The Intentional Organisation - Issue #53 - The Blind Spot
👉 HR is measuring its own AI transformation with instruments that cannot see what actually moved.
1. The Blind Spot
The standard HR read on AI is by now familiar. You commission an impact analysis and it comes back role by role: this job loses these tasks, that one gains these tools, headcount holds, a handful of hours are saved per person per week. The picture looks under control. The trouble is that the picture is drawn at the wrong scale.
What AI actually does to Work does not happen inside the job. It happens between jobs, and across the edges of the organisation. Production is a sequence of steps that firms continually re-bundle — some done by people, some by machines, some pushed outside the firm altogether. Work shifts to the customer who now self-serves, to the machine that now drafts the first version, to the competitor whose platform quietly absorbs a function you used to own. A job-level analysis is structurally blind to all of this, because the unit it measures — the job — is not the unit that is moving.
You can already see the gap in the data we do have. A Berkeley study of how AI actually lands on knowledge workers found that it tends to intensify work rather than reduce it: scope widens, expectations rise, the day stretches, and 83% of those studied reported a heavier load rather than a lighter one. A job-level analysis that records “four hours saved” and a system-level reality of “workload up and climbing” are describing the same deployment. Only one of them is true to the people living inside it.
This is the first half of HR’s blind spot: the wrong unit of analysis. The second half is the wrong default question.
The wrong question
Most people analytics begins with the data. What patterns are in our engagement scores, our attrition curves, our usage logs? It is the natural posture of a function that has finally been handed dashboards. But people analytics earns its keep only when it changes a decision, and starting from the data rather than from the decision is the field’s oldest failure. The right opening question is not “what patterns exist in our data?” but “what decision needs to be better, and what would we need to know to make it?”
AI makes this worse, not better. A system that finds patterns at near-zero cost will hand you an infinite supply of them. The scarce resource was never the pattern; it was the clarity about which decision the pattern is meant to serve. Expand the supply of answers and the absence of a question becomes the binding constraint.
What the instruments cannot see
Put the wrong unit and the wrong question together and you get a function that measures tasks and time-saved with rising precision while missing the thing that actually moved: accountability.
Back in 2022 I argued that accountability is not responsibility. Responsibility can be delegated through a clean process, but accountability can only be personally accepted. AI introduces a harder version of the same problem for the decision-making processes HR is supposed to govern. When an automated process mediates a decision and then fails, accountability does not disappear; it relocates to the nearest human — regardless of whether that person had any real control over the outcome. Madeleine Elish calls these moral crumple zones: the human in the loop absorbs the blame the way a car’s crumple zone absorbs an impact, to protect the integrity of the system.
That is the displacement HR’s instruments are not built to detect. A role can look untouched on a job-level dashboard — same title, same tasks, a few hours saved — while quietly becoming an accountability sink — Dan Davies’s term, in The Unaccountability Machine, for the place in a system where responsibility drains and pools — for decisions a model now makes and no one formally owns. The 2022 problem was that accountability is hard to give away. The 2026 problem is that it can be loaded onto someone who never accepted it and cannot see it coming.
A clear example is the many instances of automated processing in Rewards. Letting an LLM learn reward practices from past decisions risks deploying the same biases that modern processes are designed to avoid. And if it does, who is accountable?
The honest counter-argument
The strongest objection to all this is that the job is exactly the right unit, precisely because the job is where everything attaches: employment law, reward, headcount, skills, the contract itself. “Work shifting across boundaries” is real but too diffuse to manage, and the answer is not to abandon job-level analytics but to mature it. On this reading, AI in HR is only as strong as the people-analytics foundation beneath it, and the discipline is becoming exactly the capability that lets an organisation decide, deliberately, where human judgement stays in control.
I think the maturity case is half right, and the half it gets wrong is the dangerous one. Maturing your measurement of the wrong unit does not close the blind spot — it measures the wrong thing more precisely, and lends it more authority. A more sophisticated job-level dashboard is still a job-level dashboard.
