The Workforce That Stays

I've been sitting on this piece longer than I sat on other Blueprint instalments. Most of what I've spoken about so far has been about the firm, the logic of why firms exist, the boundary of the team, the way agentic systems push that boundary around. These subjects are interesting because they're structural; it lets you talk about institutions in the way that political economy lets you talk about states. What I want to consider here, is the workforce on the other side of that boundary. Not in aggregate. Not as a labour market. As the people, who are about to find out what the firm thinks of them.

There is a phrase I keep hearing in these conversations. AI won't replace people, but people who use AI will replace people who don't. It is offered as leadership in roughly every executive forum I've sat in over the last three years. It functions, almost without fail, as a threat in a friendlier voice. It tells you something about how the conversation around AI workforce transformation has gone wrong before it has started, because what it lets the speaker do is sidestep the asymmetric, awkward, costly fact at the centre of the actual work: that this transition will not be experienced symmetrically by the people inside it.

The frame I think we ought to be working from is this. AI workforce transformation, taken seriously, is a renegotiation of the labour contract inside the firm. That contract has implicit terms. You will be valued for what you know. Your career will accumulate. Your function will be defined enough that you can become expert in it. We will train you. We will pay you in some predictable relationship to your tenure and your scope. We will not surprise you. Each of those terms is up for question at the same time, and the discourse hasn't fully reckoned with that.

You can see the problem most clearly in the asymmetry of impact. The flattering version of the AI workforce conversation says we're all going to be augmented; the doomer version says we're all going to be replaced. Both miss the point. The transition isn't symmetric across the workforce or across roles. Some functions compress so sharply that the firm probably won't notice when it stops hiring into them. Some functions expand because orchestration of agents creates leverage in a way that wasn't possible before. Some functions emerge that didn't previously exist, such as workflow architect, agent supervisor, model evaluator, and they're being filled, today, by people who happen to have the right disposition rather than the right credential. The cost of pretending the impact is symmetric is borne, eventually, by the workforce that watches its peers shuffle through the asymmetry and concludes, correctly, that the firm has been less than candid with them.

Recent announcements have made the framing plain. Standard Chartered confirmed this week that it will cut 7,800 back-office roles by 2030 in pursuit of AI-driven efficiency, with its chief executive Bill Winters explaining that this was “not cost-cutting” but “replacing, in some cases, lower-value human capital with the financial capital and the investment capital we're putting in.” Lower-value human capital. That phrase is the asymmetry given voice at the level of the press conference, and it is, the speaker presumably believed, the strong version of the framing. Meta, is beginning its ten-percent reduction (around eight thousand roles) and has been reorganising teams into AI-focused pods and inventing new role categories called AI builder, AI pod lead, AI org lead, while Mark Zuckerberg explains publicly that “projects that used to require big teams now be accomplished by a single very talented person.” Each of these is a sentence the speaker thought sounded like leadership. However, they land inside the firm as a description of the asymmetry, and of where the speaker has decided the firm sits in it.

There is a more uncomfortable version of this argument we also have to name. The roles most exposed to compression in this wave are not the ones the public conversation focused on three years ago. They are roles in the middle of the knowledge‐work distribution, the ones that absorbed new graduates and let them learn by doing. The junior analyst. The trainee solicitor. The entry‐level consultant. The first‐year researcher. The work those roles did, is the work agentic systems do most readily, because it tends to be bounded, structured and observable. Which means the apprenticeship layer of professional life is being thinned at exactly the moment the senior layer is being asked to do more. We are eating the seed of professional development without quite admitting to ourselves that's what we're doing.

I don't have a clean answer to this. I notice that the firms I respect most on the question are starting to invest in deliberate apprenticeship structures, paid time on synthetic problems, supervised work alongside agents, structured coaching from seniors. Other firms are pretending that the absence of the entry‐level layer is itself a productivity win. I think we'll look back in a decade and find that the firms that handled the apprenticeship problem well (the AI native class by the way) were the ones that took the long view of their workforce as an institution rather than a cost line. But that's a directional belief; I can't prove it yet.

What I can say with more confidence is this. The binding constraint on whether AI workforce transformation works at the level of the firm, is trust. Not capability, not platform, not budget (although they do matter). The workforce reads what you do, not what you say. It reads how you treat the people whose roles compress as carefully as how you treat the people whose roles expand. It reads whether the new role descriptions and performance frameworks land coherently, or whether HR is improvising in public while leadership is on a roadshow. It reads whether the bad‐day incident, the agent that did something embarrassing, the workflow that fell over, gets handled with the calm structure of a serious institution or the defensive theatre of one that wasn't ready.

Trust compounds very slowly and discounts very quickly, and the asymmetry between those rates is most of the operational point. A transformation that loses trust early spends the rest of its life trying to rebuild what could have been preserved by being honest in the first month.

You can watch this play out, in real time and in public, across some of the most observed firms in technology. At Block, Jack Dorsey has mandated AI use across the workforce, integrated AI fluency into performance reviews, and requires employees send weekly emails listing their five most recent accomplishments, which he then summarises using generative AI. The internal response has been a mix of compliance and quiet exhaustion. “Top‐down mandates to use large language models are crazy,” one employee told Wired. “If the tool were good, we'd all just use it”. At Amazon, more than eighty percent of developers are now required to use AI tools each week, with their token consumption tracked on internal leaderboards. The workforce has responded by coining a practice they call tokenmaxxing, running the in‐house agent platform on unnecessary tasks for the sole purpose of inflating their scores. It’s textbook Goodhart's Law. The measure becoming the target and the target ceasing to measure what it was supposed to. The deeper point, though, is the trust one. When mandate and surveillance replace design and honesty, the workforce learns to game the metric while quietly losing faith in the institution. The productivity claims start to look like the dashboards that produced them, confident, large, mostly performative. The gap between what the firm believes about its AI deployment and what is actually happening inside it widens every month.

Two things follow.

The first is that the language we use about AI inside the firm matters more than it appears to. The casual habit of saying the AI decided, or the AI is going to take that on, or the model is responsible for this output, is the language of moral evasion. Tools do not carry moral weight; people do, and institutions do. When a workflow includes an agent that issues a credit decision or screens a CV or routes a customer, the agent is not the decider. The leader who designed the workflow is the decider. The operator who deployed the agent is the decider. The board that approved the deployment is the decider. Every consequential outcome the system produces is owned, ultimately, by a human. That has to be visible in the policy, in the product copy, in the disciplinary framework, and in the way the firm explains itself to its regulator. Lazy framing here will produce, in time, a generation of decisions for which no one feels accountable, and the discovery, when something serious goes wrong, that the accountability vacuum is itself the failure.

The second is that the modular, deliberate, sequenced approach to workforce transformation isn't a slow approach. It is the only one that holds trust together. The temptation, particularly under pressure from a board or a competitor or a quarterly cycle, is to deploy fast and figure out the people work later. The pattern this generates is invariant: the platform lands, the workforce doesn't absorb it, governance scrambles, an incident occurs, the programme halts. The decade of advantage that AI workforce transformation could produce gets consumed by the eighteen months you lost handling the avoidable incident. Modules, bounded units of work with defined supervision, instrumented outcomes, and named owners, are how you deploy at pace without losing the constraint that makes pace worth pursuing.

A friend who runs operations at a large institution said to me recently that he had spent more time in the last twelve months on the cultural side of AI deployment than on anything else. A lot of the time it feels like we assume that the technology is the hard part of transformation. With AI, the technology is the cheap part. The hard part is everything downstream of it. The workflows that have to be rewritten, the roles that have to be redrawn, the performance frameworks that need to catch up, the comp model that needs to recognise where value has moved, and the workforce, the actual people, who need to be told the truth about what is being asked of them.

What we owe each other through this transition is what we always owe each other in moments of institutional change: candour, structure, and the discipline to design rather than improvise. We owe the workforce that stays the explanation of what their role is becoming and the support to grow into it. We owe the workforce that doesn't stay the kind of treatment that the workforce that does will recognise as decent, because the latter is watching the former, and the trust account is single, not separate. We owe our regulators and our boards a governance posture that is principles‐based, embedded in workflows, and owned by the operators rather than outsourced to a committee. And we owe ourselves, as the people running the work, the honesty to admit when our framing is comfortable but evasive, and to choose the framing that is actually true.

This isn't a productivity argument. There is a productivity argument, and it is real, and I think it is larger than the public conversation has grasped. The productivity argument arrives, or fails to arrive, through the people inside the firm. The firms that will own the next decade are the ones that take that point seriously enough to design the transition rather than apologise for it. I don't know how this will land in any given firm. I do know that the leaders I've watched most carefully through it, share a discipline. They tell people the truth. They sequence the work. They build governance into the workflow rather than around it.

Part of the Blueprint AI series, this post reflects my personal views and is independent of my professional role.