Scale

Deciding what should scale

The objective is a better outcome. The design question is where additional people, coordination and capital improve that outcome—and where AI allows a more concentrated human core to achieve it, with humans deciding, approving and able to override throughout.

How the workforce was reshaped

  1. 1776The division of labourSmith's pin factory: dividing work into specialised tasks multiplies what each person produces.
  2. 1990Redesign the work, don't automate itHammer: automating an outdated process only speeds up its faults; the process itself has to be reengineered.
  3. 1998Business Workflow RedesignBart de Graaff, WFM '98: use the technology itself to analyse workflows and keep them current, not documentation that only fills cabinets.
  4. 2014Self-management beyond hierarchyLaloux: self-managing teams take on the coordination that hierarchy used to carry.
  5. NowAI agents join the workforceFor work that can be specified and verified, execution capacity decouples from headcount.

Each wave of technology has reshaped the workforce. The industrial revolution moved work into factories and divided it into specialised tasks. The computer automated calculation and record-keeping, workflow systems carried redesigned processes from desk to desk, the knowledge industry made expertise the scarce resource, and digitalization put the enterprise's work into systems and data. AI now crowns that sequence and disrupts it: for the first time, the technology does not only support the work but carries part of it.

An execution model should follow the work. Some activities need reach, local knowledge and the capacity to act in many places at once. Others need a few people with deep context directing systems that can produce and verify more than those people could do alone.

Capital commits resources, strategy sets the intended outcome, and structure determines how capability and authority are organized. Scale amplifies the result. AI changes the execution capacity within that chain; it does not settle the structural decision for you.

Most real transformations contain both kinds of work. Software may be developed by a concentrated team while migration, adoption, regulation and physical rollout require coordinated delivery at scale. The question is how control stays concentrated where it matters while execution grows around it.

Three execution architectures make that choice explicit. An initiative can combine them across workstreams, provided the boundaries and decision rights are deliberate.

In each of them, humans stay in the loop: they decide, approve and can override what machines do. That only holds if the technology is safe and governed by construction, so the same thread runs through every section below.

Scale-Inverted

When the work can be specified and verified, and the constraint is judgment rather than hands.

When this fits →

Hybrid

When some of the work inverts and some does not — most material enterprise transformation.

When this fits →

Consolidated

When the difficulty is migration, adoption, regulation or risk absorption, and capacity is the constraint.

When this fits →

Prerequisites To The Workforce Transformation

Four foundations come before work moves to agents.

Without them, agents scale the existing data problems, authority gaps and control weaknesses along with the work. Each depends on the one before it: agents act on data, within authorities, through an orchestration layer, inside an operating model.

“Don't automate, obliterate.”
Michael Hammer, “Reengineering Work: Don't Automate, Obliterate”, Harvard Business Review, 1990 ↗

None of this is new. Every earlier wave of workforce change taught the same lesson, and agents raise its stakes.

Time-and-motion studies
Work was observed and measured before it was changed. Redesign started from how the work was actually done, not from how it was assumed to be done.
Business process redesign
Automating a broken process made it fail faster. The gains came from redesigning the process first, then deciding what the technology should carry.
Workflow implementations
A workflow system routes exactly the work, data and authorities it is given. Implementations stalled where roles, ownership and exceptions had not been settled before go-live.
Garbage in, garbage out
The oldest rule in computing holds with more force for agents, which act on the data and authority they find, at machine speed and without pausing to doubt them.

  1. 1

    Data cleansing and data ownership

    Agents act on the data they are given. Records have to be cleaned, reconciled and current before work moves onto them, and every data set needs a named owner who answers for its quality and decides who, and what, may use it.

  2. 2

    Governance and the assignment of authorities

    Before an agent acts, it has to be settled what it may do, on whose authority and up to what limit. Authorities are assigned explicitly to people and delegated to agents within bounds, so every action traces back to someone entitled to authorize it.

  3. 3

    Orchestration with inherent safety

    The orchestration layer routes work between agents and people, and it must enforce the boundaries itself: permissions checked at the point of action, approvals that cannot be skipped, a complete audit trail and a safe stop when something falls outside the design.

  4. 4

    The updated operating model

    Roles, decision rights, incentives, funding and performance measures change with the workforce. A new execution system run inside the old operating model is gradually reabsorbed by it.

Three Execution Architectures

Selected by the characteristics of the work, not by ideology or provider economics.

The same enterprise may need all three. Their differences concern the source of execution capacity, the way authority is held and the costs that come with each choice.

Scale-Inverted

Concentrated human design and control, machine leverage for execution.

For example, a defined digital workflow with explicit inputs, bounded actions and reliable acceptance checks can be designed and directed by a small team. The team owns the context and the verification; machines extend its execution capacity.

Fits
Work that can be specified, verified and repeated, where the constraint is judgment rather than hands.
Costs
Depends on a small number of people holding deep context. Verification becomes the bottleneck, and losing a key person is an architectural risk rather than a staffing one.
Humans in the loop
People set the design and the acceptance criteria, approve at designed checkpoints and can override any agent run. With fewer people, each checkpoint carries more weight, so the system has to enforce it rather than rely on someone remembering it.

Hybrid

Concentrated design and value authority, with deliberately scaled internal, SI or specialist execution where that genuinely adds value.

An enterprise platform change may concentrate architecture, product judgment and verification while using larger delivery teams for integration, migration and adoption. The client retains design and value authority; the delivery footprint follows the work.

Fits
Most material enterprise transformation, where some work inverts and some does not.
Costs
The boundary has to be maintained. Without it, scaled execution gradually reabsorbs design authority and the architecture becomes Consolidated by drift rather than by decision.
Humans in the loop
People hold the decisions on both sides of the boundary: approval gates where agents execute, management authority where teams scale, and one escalation route across the two.

Consolidated

Substantial coordinated human scale, with control distributed through management structure.

A regulated rollout across operating sites may need specialist teams working in parallel, local decision-makers and a substantial transition organization. AI can improve preparation and monitoring while the physical and institutional work still requires human capacity.

Fits
Work whose difficulty is migration, adoption, regulation, physical rollout, local context or risk absorption — where capacity is the constraint.
Costs
Coordination cost is real and can grow faster than delivery capacity. It buys the ability to act in many places at once, and that is worth paying for when the work genuinely requires it.
Humans in the loop
People decide and approve through the management structure. Agents and Edge AI prepare, monitor and alert; a person acts on what they surface.

Humans In The Loop

People decide, approve and override. The technology has to make that real.

Whatever the balance in a layer, three things stay with people. They are designed into the work at the points where they matter, not left to the judgment of whoever happens to be watching.

Decide
People set the objective, the boundaries and the trade-offs. Agents prepare options and evidence; the choice, and the accountability for it, stay with a named person.
Approve
Consequential actions, such as a payment, a change to a live system, a message to a customer or a safety-relevant instruction, wait for a person at a designed point, sized to what an error would cost.
Override
People can stop, correct or reverse machine-led work at any time, without asking the system being overridden for permission.

Humans in the loop works only if the technology is inherently safe and governed. An approval that can be bypassed, an audit trail that can be edited or an agent that can widen its own permissions turns oversight into ceremony. So actions are bounded before they run, authority is checked at the point of use rather than described in a policy, every action is attributable and reviewable, and when a check cannot be completed the system stops rather than proceeds. Governance is designed into the execution system, not added after it (see Control).

The Workforce Behind Each Architecture

The balance between human and agentic workforce is a design decision, layer by layer.

Each architecture implies a different workforce. The diamond-shaped workforce describes where people sit once agents take on part of the work: the total work stays the same pyramid, human effort concentrates in a diamond inside it, and agents carry the difference. How far that concentration should go follows from the work, layer by layer, and it differs by architecture.

Scale-Inverted workforce. The dashed pyramid is the total work. The human workforce is a narrow diamond inside it: it shares the apex, its vertical sides end at the Expertise layer and it rests on a very short base. The hatched agentic workforce fills most of the pyramid from Expertise down.
Scale-InvertedThe narrowest diamond. Agents carry most of the work from Expertise down; people hold Direction and deep Expertise, and engineer, orchestrate and operate the agents. At the frontline, connected tools, Edge AI and agentic support work beside the people on site.
Hybrid workforce. The dashed pyramid is the total work. Direction and Coordination form the shared human apex. Below them a marked vertical boundary divides the pyramid. On the left, labelled inverts, is the diamond's side with the hatched agentic workforce beside it. On the right, labelled scales, a grey human segment of the pyramid carries migration, adoption and rollout.invertsscales
HybridA diamond for the workstreams that invert, standing inside a human pyramid segment for migration, adoption and rollout. Design and value authority stay concentrated above the shoulders; the boundary below them is marked and maintained. On the ground, people carry rollout with Edge AI, connected tools and a digital twin.
Consolidated workforce. The dashed pyramid is the total work, and the human workforce fills almost all of it. Only the lower corners are hatched: agents augmenting the base of the work.
ConsolidatedClose to the traditional pyramid. Agents augment the base: Digital assistants in operations and, at the frontline, Edge AI, connected tools and agentic support beside people at many sites. Human capacity stays dominant because capacity is the constraint.