The capability HR is building is real; it is simply pointed at the question that is easiest to answer rather than the one that matters.
What this looks like in practice
In Issue #52 I described the “Gioia for HR” pilot at Campari — a digital twin built with the HRBPs who actually do the work, on the principle that every AI use case is designed with today’s job holder. The measurement discipline is the same move turned on the analysis itself: before asking what the data shows, we try to name the decision the use case is meant to improve, and who stays accountable for it once the model is in the loop. It is harder than commissioning another dashboard, and we are early…. But it changes what you look at — not hours saved, but accountability located.
**The question worth asking**
Issue #51 named the governance gap]; Issue #52 made it human. This issue is about the instruments themselves. HR cannot redesign what it cannot see, and right now it is measuring its own transformation at the wrong scale, opening with the wrong question, and missing the one variable — accountability — that AI is quietly moving around the system.
Some are already redrawing the mandate: Richard Rosenow argues that HR is really a regulatory control function, whose real job is to decide where the variability of a workforce gets absorbed — and the accountability sink is what forms when that job goes undone. That is the operating-model question Issue #54 takes up. Before then, one question is worth sitting with.
The next time an AI-impact analysis lands on your desk and reassures you that the jobs are holding and the hours are saved — what is it not showing you, and who is now accountable for the part it left out?
Sergio
References
Elish, M. C. (2019). Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction. Engaging Science, Technology, and Society, 5, 40–60. https://doi.org/10.17351/ests2019.260
Demirer, M., Horton, J. J., Immorlica, N., & Lucier, B. (2026). Chaining Tasks, Redefining Work: A Theory of AI Automation (NBER working paper).
Davies, D. (2024). The Unaccountability Machine: Why Big Systems Make Terrible Decisions — and How the World Lost Its Mind. Profile Books.
2. Site Updates
I wrote a new Essay : Organisational Grammar: The Atomic Elements of Designed Coordination — introduces the six atomic elements (Role, Skill, Group, Decision, Obligation, Accountability) as the underlying vocabulary of any organisation design work. If this issue’s argument — that the skill is the right unit for Work Design decisions — resonates, the Grammar piece is the structural companion.
The Laws of Organisation Design series continues on the blog: the Law of Requisite Variety and the Law of Alignment are both up. The Laws series is the slower, more structural companion to this newsletter’s arc.
3. Reading Suggestions
‘Can a machine do this job?’ is the wrong question — Carl Benedikt Frey (Financial Times). The sharpest statement of this issue’s first move: AI rarely deletes a job outright, it shifts the work to the customer who now self-serves. “Can a machine do this job?” is the wrong question; “can the customer do without it?” is the real one.
The goal isn’t an AI-native organization, it’s an AI-legible one — Iain Roberts. AI fails where authority, accountability and process are not *legible* — machine-readable. A useful mirror to this issue: you cannot measure what you have not made legible, and accountability is the first thing to blur.
When Everyone Uses AI, Companies Risk Losing Critical Skills — Charikleia Kaffe et al. (BCG). The capability cost of the residual-task problem: lean too hard on AI and judgement quietly atrophies. The antidote is design — workflows that keep people challenging the output rather than rubber-stamping it.
Every hire is a bet made in the dark — Andrew Marritt (Working Ideas). A long, careful piece on the irreducible uncertainty in hiring. Marritt is the analytics voice behind this issue’s “what decision needs to be better?” — here he shows what taking a single decision seriously actually looks like.
Are Apprentices an Endangered Species? — Luis Rayo (Kellogg Insight). If AI absorbs the basic tasks juniors once learned on, where does the next generation of judgement come from? The accountability question, pushed one career-stage upstream.
4. The (un) Intentional Organisation 😁
5. Keeping in Touch
Don’t hesitate to reach out by directly hitting “reply” to this newsletter or using my blog’s contact form.
I welcome any feedback on this newsletter and the content of my articles.
Find me also on: