what a CIM was. Domain expertise matters more than software sophistication.

This is the closest match to our entire business model. We are not a tool company. We are a behavior change company that uses AI agents as the delivery mechanism. The Human-AI Learning Architecture (HALA) is built on eleven principles of agent pedagogy — we codify the process of turning knowledge into capability, not just the tools.

A concrete example: we are running experiments on whether giving an AI agent documentation is sufficient, or whether the agent also needs a structured onboarding process — the same way a new human hire needs onboarding, not just access to the wiki. Sivulka's thesis supports our hypothesis: if process engineering matters more than software for humans, then process engineering for agents likely matters more than giving them better prompts.

Two weeks of coached practice beats twelve weeks of courses. Not because courses are bad, but because behavior change requires repetition, feedback, and accountability — not information.

Unprompted

This is Sivulka's most forward-looking pillar, and the one where our thinking goes furthest beyond his formulation.

Sivulka argues that humans are the bottleneck — prompting an AGI is like "hooking an electric motor into a power loom." He envisions agents that continuously monitor data, spot anomalies, and alert before anyone asks. An agent watching a portfolio, catching a deterioration in working capital cycles, warning about covenant breaches before the fund notices.

That is one level of "unprompted." We see three.

Level 1: Monitor. This is what Sivulka describes. Agents that watch data and alert on anomalies. We have these in our architecture — automated scans, signal agents that detect threshold breaches and escalate. Important, but reactive. The agent finds problems in existing data.

Level 2: Learner. This is where our work goes further. The Learner agent does not just watch — it actively creates new knowledge that did not exist in the organization before. It processes external content (courses, research, industry reports), maps it against the organization's model of the world, finds points of divergence, and brings back specific questions. The human does not need to know what to ask. The agent identifies that a concept from an external source contradicts or extends an assumption in the company's worldview — and prepares the question.

This is not monitoring. It is knowledge creation. The agent changes the organization's model of the world, not just alerts about deviations within it.

Level 3: Builder. The agent that takes a specification and autonomously executes to completion. Takes a product requirements document and ships working code. Takes a research brief and produces a complete analysis. The human defines what needs to be built. The agent builds it.

Sivulka sees Level 1. We are building across all three. The Learner level is, we believe, the most transformative — and the least discussed. Most conversations about autonomous agents focus on monitoring (find problems) or building (execute tasks). The idea of an agent that autonomously learns, synthesizes, and brings new understanding back to the organization is a different category entirely.

Where We Are Honest About Gaps

Sivulka's framework also exposed three gaps in our own work. We believe in stating these openly.

Gap 1: Are we redesigning or electrifying? The 1890s metaphor forces a sharp question: which parts of our system are genuinely new designs, and which are faster versions of old processes? Our divergence tracking, our organizational knowledge architecture, our multi-phase agent lifecycle — these are redesigns. But some of what we do (faster research, better summaries, automated preparation) is still electrification — doing the old thing faster. We need to keep asking ourselves this question honestly.

Gap 2: Multi-agent coordination at scale. We have strong coordination for a small team with a defined agent set. But when you scale to ten founders, each with their own agent constellation, you get a new coordination problem: how do the agents' models of the world interact? If eight out of ten agents independently flag the same divergence, that is a signal about the methodology, not just about individual organizations. We do not have this aggregation layer yet. It is on the roadmap.

Gap 3: From reactive to proactive. Our agents respond to triggers — a task arrives, a session starts, a threshold is crossed. Sivulka pushes toward something more radical: agents that initiate without any trigger. An agent that notices a gap in the organization's worldmodel has not been updated for three weeks and starts preparing materials. An agent that detects that a hypothesis flagged as "needs revisiting" has been dormant too long and schedules a review. We are moving in this direction, but we are not there yet.

Where We Go Further

Despite these gaps, our work goes further than Sivulka's framework in three dimensions that his enterprise-scale view does not address.

The human layer. Technology can be redesigned with architecture diagrams. Organizations can be restructured with new processes. But the person at the center — the founder, the executive — has to change how they think, decide, and delegate. That is a behavioral challenge, not a technical one. You cannot redesign the factory if the factory owner still thinks in terms of steam engines.

Structured disagreement. Our agents do not just execute — they track their own disagreements with the human's direction, flag them formally, and preserve the record. This goes beyond Sivulka's "say no" requirement. Our agents say no, record why they said no, and track whether the disagreement was resolved or overridden. The system remembers dissent.

Knowledge creation, not just knowledge application. Sivulka's framework assumes knowledge exists and needs to be applied better. Our Learner agents create knowledge that did not previously exist in the organization — by synthesizing external inputs against the company's specific context and surfacing insights that no one thought to ask about.

This is why our work starts with the individual. Not because individual productivity is the end goal — Sivulka is right that it isn't — but because individual transformation is the prerequisite for institutional transformation. The founder who has crossed the Chasm personally has the mental model to redesign their organization. The founder who hasn't will install electric motors into the old floor plan and wonder why nothing changed.

The Validation

Reading Sivulka's piece was validating in a specific way. Not because he confirmed our methodology — he is solving a different problem at a different scale. But because his diagnosis of the gap between individual AI and institutional AI is exactly the gap we see every day in our work with founders.

The electrification metaphor is perfect. We've been saying "tool adoption is not behavior change." Sivulka says "electrification is not factory redesign." Same insight, different language.

What his article also gave us is a sharper view of our own gaps. The conformity bias problem, the reactive-versus-proactive divide, the question of whether we are truly redesigning or just electrifying faster — these are the questions we are now working on.

We have our electricity. The tools are here. The models are extraordinary. The question is no longer "which AI should I use?" The question is: "Am I willing to redesign how I work?"

That redesign starts with you. Not your team. Not your organization. You.

And it starts with knowing where you actually stand.