AI workforce transformation cannot be delivered as a posture. It has to be delivered as an institutional architecture, and the firm that gets the architecture wrong will lose the trust of its workforce whatever its leaders say.

The architecture is the harder of the two problems and the less discussed. The visible problem is candour, whether the leadership tells its workforce the truth about what is being asked of them. The deeper problem is whether the institution underneath is capable of delivering on the rhetoric. This piece is about delivering on these promises.

The first thing to prevent is fragmentation. If every division of a large firm is allowed to redesign jobs in response to AI on its own timetable, with its own principles, the firm ends up with an inconsistent job architecture across divisions inside eighteen months. The marketing function in one business unit decides that AI augments analysts and reduces the need for managers. Another decides AI augments managers and absorbs the analyst layer. The two are not reconcilable, not internally consistent, and not legible to a workforce that has to make career decisions across them. The cost of this fragmentation is not theoretical. It shows up in talent mobility, in compensation consistency, in the integrity of the performance framework, and in the firm’s ability to redeploy people across the businesses without renegotiating from scratch every time.

The instinct in most firms is to address this with policy. Issue a set of principles, distribute them through HR, trust that divisions will interpret them consistently. The instinct does not work. Principles without architecture get translated unevenly by people with different incentives, different timelines, different reading of what the principles actually require. The variance compounds. What you need is not better principles. You need a design authority, a function that has the right and the responsibility to refuse a job redesign that does not fit the enterprise architecture, even when the division wants to proceed.

Most firms do not have this function. They have HR, which interprets workforce policy but does not own architecture. They have a CTO function, which owns technology architecture but not workforce. They have transformation offices, which run programmes but do not hold standards. The function I am describing is the workforce equivalent of an enterprise architect, someone who can say no to a divisional redesign that breaks the firm’s consistency, and who has the seniority and the analytical depth to make the call defensibly. The discipline of enterprise technology architecture took shape in the 2000s and 2010s after a decade of fragmented application deployments produced inconsistencies that became expensive to unwind. The workforce equivalent is at the point of emergence now, and the firms that build it early will avoid a structurally similar unwind in their job architecture five years from now.

The standard rebuttal to centralised authority is that it cannot move fast enough to be useful. Divisions are closer to the work, they know their populations, they need to act on their own timelines. The rebuttal is correct as far as it goes. What it misses is that the question is not whether centralised authority is faster than distributed action. The question is whether the firm wants a coherent job architecture in five years’ time. If it does, the architecture has to be set centrally, and the execution has to happen federally.

This is the structural argument for what I will call architectural federation. I prefer the term to hub and spoke, which is the more common shorthand, because hub and spoke describes a topology while federation describes a relationship. The federation is one in which authority is shared, standards are common, and execution sits with the local function rather than being pulled into the centre. That is the institutional form the AI workforce transition requires. Within it, the hub does three things. It sets principles, it provides patterns, and it adjudicates exceptions. The spokes, the divisional and country people partners who actually advise the business, do the work of applying the architecture to specific populations. Neither layer can do the other’s job. A hub that tries to execute against every divisional redesign becomes a bottleneck. A spoke that tries to set its own principles produces the fragmentation I described above.

The interesting structural fact about this model is that the hub is small. Architecture is a thinking function, not a doing function. The doing happens in the spokes. A hub of fifteen people, deployed correctly, can hold the architecture for an enterprise of seventy thousand. The doing capacity comes from the federated network. The leverage in the model is that the federated network has been there all along, every large firm already has business people partners, country people partners, and divisional HR leads, and the question is whether they are working off a coherent architecture or improvising in parallel.

Trust has an architectural shape worth setting out directly. Workforce trust compounds slowly and discounts quickly because the workforce reads the firm’s actual behaviour over time, across many small decisions, and constructs from those decisions a model of whether the firm means what it says. Posture cannot hold this. A leader’s speech can shape expectations for a quarter. The way the firm handles the next twelve job redesigns will shape them for a decade.