Layer by layer

The same layers, held differently. Each row is a layer of the diamond with its agentic counterpart; each column shows who carries that layer's work under one architecture.

LayerScale-InvertedHybridConsolidated
Directionno agentic counterpartHuman-ledA small team owns the outcome and answers for it.Human-ledThe client holds design and value authority across the boundary.Human-ledHeld through the management structure.
Coordinationno agentic counterpartHuman-ledLean: agents absorb routine handoffs; the decisions stay human.Human-ledLean inside the workstreams that invert, fuller where delivery scales.Human-ledSubstantial, because many teams act in parallel.
ExpertiseSpecialist agentsSharedSpecialists frame and judge; Specialist agents prepare and check.SharedArchitecture, product judgment and verification stay concentrated; Specialist agents prepare the work.Human-ledSpecialist teams in parallel; Specialist agents help with preparation.
Agent engineeringBuilder agentsAgent-ledBuilder agents write code, tests and evaluations under engineering review.SharedBuilder agents in the workstreams that invert; delivery teams for integration.Human-ledConventional engineering; Builder agents are tools, not the delivery model.
Agent orchestrationAI orchestratorsAgent-ledAI orchestrators route and schedule work inside the design people set.SharedAI orchestrators on one side of the boundary, managed teams on the other.Human-ledWork is planned through the management structure; AI supports monitoring.
Agent operationsDigital assistantsAgent-ledDigital assistants execute; people step in at designed points.SharedDigital assistants where work inverts; people carry migration and adoption.SharedDigital assistants augment a large human delivery base.
Physical / frontline workEdge AISharedIn scope: agents prepare and report the work, connected tools and Edge AI check it as it is done; people do it and decide on site.SharedEdge AI and connected tools where the work inverts; rollout on the ground stays with people, fed by a digital twin.Human-ledPeople at many sites, supported by agentic workflows, connected tools, smart cameras and data at the point of work.

The shapes and the table are BdG's framing of how the diamond applies to each architecture. The proportions are illustrative, not measurements.

Finding the balance

Four questions place a layer. Ask them of the work in each workstream rather than of the organization as a whole; the same layer can land differently in two workstreams, which is what makes an architecture Hybrid.

Can the work be specified?

Toward agents whenInputs, bounded actions and acceptance criteria can be written down.

Stays human-held whenThe work is still being framed, or the objective moves as it is done.

Can it be verified independently of whoever, or whatever, produced it?

Toward agents whenA person or a separate check can establish that the result is right.

Stays human-held whenOnly the producer can judge it, so more agent output means more to inspect, not more done.

What does an error cost, and who answers for it?

Toward agents whenErrors are caught before they matter and can be reversed.

Stays human-held whenErrors are costly, regulated or irreversible: a person holds the decision, and agents prepare it.

Does it need local context, presence or relationships?

Toward agents whenIt can be done from the information available, wherever the work sits.

Stays human-held whenIt needs people on site, local judgment or trust. Agents and Edge AI augment that capacity rather than replace it.

Verification carries the most weight. Where a layer moves toward agents, someone still has to establish that the work is correct, independently of whoever produced it; that becomes the load-bearing control (see Control). Nor is the balance a headcount target. As with Execution Density, the measure is the outcome that concentrated human judgment can direct, not the number of people removed.

The full argument: the diamond-shaped workforce, layer by layer →

Physical And Frontline Work

The frontline is in scope, and it changes too.

Much of an enterprise's value, risk and workforce sits on the ground: in plants, networks, sites and the field. That work still needs people, and every architecture stands on it. What changes is what those people are given and what they carry: agentic support, connected tools and Edge AI take on preparation, checking and some tasks outright, while the person on site keeps the decisions that have physical consequences.

Agentic support for frontline workflows
Agents prepare the work before a shift, such as permits, parts, schedules and the asset's history, take on the reporting afterwards and answer questions in the field, so the person on site spends their time on the task rather than on the administration around it.Humans in the loop: The worker decides how the job is done on site. The agent supports; it does not direct.
Intelligent connected tools
Tools that know where and how they are used: a connected drill that confirms it is at the specified location and setting before it runs, or a torque tool that records each fastening against the work order. Quality is checked as the work is done rather than inspected afterwards.Humans in the loop: The tool can hold a step that does not match the specification; the worker and supervisor decide what happens next.
Cameras with inference on the device
Smart cameras run inference on the camera itself, so images need not leave the site: helmet and PPE compliance at the entrance to a work zone, weapon detection, or confirming that a person is authorised to enter a restricted area. Privacy safeguards are designed in from the start: processing on the device, minimal retention, no identification beyond what the purpose requires, and lawful, proportionate use.Humans in the loop: A camera raises an alert; a person reviews it and decides on any action. No consequential decision about a person is taken by the camera alone.
Data provided where it is needed
The right information at the point of work: task-specific training just before the task, step-wise standard operating procedures on a handheld or headset, and the asset's history and hazards in view as the work starts.Humans in the loop: Procedures are written and approved by the people accountable for them. The system delivers them; it does not rewrite them.
Digital twins with a feedback loop
A digital twin of a site, network or asset combines live data from connected tools, sensors and cameras with the engineering model, so problems can be seen, simulated and planned remotely. The loop closes back to the people on site: what the twin learns becomes updated instructions, priorities and warnings for local workers, and what they observe corrects the twin.Humans in the loop: Changes planned in the twin reach the physical world only through people with the authority to make them.

