Your Organisation Was Designed for Humans. That’s About to Be a Problem.

By Peter Inge, CEO & Co-Founder, insightfactory.ai

Two Perspectives That Crystallise the Moment

In the space of a few days recently, two people I have enormous respect for put out perspectives on AI and organisations that sit at completely different ends of the spectrum. Neither was responding to the other. They were each working through the same question from their own vantage point. But reading them side by side is genuinely fascinating, because the contrast illuminates something important about where we are right now and, more importantly, where the hard work actually lies.

Jack Dorsey, co-authoring with Roelof Botha, former Sequoia Capital managing partner and Block’s lead independent director, in their essay From Hierarchy to Intelligence, made the case that the entire premise of corporate hierarchy is reaching its expiry date. The org chart, they argue, was always just an information routing protocol born from the Roman legions and refined through Prussian military structures and the American railroad era. AI can now perform that coordination function better, faster, and without the latency and distortion that comes from relaying decisions through layers of management. Block, Dorsey’s payments company, cut approximately 4,000 positions, roughly 40% of its workforce, and reorganised around three roles: individual contributors, directly responsible individuals (DRIs), and player-coaches. It is a genuinely bold vision for what a company could look like when you strip away the layers that exist primarily to route information.

At the same time, the essay is honest that this is early. Dorsey and Botha wrote that Block is “in the early stages of this transition” and that “parts of it will likely break before they work.” That candour matters. Reports from current and former Block employees suggest the reality is still catching up to the vision, with roughly 95% of AI-generated code changes still requiring human modification, and AI tools not yet able to lead in regulated areas like banking. But the direction Dorsey is pointing is provocative and worth taking seriously, even if the destination is further away than the essay’s confidence might suggest.

Around the same time, Marc Andreessen shared a very different perspective on the 20VC podcast. Andreessen, whose thinking on technology and markets I’ve followed for years, argued that the labour displacement narrative is fundamentally wrong. He invoked the “lump of labour” fallacy, the economic principle that there is not a fixed amount of work in the economy and that technology historically creates more work than it destroys. In his view, most large companies are overstaffed by 25% to 75% following the pandemic hiring binge, and AI is simply the convenient justification for corrections that were overdue regardless.

Andreessen’s framing is grounded in a long-run economic truth that has held across every previous technology wave. And his observation about pandemic-era overstaffing resonates with anyone who watched headcount balloon across the tech sector between 2020 and 2022. Where I think his perspective is most valuable is in the reminder that not every layoff attributed to AI is actually caused by AI, a point that even OpenAI’s Sam Altman has echoed in cautioning against “AI washing.”

Where I find myself wanting to extend the conversation is on the transition period. The lump of labour fallacy holds over decades. But it says less about the next five years, which is where leaders need to make decisions. When Anthropic’s Dario Amodei warns that AI could eliminate 50% of entry-level white-collar jobs within five years, potentially pushing unemployment to 10–20%, the long-run economic argument doesn’t help the 24-year-old graduating into a contracting job market right now. The short-term pain of technological transitions has always been real, even when the long-term outcomes are positive. The question is how we manage the transition, not whether the destination is good.

What I find so stimulating about these two perspectives is that they are both grounded in genuine insight about how technology changes organisations, but they lead to radically different conclusions about urgency, scope, and the nature of the change ahead. The reality, I believe, lives somewhere in the space between them. And exploring that space is what this piece is about.

What the Evidence Actually Shows

The most rigorous evidence we have comes from the Stanford Digital Economy Lab. Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen analysed payroll data from ADP covering millions of American workers through mid-2025. Their findings, published in a paper titled Canaries in the Coal Mine, are striking and specific.

Employment for workers aged 22 to 25 in the most AI-exposed occupations, such as software development and customer service, declined by approximately 13% since the release of ChatGPT in late 2022. In software engineering and customer service specifically, entry-level employment fell by nearly 20%. Meanwhile, employment for older workers in the same fields held steady or grew. The divergence was sharp and concentrated.

The critical distinction, and this is the part most commentators miss, is between automation and augmentation. Where AI automates codified, repetitive tasks, the kind of “book learning” that overlaps with what universities teach and what junior workers know, employment declines. Where AI augments work, adding capability without replacing the worker, employment holds or grows.

