The Rise of Digital Labor

Why Enterprises - and Nations, Must Prepare for a World Where Labor Is No Longer Exclusively Human

Every major economic transformation has been defined by a new source of productive capacity. Mechanization multiplied physical labor. Computing multiplied information processing. AI is beginning to multiply knowledge work, and the enterprises absorbing it are finding that their existing categories do not fit.

Most organizations still classify AI as software. Software is specified, purchased, deployed, and maintained. A system that self-initiates, holds objectives across time, retains memory, develops its own strategies, and improves through operating experience is not behaving like software. It is behaving like labor, and labor is recruited, governed, measured, and developed. The apparatus for one does not transfer to the other.

This paper runs further forward than the others in this series. The first four observations describe what is already underway inside enterprises. The last three describe institutional and geopolitical consequences that are less certain and in places contested, included because the planning horizon for infrastructure, licensing, and fiscal exposure is longer than the horizon of confidence.

They lead to one conclusion. The difficult problems of digital labor are not technical and never were. They are problems of management, accounting, and institutional design, and each of them is often owned by nobody.

1. What Makes an Agent Labor Rather Than a Tool

• "Those entities are living alongside humans and not in their own system."
• "It is a very big organizational change to move them to the org chart, from the IT department."

The distinction between a tool and a worker is not about capability. It is a set of properties: the ability to initiate rather than wait, to hold a long-horizon objective and decompose it into what to do next, to retain memory of what was done and why, to develop its own strategies, and to remain aware of the environment it operates in. A system with those properties does not need a person to decide what it should do this hour. That is the threshold.

The organizational consequence is a move from the IT department to the org chart, and it is more disruptive than it sounds. Almost every enterprise runs a parallel governance track for AI: separate policies, separate approvals, separate owners. That track is the artefact of treating agents as systems. Bringing them into the org chart means bringing them inside the standardization and performance discipline that already govern people. There are still managers and still approvals. The difference is that they are the existing ones.

The capacity argument underneath is the one most often missed. Under the tools model the human remains the ceiling, because a person with excellent tools is a faster person and the organization is still bounded by how many people it has. Agents use the same tools people use. What changes is that capacity stops being limited by human labor, and that is the entire economic difference.

Questions:

• Do our AI deployments initiate work, or do they wait to be asked?
• Are we running a separate approval structure for agents, and what is the argument for keeping it separate?
• Is our capacity still bounded by headcount, and if so, what have we actually changed?

2. The Economics Turn on Autonomy, and Then on Experience

• "They are moving from being a junior that read all the books in the world to someone that actually runs operational decisions."
• "Most people would prefer a surgeon with twenty years of experience over one who just started but read everything."
• "It can go from zero to a hundred times in a few days, and it can take a few years."

An assistive deployment produces a faster person, and its returns are bounded by that person's hours. An autonomous deployment produces capacity that is not bounded by anyone's hours. These are different economic objects, and the difference in return between them is not incremental. Any enterprise disappointed by its results should first establish which of the two it built, because most have built the first while expecting the economics of the second.

The economics are also not fixed at deployment. They improve along two curves: model capability rises through external investment, and operating experience rises through use. An agent at deployment is a junior who has read every book in the field. Operating experience converts it into something that has run the process, met the exceptions no training data anticipated, and adjusted. Enterprises pricing the deployment-day position are treating a floor as a result.

The rate at which that experience accrues is the variable nobody plans around, and it differs by orders of magnitude across processes that look superficially similar. Two things govern it. Feedback volume: a service process handling millions of interactions daily generates enormous outcome data, and an agent inside it develops operating judgment quickly. And cycle time: a trading strategy evaluable only weekly, requiring an understanding of seasonality and shifting market dynamics before results mean anything, produces a fraction of that signal in a year. Same technology, same intent, entirely different rate of return. This makes experience velocity a deployment criterion, and most organizations do not use it. They select by visibility, executive sponsorship, or whichever business case was easiest to write. It also reframes patience: a slow-feedback deployment showing modest results after two quarters is behaving exactly as its structure dictates, and enterprises that do not distinguish the two will cancel their long-horizon work at the point of maximum discouragement and minimum information.