Below the ground line

Physical and frontline work changes too.

Two rows of physical and frontline tasks. Before: every task is performed by a person. After: most tasks are still performed by people, now augmented by Edge AI and smarter devices, and a portion is handled by Edge AI devices such as smart cameras. Illustrative, not to scale.BeforeAfterPeople, augmented by Edge AIEdge AI devices
  • Performed by people.
  • Augmented by Edge AI and smarter devices.
  • Handled by Edge AI devices, such as smart cameras.
Illustrative, not to scale.

Execution Density

The transformation outcome created from a concentrated amount of human judgment.

Human effectiveness creates direction. AI creates leverage. Execution Density describes the meaningful transformation outcome that results from a concentrated amount of human judgment, creativity, accountability and decision effort. It increases when the same human capability can direct better work, resolve more of the right decisions and carry its intent through to delivery.

This is not a headcount-reduction metric.

FTE arithmetic misses the value of clearer judgment and better outcomes. A programme can raise density by removing coordination cost and clarifying decision rights without changing headcount at all. Equally, removing people without protecting context and verification can reduce the organization’s ability to execute.

Concentrated Control, Scaled Execution

A Hybrid architecture only stays Hybrid if the boundary is maintained. Design and value authority stay concentrated; execution scales around them. Without a maintained boundary, scaled execution gradually reabsorbs design authority and the architecture becomes Consolidated by drift rather than by decision.

This matters especially where internal teams, SIs and specialists share delivery. Clear decision rights let those teams act without renegotiating the objective at every handoff. Independent verification and escalation keep the intended outcome visible when schedules or commercial pressures pull another way.

–
Decision rights: Who decides what, explicitly, so escalation is a designed path rather than a symptom of ambiguity.
–
Sequencing: Development follows a defined order, not the loudest commercial pressure of the quarter.
–
Incentives: Authority, incentives and accountability reinforce each other rather than pulling apart.
–
Verification: Someone can establish that the work is correct, independently of whoever produced it. Where execution is machine-led this stops being a quality step and becomes the load-bearing one.
–
Escalation: A defined route for the decisions that exceed the level currently holding them.
–
Governance: Oversight proportionate to consequence, so it constrains what matters without adding latency to what does not.
–
Override: People can stop, correct or reverse machine-led work at any point. Override is a property of the system, not a permission the overridden system grants.

Two-Speed Transformation

AI-led redesign can learn faster than a large transformation can safely change. A team may test a different approach to development or support in weeks while an enterprise migration is committed to a much longer sequence. Both speeds are legitimate; their connection has to be designed.

The fast AI-led redesign and learning cycles have to feed the main transformation — the same decision rights, the same sequencing, the same governance. A useful result needs a route into the target operating model, an owner who can authorize the change and a controlled transition from the current delivery path. Otherwise the organization learns in one place and keeps executing the old model in another: innovation theatre with no route into delivery.

What The Design Has To Survive

These forces act on any architecture. They are not the argument that a problem exists — that is Execution — but the conditions a design is chosen under.

Capital pressure

Growth expectations accelerate faster than operational capacity, pushing commercialization ahead of delivery readiness.

Governance friction

Decision rights migrate as companies grow. Additional layers slow execution unless authority and escalation stay explicit.

Commercial sequencing

Deal-driven development fragments the roadmap and erodes scalability where sequence control is absent.

Organizational scaling

Coordination overhead rises with team size, with or without AI. Misaligned structure raises decision latency.

From pyramid to diamond: the shape of human effort →

Technology architecture

Early decisions optimize for speed; at scale the shortcuts constrain delivery. AI amplifies the constraint rather than removing it.

Next / EngageMake the choice—and hold it through delivery. →

Choose the senior involvement your transformation needs.

The economic argument behind these choices: how AI changes execution and scale →