Brynjolfsson himself noted the irony: senior coders, who have tacit knowledge that AI cannot replicate, are more familiar with AI tools than the juniors they are displacing. This is not mass unemployment in the aggregate. Overall employment continues to grow. What we are seeing is a reshaping, a reallocation of where human labour creates value. In Brynjolfsson’s words, it is the “fastest, broadest change” he has ever seen in the workplace, second only to the pandemic’s shift to remote work.

This pattern, fewer junior resources, more demand for experienced judgment, is a structural signal, not a cyclical blip.

The Inverted Pyramid: Fewer Juniors, More Orchestrators

If the evidence points anywhere, it points to an inversion of the traditional organisational pyramid.

For decades, organisations scaled by adding layers. Entry-level workers performed codified tasks. Middle managers aggregated, routed, and filtered information. Senior leaders made decisions based on what survived the filtering process. This was not just a management philosophy; it was an information processing architecture built around human cognitive limits. As Dorsey and Botha correctly noted, one human can effectively manage three to eight others. Scale beyond that, and you need another layer. Each layer adds latency, cost, and distortion.

Agentic AI breaks this constraint. When agents can triage, synthesise, monitor, and execute routine knowledge work, the coordination role of middle management becomes a technology function rather than a human one. The MIT Sloan Management Review and Boston Consulting Group’s 2025 research on the emerging agentic enterprise found that among organisations with extensive agentic AI adoption, 45% expect reductions in middle management layers within three years. This is not speculation. It is what companies at the leading edge are already planning.

McKinsey’s recent work on the “agentic organisation” reaches a similar conclusion. Most companies have added at least one management layer over the past decade, some two or three. This is expensive and slows decision making. AI creates the possibility for leaders to manage across significantly broader scopes, flattening structures and accelerating decisions.

Gartner predicts that by 2026, 20% of organisations will use AI to flatten their structures, eliminating more than half of current middle management positions in those organisations.

The shape of this new organisation is not a pyramid. It is something closer to a hub-and-spoke model. A smaller number of senior, experienced humans, operating as orchestrators, marshalling fleets of AI agents to do the knowledge work that was previously done by layers of junior and mid-level staff. The Deloitte State of AI in Enterprise 2026 report captures this well: organisational structures are beginning to flatten as AI absorbs routine execution tasks. Roles, skills, and career paths need to be rebuilt, not simply adjusted.

This is the inverted pyramid. Fewer entry-level roles. Fewer middle managers. More senior decision makers with broader spans of control, supported by agentic systems that provide real-time intelligence, execute workflows, and surface exceptions for human judgment.

The talent pipeline problem we cannot ignore

There is an obvious and uncomfortable question here that proponents of the inverted pyramid, myself included, need to confront honestly: if you eliminate the bottom of the funnel, where do future senior orchestrators come from?

Today’s experienced professionals gained their tacit knowledge, the judgment and pattern recognition that makes them valuable and that AI cannot yet replicate, by spending years doing the junior work that AI is now displacing. If entry-level roles shrink dramatically, we risk destroying the apprenticeship pipeline that produces the very expertise organisations need more of. A system that eats its own seed corn is not sustainable over a generation.

I do not think this is a reason to slow down. But it is a reason to redesign entry-level roles rather than simply eliminate them. The junior worker of the future may not write code from scratch or process claims manually. They may instead learn by orchestrating agents, reviewing agent outputs, handling escalations, and building the judgment that comes from seeing where AI fails and why. This is a fundamentally different apprenticeship model, and universities, employers, and professional bodies have barely begun to think about it. Any honest assessment of the inverted pyramid has to acknowledge that we do not yet have the talent development model to sustain it. The organisations that figure this out first will have a compounding advantage. The ones that simply cut junior headcount and celebrate the margin improvement may find, five years from now, that they have no pipeline for the senior orchestrators their model depends on.

Where It Will Land Fast, and Where It Won’t

Not all parts of an organisation will experience this shift at the same speed or in the same way. The rate of change depends on three variables: how codified the work is, how high the stakes of decisions are, and how entrenched the incumbent governance structures are.

Fast movers: Software engineering and technical production

Software engineering is the canary in the coal mine, and the data confirms it. Entry-level developer roles have already declined significantly. AI coding assistants are generating production-quality code, and Dario Amodei has stated publicly that some of his own engineers “don’t write any code anymore,” having transitioned to reviewing and directing AI-generated output. The engineer of 2026, as CIO magazine noted, spends less time writing foundational code and more time orchestrating a portfolio of AI agents, reusable components, and external services. The core skill becomes systems thinking, not syntax.