The competitive corollary belongs in front of a board. An enterprise starting two years later does not simply lag by two years of installation. Its agents are less capable at the same task, because a competitor's have two years of operating history and its own have none. The gap widens rather than narrows, because the experienced population keeps learning throughout the period the late starter spends deploying.

Questions:

• Have we built assistance or autonomy, and would our results look different if we were honest about which?
• For each deployment, what is the feedback volume and cycle time, and did either factor into where we started?
• If a competitor started two years before us, what would they have that we could not buy?

3. Where the Displacement Lands, and How It Gets Booked

• "We are not necessarily reducing employees. We are reducing contractors and offsetting future hiring."
• "We can prove the productivity. We cannot find it in the P&L."

The public conversation about AI and work is conducted in redundancies. What is actually happening inside enterprises is different, narrower, and worth stating precisely, because the difference determines both where the value shows up and who absorbs the adjustment.

Broad headcount in the aggregate is not falling. Sourced and contracted capacity is, and in specific functions such as the service desk the reduction is direct. Beyond that, the dominant pattern is hiring offset: requisitions revised before they open, growth absorbed rather than staffed. Flexible resourcing is the deliberate shock absorber, and several enterprises hold a substantial proportion of their workforce in flexible arrangements precisely so that adjustment never requires touching permanent staff. None of this produces a visible workforce event, which is why the public account and the internal reality have diverged so far.

That invisibility is also why the value is so hard to book. Under the tools model the gain is time, saved by individuals, spread thinly, in units too small to redeploy, so nothing is released and nothing appears. Under the labor model the gain is capacity, which can be counted and allocated. The enterprises that have solved the accounting separate direct value, which finance will sign off and the company will state publicly, from indirect value, which gets a figure but is never claimed. The rule that makes the direct category defensible is stricter than it looks: released capacity only counts if it is offset against a plan finance had already approved. Hiring avoided is real if the requisitions existed and are now unfilled. It is not real if the roles were aspirational. Without that rule an enterprise is arguing about savings nobody can locate.

The exposure underneath is the talent pipeline, and it has a long fuse. The capacity being displaced first is disproportionately entry-level, contracted, and junior work, which is where people historically entered industries and built the judgment that made them senior. Removing it produces no visible unemployment and a capability gap that surfaces years later, by which point the cause is hard to establish. The compensation market is already showing the strain in a dumbbell shape, with senior people who can build and direct this work commanding a premium, early-career people who grew up with these tools in demand, and the middle under pressure. The disciplined response is to preserve some junior roles deliberately, not because the work cannot be automated, but because the enterprise still needs a mechanism for producing its own experts.

Questions:

• Would our finance function sign off on the value we claim, and against which approved plan?
• How much of our reduction was a decision, and how much an accumulation of separate renewal calls?
• If we remove the work where our people historically learned the business, what replaces it?

4. Management Flattens Because Coordination Was Most of It

• "We promoted them to coordinate. That is now the part we do not need."
• "Everything is tied to scope, to the number of people you are managing."

Coordination is the substantial majority of what management layers do: routing work, chasing status, reconciling handoffs, aggregating information upward and translating direction downward. It is also the part AI does well, including coordination between agents and between people inside a mixed team. Remove most of that load and span of control widens substantially, layers become unnecessary, and the middle of the organization contracts. What remains is the work that was always higher value and always crowded out: mentoring, developing people, judgment on exceptions, accountability for outcomes.

The blocker is not capability. It is that scope, status, and compensation are indexed to headcount. A leader of sixty thousand is regarded as important because of the sixty thousand rather than the return generated, and organizations have inflated structure for decades partly because that is what the reward system recognized. Nothing changes in the operating logic until that indexing changes, which means rebuilding job architecture, compensation bands, and promotion criteria together.

The historical precedent is exact. The manager of ten thousand manual workers did not survive mechanization as a role, and the plants that did not restructure became boutiques or closed. The honest test today is the counterfactual: very few enterprises building themselves now would produce the organization they currently have. That gap persists because nobody owns it. Individual leaders optimize within their boundaries and are rewarded for it, while the redesign that matters crosses every boundary and threatens the people best placed to execute it.