This is why architecture matters more than rhetoric. If the firm has architectural standards for job redesign, clear principles about what AI augments and what it automates, clear guardrails on how productivity gains are recognised in compensation and capacity, clear escalation paths when a redesign is contested, then the workforce can read those standards and predict, with reasonable accuracy, what is going to happen in their own case. Predictability is what trust actually rests on. Without architecture, every division’s behaviour becomes a separate signal, and the workforce ends up extrapolating from the most aggressive case. This is why the federation has to be architectural and not merely coordinated. Federation without shared standards reverts to a set of divisional fiefdoms with a coordinating function pasted on top, which is what most firms have today, and what the workforce already reads accurately.

The same logic applies to governance. When something goes wrong with an AI-augmented workflow, and it will, in every firm, repeatedly, the workforce reads whether the response was institutional or improvised. An institutional response requires that the architecture had a place for the incident before it happened. Improvisation, even when it is well-intentioned, signals that the firm did not think the question was important enough to design for in advance. The architecture has to include the escalation framework, the human oversight model, the accountability chain, and the disclosure norms, because each of these is a place where trust can be preserved or lost depending on how it was set up before the incident occurred. The EU AI Act will codify parts of this for HR and people decisions over the next year. The firms that will absorb the regulation calmly are the ones that already had the architecture in place. The firms that will struggle are the ones that planned to bolt governance on after the fact.

The apprenticeship problem belongs in the same frame. It is one of the clearest cases where divisional incentives produce a firm-level cost that no single division can solve, and where the architecture has the most decisive leverage.

The reason firms do not solve the apprenticeship problem on their own is not that they do not understand it. Most senior leaders I speak to understand it perfectly well. The reason it does not get solved is that the cost of preserving the apprenticeship layer falls on the division that hires juniors, and the benefit of preserving it accrues to the firm as a whole, several years later. The division has no rational incentive to absorb that cost unilaterally. Without architectural intervention, the apprenticeship layer thins, division by division, even when every leader involved would prefer that it did not.

The architectural intervention is not complicated to describe. The firm sets a principle that says junior hiring will be maintained at a level that preserves the institutional pipeline, defines what that level is empirically, and holds divisions to it through the same architecture that governs other workforce decisions. This is exactly the kind of decision that requires central authority because no single division can carry it on its own. The architecture exists to make the firm-level decision deliverable.

I do not want to overstate the case. The architectural intervention does not solve the apprenticeship problem on its own. It still requires deliberate apprenticeship structures, supervised work alongside agents, structured coaching from seniors, paid time on synthetic problems, which are operational rather than architectural. But the architecture is the precondition. Without it, the operational work happens in patches, in the divisions that happen to have leaders who care, and the firm-level outcome is what the worst-performing division produces.

I will try to summarise the architecture as concisely as I can.

It is a design authority for job and organisational decisions in an AI-augmented environment, organised as an architectural federation. It defines principles for task allocation between humans and AI with sufficient specificity that divisions can apply them without each interpreting them differently. It owns the productivity translation question, how AI gains are recognised in compensation, capacity, and investment, at the enterprise level, so the firm has a coherent answer rather than thirty divisional answers. It governs the apprenticeship pipeline as a firm-level asset, not a divisional cost line. It sets the accountability framework for AI-augmented decisions, with the human ownership chain made visible in policy, in product, and in regulatory disclosure. The federation is held by a small hub that sets the standards and adjudicates exceptions, and executed by the network of people partners who apply the standards to specific business contexts.

This is not, on the face of it, an exciting institutional form. It does not produce a launch announcement. It is, however, the form that lets the firm deliver on the commitments it makes to its workforce during the AI transition. Without it, the commitments are posture. With it, they have the institutional weight to be credible.

The workforce that stays is reading whether the firm is building this architecture. They will know, before the architecture is named, whether it is being built. They will tell each other, and they will draw conclusions, and those conclusions will harden into the operational reality of the firm’s AI transition long before any board paper acknowledges that this is what has happened. The architecture is the institution behind the candour. The firm that means it, builds it. The firm that does not, will be found out by the workforce well before it is found out by anyone else.