To be precise about what the data shows: this clearly means fewer junior software engineers. Brynjolfsson’s data is unambiguous on that point. The aggregate effect depends on how much new demand the productivity gains create. Firms that adopt AI-assisted engineering see 20% to 40% reductions in operating costs and 12 to 14 point increases in EBITDA margins, according to McKinsey. In many organisations, the result is dramatically higher output per engineer: more revenue, more product shipped, more value created per person on the team. Whether that translates to stable total headcount or reduced total headcount depends on whether the organisation’s appetite for building grows to match the new capacity. In growth-oriented tech companies, it likely does. In a corporate IT function running steady-state systems, it likely does not. The honest answer is: it depends on the objective function, and it will vary by industry, by company, and by team.

Fast movers: Service centres and outsourced knowledge work

The BPO industry, valued at over $300 billion in 2024 and forecast to exceed $525 billion by 2030, is arguably the most exposed sector to agentic disruption. BPOs were created specifically to provide cost-effective execution of high-volume, repetitive work: data entry, call centre operations, revenue cycle management, invoice reconciliation, claims processing. The original thesis was straightforward: centralise and offshore repetitive knowledge work for cost reduction and scale.

That thesis has not changed. What has changed, fundamentally, is the execution mechanism. And this is where the disruption is existential rather than evolutionary.

Interestingly, Andreessen Horowitz published a detailed analysis arguing that modern AI makes “productising and unbundling the BPO” entirely possible. Foundation models now handle data extraction, research, and complex reasoning. Voice AI agents are mature enough for large-scale production. Browser agents are close behind. The shift is from low-cost human labour to near-zero marginal cost agentic systems. This sits alongside Andreessen’s broader point about the lump of labour fallacy in an interesting way: the BPO sector may not disappear in aggregate economic terms, but it will be profoundly transformed in form.

Vinod Khosla warned at the India AI Impact Summit 2026 that India’s IT services and BPO sectors could “almost completely disappear” within five years. Indian BPO startups like LimeChat already deploy AI agents that handle up to 95% of customer queries without human assistance. As one displaced worker told Gulf News: “I was told I am the first one who has been replaced by AI.”

Behind statistics like this are real people, often young graduates in cities like Bengaluru, Hyderabad, and Manila, for whom BPO work was the first rung on the economic ladder. The human cost of this transition will not be distributed evenly, and the communities that built their economies around outsourced knowledge work face a particularly acute challenge. Leaders making decisions about agentic deployment need to hold that reality alongside the efficiency case.

The implication for enterprises is straightforward. If you are outsourcing work today because it is repetitive, codified, and rules-based, that work is in the first wave of agentic displacement. The question is not whether to bring it back in-house, but whether to bring it back as an agentic capability.

Slow movers: Tacit decision making and governance-heavy domains

Here is where Andreessen’s perspective on the pace of change has the most practical relevance.

Large parts of enterprise decision making involve what we might call perceived tacit nuance: decisions where the logic is not easily codified, the stakes are high, and the approval structures are deeply embedded in organisational culture and regulation. Think compliance, regulatory interpretation, capital allocation, strategic planning, people decisions, risk management.

In these domains, the challenge is not primarily technical. AI is increasingly capable of performing sophisticated analysis and generating recommendations. The challenge is structural and political. These are areas where incumbent approval structures and governance exist not just for quality assurance but as mechanisms of organisational power. When you propose automating a compliance review process, you are not just suggesting a technology change. You are proposing to disintermediate the people whose authority and career progression depend on being in that approval chain.

This is where Dorsey’s vision for Block encounters the reality of most enterprises. Block can restructure aggressively because it is a tech-first company with a founder-CEO who can set the direction. Most enterprises have boards, unions, regulators, legacy systems, multi-stakeholder governance frameworks, and decades of institutional muscle memory. For these organisations, the argument for risk will always be easier to make than the argument for benefit. And the risk argument, “What if the AI gets it wrong on a regulatory matter?”, is powerful precisely because it is hard to refute in the abstract.

Fortune’s reporting from Davos 2026 found that few C-suite leaders concurred with Amodei’s prediction that 50% of entry-level white-collar jobs would be eliminated within five years. Not because they disagreed with the technology trajectory, but because they understood the implementation reality. Amodei himself projected that 90% of code would be AI-written by end of 2025. That was true for Anthropic. At other software companies, it was 25% to 40%. The gap between what is technically possible and what organisations actually do is enormous.

A caveat here: the “slow mover” categorisation assumes AI capability improves incrementally in these domains. It may not. AI capability is advancing on exponential curves, not linear ones. A legal AI that can interpret regulation with genuine reliability, or a compliance agent that can navigate multi-jurisdictional frameworks with confidence, would blow through the “slow mover” category overnight. Leaders should plan for the expected timeline but not assume it is the only one.

The Uncomfortable Middle: This Is Not a Binary Choice

The real insight is that this is not a binary choice between rapid restructuring and patient scepticism. The impact of agentic AI on organisations will follow a familiar pattern of technology diffusion, but at an unfamiliar speed.

The Duke University/Federal Reserve survey of 750 senior finance leaders found that in 2025, the impact of AI on employment was negligible. But they expect it to grow in 2026 and beyond, with AI potentially reducing overall US employment by approximately 0.4%, or about 500,000 jobs. That is not mass unemployment. But as the researchers noted, 42,000 fewer jobs per month is not trivial.

The more revealing statistic comes from Mercer’s Global Talent Trends 2026 report: 40% of employees now fear losing their jobs to AI, up from 28% in 2024. Whether or not those fears are justified in the aggregate, the perception is reshaping how people relate to their organisations and their careers. Leaders who ignore this are building organisational change on a foundation of anxiety, which is a poor basis for the kind of adaptive, high-trust culture that agentic transformation demands.

My view is that the impact follows a three-horizon model, with the important qualification that these horizons may compress or overlap as AI capability accelerates:

Horizon 1 (Now to 12 months): Yield and productivity gains. The immediate impact is not fewer people but more output per person. Organisations that adopt AI-assisted workflows see meaningful improvements in what I would call yield: revenue per employee, throughput per team, margin per engagement. Software teams ship faster. Analysts produce more. Customer service scales without headcount growth. The result is better margins, not mass layoffs. Employment growth slows, particularly for entry-level roles, but it does not collapse.

Horizon 2 (12 to 36 months): Structural reshaping of specific functions. BPOs, shared service centres, and functions built around codified knowledge work begin to contract meaningfully. Companies that outsourced for cost arbitrage discover they can achieve better outcomes with agentic systems at lower cost. The inverted pyramid becomes visible: fewer juniors, fewer middle managers, more experienced orchestrators. Companies begin redesigning processes around agents rather than layering AI onto human workflows.

Horizon 3 (36 months and beyond): Deep organisational redesign. The harder, governance-heavy, tacit-knowledge domains begin to shift. Not because the technology forces it, but because competitive pressure from Horizon 1 and 2 organisations makes the status quo untenable. Enterprises that have built the data infrastructure, governance frameworks, and agentic capabilities begin to restructure decision-making architectures. The org chart gives way to what Microsoft has called “work charts,” task-focused networks where humans and agents collaborate in fluid configurations.

Most Organisations Are Not Tech Companies

This is the single most important point in this debate, and it is the one that almost every commentator underestimates.

The organisations that are most aggressively restructuring, Block, Meta, Amazon, Atlassian, are technology companies. They were built from the ground up with digital infrastructure, engineering-led cultures, and leadership teams who deeply understand what AI can do. They have the technology platforms, the data pipelines, the engineering talent, and the organisational agility to absorb rapid change.

Most organisations are not like this. Census data shows that as recently as the third quarter of 2025, actual AI adoption among large US businesses had only reached 12%. The tools required for the displacement story to play out at scale simply have not been deployed by the vast majority of enterprises.

For a mining company in Western Australia, a healthcare provider in regional South Australia, a rail network operator, or a government department, the path from “AI is impressive” to “we have restructured our organisation around agentic systems” is long and requires solving problems that tech companies have already solved: data quality and integration, security and governance frameworks, change management, workforce reskilling, regulatory compliance, and building the basic digital infrastructure that makes agentic systems possible.

This gap between capability availability and actual deployment is critical to understand. The technology to restructure is available now. The capacity of most organisations to deploy it is not. That distinction explains why Amodei’s timeline predictions have been accurate for Anthropic and inaccurate for the broader economy. It also explains why 12% adoption is not a reason for complacency. It is a reason for urgency. The technology is available. The competitive pressure is building. And the window between “this is coming” and “this is here” is shorter than any previous technology transition, even if the majority of enterprises have not yet begun the journey.

The organisations that will benefit most are not the ones that slash headcount first. They are the ones that build the foundations for intelligent work first.

The Paradigm Shift in Process Redesign: Designing for Agents, Not Humans

This brings me to what I believe is the most underappreciated dimension of this entire discussion. Everyone is talking about what AI can do. Very few are talking about how organisations need to fundamentally rethink the way they design and redesign work.

Every major wave of enterprise transformation has been powered by a process discipline. The quality revolution of the 1980s and 1990s drew on two distinct traditions: Lean, rooted in the Toyota Production System’s focus on flow, waste elimination, and pull-based production; and Six Sigma, rooted in Motorola and GE’s focus on statistical variation reduction and the DMAIC framework. These are different disciplines with different origins, but in practice most organisations combine them, and for good reason. Lean tells you what to eliminate. Six Sigma tells you how to measure and control what remains. Together they provide a rigorous, proven approach to understanding and improving how work gets done. As Genpact noted in a recent analysis, these principles are vital for the process intelligence that powers AI.

The discipline of understanding process, of mapping how work actually flows through an organisation, of identifying where value is created and where waste accumulates, of understanding where governance and approvals must exist and where guardrails are needed, none of that goes away in an agentic world. If anything, it becomes more important. You cannot deploy an agent into a process you do not understand. You cannot define guardrails for a workflow you have not mapped. You cannot govern what you cannot see.

But here is the fundamental paradigm shift: for the past century, every process improvement methodology has been designed to optimise for human workers. Value stream maps, SIPOC diagrams, time-and-motion studies, the entire toolkit assumes that the “worker” is a person with cognitive limits, training requirements, fatigue curves, a need for breaks, and social dynamics. When we redesigned processes under Lean, we redesigned them for humans to perform better. When we introduced RPA, we automated discrete steps within processes that were still fundamentally designed for human execution.

In an agentic world, the worker performing the process is increasingly not a person. It is an AI agent. And agents are a fundamentally different type of worker. They operate at near-zero marginal cost. They do not tire. They can process information at machine speed. They can execute in parallel across hundreds of instances simultaneously. They do not need training in the traditional sense; they need configuration, prompting, and access to the right data. They can learn and adapt continuously, but they can also drift, hallucinate, and fail in ways that are categorically different from human error.

How Lean principles transform when the worker is an agent

Consider some foundational Lean concepts through an agentic lens. Takt time, the rate at which you need to complete work to meet demand, is traditionally constrained by shift hours, breaks, and cognitive load. For agents, takt time is limited only by compute capacity and API latency. That is a radical shift in what production capacity means. One-piece flow, the Lean ideal of processing items individually rather than in batches, is something humans struggle to achieve at scale. Agents can do it trivially because they parallelise. Pull-based systems, where downstream demand triggers upstream work, become far more responsive when the workers responding operate in milliseconds rather than hours.

And here is something that process practitioners need to think carefully about: the classic forms of waste, the seven wastes of muda (Transport, Inventory, Motion, Waiting, Overproduction, Overprocessing, Defects), do not simply disappear when agents replace humans. Some shift. Waiting waste largely evaporates. Motion waste becomes irrelevant in a digital context. But others persist or even intensify. Overproduction is a real risk when agents can produce output so fast and so cheaply that they flood downstream processes or customers with work nobody asked for. And critically, agentic systems introduce entirely new categories of waste that have no direct equivalent in human-centred Lean:

Hallucination waste: The agent produces plausible but incorrect output that consumes downstream review effort, rework, and potential reputational damage before it is caught. This is not a human-equivalent defect. It is a failure mode unique to generative systems.

Drift waste: The agent’s performance gradually degrades as the data environment changes around it, meaning outputs produced before drift was detected may need to be reviewed or reworked retrospectively. Unlike human skill degradation, which is gradual and observable, agent drift can be silent until it crosses a threshold.

Escalation waste: Confidence thresholds are set too conservatively, causing the agent to over-escalate to humans, which defeats the purpose of agentic execution and creates a new bottleneck at the human review layer.

Compute waste: The agent runs expensive inference on tasks that could be handled by simple business rules or lighter-weight models, consuming cost and latency without adding value.

Naming and measuring these new forms of waste is essential. A Lean to Agentic approach must extend the waste taxonomy, not just inherit it.

The claims processing example, more honestly

To make this concrete, consider a claims processing workflow. The human-optimised version has quality checks at multiple stages because humans make errors that accumulate. It has supervisory review because individual judgment varies. It has batch processing because humans work in shifts. It has handoffs between teams because human specialists have narrow domains.

The agent-optimised version is fundamentally different. An agent can process end-to-end without handoffs. It can apply consistent judgment across every single case. It does not need batch processing because it operates continuously. The quality controls shift from checking human work to monitoring agent confidence scores, detecting drift, and governing exception handling.

But let me be honest about what this does not mean. It does not mean agents handle 100% of claims autonomously. In reality, perhaps 80–85% of claims follow a standard path that an agent can handle end-to-end within well-defined guardrails. The remaining 15–20% involve regulatory edge cases, appeals, fraud patterns, or emotionally complex situations, a customer who is grieving, a dispute that requires empathy and judgment, that require human involvement. The interesting process design challenge is not the autonomous path. It is the hybrid handoff: how the agent recognises the boundary of its authority, escalates cleanly, and provides the human with full context so they can act without re-investigating from scratch. Designing that boundary, and the governance around it, is where the real process discipline lives.

How DMAIC shifts when you design for agents

The Six Sigma DMAIC framework maps naturally to the agentic lifecycle, but each phase changes meaningfully when the executor is an agent rather than a person.

In the Define phase, you still need to start with the Voice of the Customer. This is non-negotiable and it is the first question any serious process redesign must answer: does the customer care whether a human or an agent does this work? In many contexts, they do not. In some, they absolutely do: healthcare, financial advice, legal services, bereavement claims. Redesigning for agents without starting from VOC is designing from the inside out, which is the exact mistake Lean was created to prevent. Beyond VOC, you must map the current state before you can redesign. You cannot skip current-state analysis and go straight to green-field design, because some of the constraints that look like “human limitations” are actually regulatory requirements, contractual obligations, or risk mitigations that exist for good reason. The Define question then becomes: given the current process, the customer’s needs, and the regulatory landscape, what is the minimum viable set of human decision points required, and what is the decision authority of the agent? Where can it act autonomously? Where must it escalate? What constitutes a failure? These are governance questions, not technology questions.

In the Measure phase, the shift goes deeper than just tracking different KPIs. Six Sigma’s Measure phase is fundamentally about establishing a validated measurement system and understanding process capability. For agents, this means: how do you validate that a confidence score is actually measuring what you think it is measuring? What is the agent’s equivalent of process capability (Cp, Cpk), meaning how reliably does it produce outputs within specification? What is its sigma level? Beyond capability, the operational metrics themselves change. You are measuring throughput across parallel instances, confidence scores per decision, escalation rates, drift detection over time, cost per transaction, and the ratio of autonomous completions to human interventions. The measurement framework must account for the fact that agents improve continuously but can also degrade as the data environment changes around them. Measurement System Analysis, the Six Sigma discipline of validating that your measurement instrument is reliable before you trust the data, is arguably more important for agent systems than it ever was for human ones, because the measurement instruments (confidence scores, embedding similarity, retrieval accuracy) are themselves probabilistic.

In the Analyse phase, when agents underperform, the root cause is almost never “the agent needs more training” in the way a human might. A spike in escalations might stem from incomplete source data, a change in input formats, or a shift in the underlying business rules. These are process problems, not model problems, and traditional quality tools like fishbone diagrams and Pareto charts remain effective for initial diagnosis. But agent failure modes are often multivariate and interact in non-obvious ways. A model might perform well on each input dimension independently but fail on specific combinations. This is exactly the kind of problem that Six Sigma’s Design of Experiments methodology is built to investigate: systematically varying inputs to isolate which factors and interactions drive variation in output quality. The analytical rigour of Six Sigma is not just compatible with agentic systems; it is essential for understanding them.

In the Improve phase, you refine agents through prompt engineering, retraining, data pipeline improvements, or business rule updates. You build guardrails: confidence thresholds, fallback protocols, escalation logic. This is where the Lean principle of continuous improvement meets the iterative nature of agent development. But the cycle time of improvement itself collapses. Where a human process improvement project might take weeks or months to design, pilot, and roll out, an agent refinement can be tested and deployed in hours. The kaizen cycle accelerates by orders of magnitude.

In the Control phase, the shift is perhaps most profound. Traditional Six Sigma control relies on Statistical Process Control: control charts that distinguish common-cause variation (normal, inherent to the process) from special-cause variation (something has structurally changed, investigate). This distinction is exactly what you need for agent monitoring. Is this agent’s performance drift within normal variation, meaning leave it alone? Or has something structurally changed, meaning investigate and intervene? SPC applied to agent output quality, escalation rates, and confidence score distributions gives you a principled, statistically grounded way to know when to act and when to leave the system alone. As McKinsey has emphasised, governance in the agentic organisation cannot remain a periodic, paper-heavy exercise. As agents operate continuously, governance must operate continuously alongside them. SPC provides the statistical backbone for that continuous governance, and it is a discipline that already exists. We just need to apply it to a new type of worker.

The core thesis

This is a paradigm shift in how change itself needs to occur. The disciplines of Lean and Six Sigma, the rigour of understanding process, the commitment to eliminating waste, the tools for measuring and controlling variation, all of this endures. It is proven. It works. But the design target has changed. We are no longer redesigning processes to make humans more efficient. We are redesigning processes so that agents can execute them, with humans operating above the loop as governors, exception handlers, and strategic decision makers.

I think of this as the shift from Lean to Agentic. Not a replacement of Lean Six Sigma, but an evolution of it for a world where the worker on the factory floor of knowledge work is an AI agent.

The concept of the “agent factory” captures this well. The process is still the process. The objective function, delivering value to the customer, reducing waste, maintaining quality, ensuring compliance, has not changed. What has changed is who (or what) is doing the work. And that single change, from human to agent as the primary executor, cascades into every dimension of process design: how we define success, how we measure capability, how we identify failure modes, how we govern, and how we improve continuously.

The organisations that skip this discipline, that simply “plug AI into” existing human-designed workflows, will discover what every previous automation wave has taught us: automating a broken process just produces broken outcomes faster. And automating a human-optimised process with an agent leaves enormous value on the table, because you have constrained the agent to work within a design that was built around limitations it does not have.

What This Means for Leaders

The contrasting perspectives from Dorsey and Andreessen are both valuable precisely because they illuminate different dimensions of the same challenge. But neither addresses the question that matters most for the majority of business leaders: How do I get from here to there?

The answer is not to wait for consensus. The technology trajectory is clear enough. Agentic autonomy will increase dramatically in the coming months and years. Full end-to-end capability, if not already available, is arriving fast. Certain industries will reshape aggressively for cost and margin benefits. Others will pursue greater yield and productivity. The objective function varies by industry, by company, and by competitive context. But the direction does not.

What differs is the path. For the small number of tech-first companies, the path is rapid restructuring, shedding management layers, and building intelligence systems that replace coordination functions. For the vast majority of enterprises, the path requires building the infrastructure, architecture, data foundations, and process discipline to do this safely and responsibly.

The human dimension of this transition demands honest attention. There are real people in middle management roles, in BPO centres, in entry-level knowledge work positions, whose livelihoods will be affected. Leaders have a responsibility not just to their shareholders but to the people who built the organisations they are now transforming. That means investing in reskilling, redesigning junior roles for the agentic era rather than simply eliminating them, and being transparent about the direction of travel rather than pretending it is not happening. It may also mean engaging with policymakers on questions of transition support, tax policy, and social safety nets, as Amodei and others have urged. The efficiency case for agentic transformation is compelling. The human case for managing that transformation responsibly is equally so.

The organisations that will thrive are not the ones who move fastest to cut headcount. They are the ones who invest in understanding their processes deeply, redesign those processes for agentic execution, build the governance frameworks to do this safely, develop the human capability to orchestrate hybrid human-agent teams, and take seriously the responsibility of managing the transition for the people it affects.

This is not a technology project. It is an organisational paradigm shift. And like every paradigm shift before it, it requires not just new tools but a new discipline for change itself, one that respects the proven foundations of process excellence while fundamentally rethinking the design target.

The hierarchy is not dead. But it is being redesigned. The question is whether your organisation will design its future, or have it designed for you.