Questions:

• What proportion of our management layers exist to coordinate, and what would remain if coordination were handled?
• Are managers evaluated on return generated or on people managed, and which does our promotion history reward?
• If we designed this organization from scratch today, how different would it look, and has anyone senior been asked to answer seriously?

5. When Output Stops Tracking Employment

• "Even if it is a ten percent probability, it is better to assume that than to assume the opposite."

Everything to this point sits inside the enterprise, where the decisions belong to management. What follows does not, and the reason is arithmetic.

National output has been treated as a function of working-age population and productivity per worker. That identity is not a convention. It underpins growth forecasts, pension solvency models, central bank projections, and much of the economic case for immigration in ageing societies. It holds because output has always required people. Digital labor is the first productive capacity that generates output without a corresponding worker, and if it reaches a material share, the identity does not become less accurate. It stops describing the system. Measured productivity per worker rises in ways that no longer mean what the statistic was built to mean, and demographic pessimism about growth may turn out to be wrong for reasons that have nothing to do with demographics.

The fiscal consequence follows immediately and cuts both ways. Income and payroll taxes are levied on human labor, so output growing while the labor share of producing it falls erodes the base fastest in the most productive sectors. Against that, agents produce output without accruing pension or healthcare liabilities, which for an ageing economy is materially positive and may be the only available answer to a demographic problem with no politically viable alternative. Both effects are real, they run in opposite directions, and which dominates is jurisdiction-specific and unresolved. The asymmetry worth noting is that the states with the most acute demographic motive are largely not the states moving fastest.

This is the point at which digital labor stops being a private matter. Once a meaningful share of economic output is produced by something that is not a person, governments need to see it, count it, attribute it to a jurisdiction, and eventually tax it. Those requirements are what generate everything in the rest of this paper. The enterprise does not get to settle these questions and will be operating inside whatever answers emerge.

Questions:

• Which of our long-range plans assume output and headcount move together?
• If the tax or reporting treatment of AI-produced output changed materially in one of our major jurisdictions, what would be affected and how quickly could we respond?
• Have we written down a position on any of this, or are we waiting to be told?

6. Registration, and the Precedent We Already Set

• "It is the first time a non-human entity got this position, and it was a very interesting idea five hundred years ago."
• "The main liability is with the company."

If governments need to see and attribute non-human output, they need something to attach it to. Registering AI agents as entities sounds novel until it is placed against the precedent. Roughly five hundred years ago we created the corporation: a fictitious non-human entity with an identity, an accountable existence, full traceability, obligations under regulation, and the capacity to sue and be sued. Its people, products, and name can all change while the entity persists. Agents would be the second instance of a registered non-human entity, and most of the legal machinery already exists.

The reason to do it is not primarily fiscal, which is where most enterprise commentary has gone wrong. Human mobility between firms is a significant part of why economies grow, because experience circulates. An expert agent developed inside one institution never leaves its boundary. Registration is what would let accumulated agent experience move through an economy rather than remain captive. Identity, wallet, and auditability are the mechanics. Circulation is the purpose. Taxation follows from having built the mechanism rather than being the reason to build it.

Liability is less unresolved than it is usually presented. It sits with the company, as with a human employee. Much of the public debate conflates liability with a wish that someone feel remorse, which has never been a legal category. Corporations pay fines, and an agent holding a wallet can be penalized directly. The directional consequence matters: registration moves liability away from the model provider and toward whoever deployed and registered the agent.

Questions:

• If our agents had to be registered as entities with identities and wallets, could we produce the list today?
• Do we treat our agents' accumulated experience as captive to our boundary, and have we examined that?
• Where do we believe liability for agent action sits, and would our general counsel agree?

7. Three Tiers, and the License as the Lever

• "The country that created the experienced labor will gain the tax, not necessarily the country where the company is registered."
• "You cannot get a license for the bank unless your labor is sitting under this infrastructure."

The national picture is resolving into three positions rather than a spectrum. At the base sit hardware creators, where no country is independent and realistically few can contest the position over two decades. Above them sit sovereign labor powers: states that build the full stack, accumulate operating experience early, and export governed labor as a national product. Below sit renters, who moved too slowly to develop experienced labor domestically, must lease it, and inherit a dependency with a kill switch attached. The advantage in the second tier goes to smaller, centrally directed states rather than the largest economies, because concentrated decision rights and fast regulatory change let them act as sandboxes. Cheap energy and physical space decide which succeed.

The enforcement mechanism is not new law. It is licensing, and it is already in use. Banks site core systems domestically because no government accepts a foreign kill switch on critical infrastructure. The same logic extends to labor: no banking licence unless the labor runs on domestic infrastructure. The argument applies equally to hospitals, energy, and telecommunications, and it requires no novel legal theory.

The fiscal consequence inverts the usual concern. Corporate tax optimization gets easier when no human employees anchor a location, so the incentive to relocate intensifies. But if the labor itself must physically reside somewhere, revenue follows the labor's domicile rather than the company's registration. That makes hosting experienced labor a revenue strategy rather than a cost, and it is a materially different result from the one most fiscal commentary assumes.

Questions:

• Which tier are we effectively operating in, and did we choose it?
• Which of our regulated entities could face a domestic labor-siting requirement, and how far are we from meeting one?
• Where does our most experienced agent capacity physically reside, and what would it cost to move?

8. The Argument Nobody Has Won

• "We do not take associates any more, because it is faster to work this way."

One question runs underneath everything above it and has not been settled by anyone.

The case for preparing rests on assuming that agents will exceed human performance across most knowledge work. The reassurance usually offered on employment is that new roles will emerge, as they have in every prior transition. Both positions are held simultaneously and comfortably by most leaders, and they do not sit together.

The objection is specific rather than speculative. Every previous transition worked because the interval between a new kind of work appearing and the technology becoming capable of it was long enough for a generation to move into it. That interval is narrowing. If it narrows to less than the time required to retrain someone, the mechanism that resolved every prior transition stops functioning, and the evidence that it is already narrowing is visible at the entry level, where senior practitioners in several fields have stopped hiring juniors because working directly with these tools is faster than supervising a person.

If the objection holds, the consequence is not a harder transition. It is that there is no transition, because a transition has an end state and this would not. Reskilling as a bridge becomes income support as a permanent condition, which is the same instrument at a completely different scale and a materially different fiscal proposition. Nothing in the current policy conversation is scoped for that.

This should be held open rather than resolved, because the honest position is that nobody knows. What can be said is that the prior transitions took generations and were narrower, this one is being discussed in years and reaches most of knowledge work, and the people most confident that new roles will emerge are generally not the people who have looked closely at how quickly the interval is closing.

Questions:

• Do we hold both positions at once, that agents will exceed human performance and that new roles will absorb the displacement, and has anyone senior been asked to reconcile them?
• How long does it take us to retrain someone, and how does that compare with how quickly the work we would retrain them into is changing?
• If the reskilling case does not hold, what does our workforce plan become, and who would be responsible for saying so first?

What This Adds Up To

The framing of AI as a technology problem has survived longer than it should have, because it is the comfortable version. It puts the difficulty inside a function that can be resourced and held accountable, and it implies the problem is bounded.

Digital labor breaks that framing. The hard questions are what a manager is for once coordination is handled, how value gets recognized when the gain arrives as capacity rather than cost, what an entity is when it can act and be liable, where labor is domiciled when it does not physically go anywhere, and whether the mechanism that absorbed every previous wave of displacement still works. Not one of those is a technology question, and not one of them has an owner in a typical enterprise.

The most valuable workforce may no longer be defined by the people an organization employs. It will be defined by how deliberately it combines human judgment with digital capacity, and by whether it decided that combination or simply arrived at it.

© 2026 Executive Technology Board. All rights reserved. This document is the proprietary work product of the Executive Technology Board and is intended for the use of board members and authorized recipients. No part of this document may be reproduced, distributed, quoted, or republished, in whole or in part, without the prior written consent of the Executive Technology Board.

Executive Technology Board (c)