The Human-Centered Alibi

There is a genuine finding underneath the consultancy language: enterprise AI programmes fail more often for organisational reasons than for model reasons. Adoption stalls when people cannot trace a recommendation, when the system does not know internal vocabulary, when nobody can say which human is accountable for the output. That is not marketing. It matches what anyone who has watched a deployment from the inside already knows.

The danger is that a correct diagnosis is also a portable one. "Humans in the loop" is the single most transferable sentence in the current market, because it can be attached to almost any programme without changing the programme. It absolves, it reassures the board, it satisfies a procurement checkbox, and it costs nothing to say. The techno-fabulists are not bad at this argument. They are the best at it, because they have the most to gain from being trusted while consolidating.

So the task is not to argue against human-centred AI. It is to make the phrase falsifiable — to write down what a firm would have to be doing differently if the claim were true, so that the claim stops functioning as an alibi.

The current leadership playbook, as circulated in early 2026, has four moves: senior leaders should personally use the tools rather than delegate the vision; roles should be redesigned as routine work shifts to agents; management layers should be reduced because agents widen a manager's span of control; and human leaders should retain responsibility for strategy, ethics, and hard calls.

Three of the four are defensible on their own terms. Leaders who use the instrument understand its failure modes. Role redesign is honest about the fact that the work changes. Retained human accountability is the only version of this that survives contact with a regulator.

The fourth is a load-bearing wall dressed as a productivity note. "Flatten structures" means fewer people between a decision and its execution — which is also fewer people positioned to say no. Span of control is not a neutral quantity: middle management, for all its cost, is where an organisation historically stored its friction, its institutional memory, and its dissent. Remove the layer and you have not only removed overhead; you have removed the mechanism by which the third principle — preserved human judgement — was actually exercised. Principle four quietly defunds principle three.

Note what happens to the sequence when it is stated plainly: the humane principles justify the adoption, the structural principle harvests it, and the accountability principle is left with no organ to act through. That is not hypocrisy. It is an unexamined interaction between two recommendations that were written by different concerns.

The vendor version of the humane strategy usually resolves to four principles: contextualisation, collaboration, transparency, and continuous learning. Each is a real engineering requirement. None of them is comprehension in the sense used in the companion essays.

Contextualisation means the system knows your product names, your workflows, and which internal expert has solved the problem before. That is retrieval quality. It reduces irrelevance. It does not tell you whether the underlying claim about the world is true, whether the benchmark survives a tuned classical baseline, or whether the biological target is a distribution rather than a structure.

This is where a knowledge-graph vendor and a frontier-lab founder can use the same word to mean two different sizes of thing — and where a buyer can be fully satisfied on human-centredness while remaining short on all three comprehensions. A firm can have excellent internal expert routing and still be mis-pricing its roadmap by two years.

Keep the two audits separate: contextualisation asks whether the answer is relevant here; comprehension asks whether the answer is true anywhere.

The test of a humane AI strategy is whether any of its commitments is expensive to fake. The following six are. Each has a documentary answer, a date, and a party who can be held to it.

First: name the last decision a human overrode, with the date and the person. A firm that has preserved judgement has a log of refusals. A firm that has only claimed it has none, because refusal leaves paperwork and alibis do not.

Second: after flattening, who now carries escalation? If the answer is a channel rather than a named role with time in their week to use it, the layer was removed and the function was not rehoused.

Third: show the training spend per redesigned role, and how it moved year over year. Role redesign without a learning line is headcount reduction with better vocabulary.

Fourth: who defines the outcome in the outcome-based contract, and who measures it? If those are the same party, the model has transferred risk in the slide and retained it in the ledger.

Fifth: what does the system refuse to answer, and where is that list published? A model with no published refusal surface is a model whose limits were never characterised — which is the operational signature of sycophancy, priced elsewhere on this site.

Sixth: which claim in the last public roadmap has since been retired, and on what evidence? A firm that has never retired a claim is not being consistent; it is not measuring.

There is a market reason to want this audit, and it is not reputational. When a claim is free to assert, it stops carrying information, and the firms that actually did the work lose their ability to charge for it. Cheap humane language is a commons problem: every unearned assertion lowers the value of every earned one, until buyers discount the whole category and price on demo quality instead.

A firm that has genuinely rehoused escalation, funded role redesign, published a refusal surface, and retired a roadmap claim has assets that a fabulist cannot manufacture in a quarter. Its interest is in a screen sharp enough to separate it from the imitation. The audit is not a constraint on the honest operator; it is the only mechanism by which the honest operator gets paid the difference.

Which reverses the usual posture. The people best served by making "human-centred" falsifiable are not the critics. They are the firms that meant it.

This is not a claim that any specific firm, consultancy, or vendor is acting in bad faith. The documents cited are ordinary professional output and the diagnoses in them are largely correct. The argument is about an interaction between recommendations, and about the asymmetry that lets a correct sentence be borrowed by a programme it does not describe.

Nor is it a claim that flattening is always harmful, that agents cannot widen a span of control usefully, or that consolidation is the destined outcome of AI adoption. Nothing here is ordained. What is claimed is narrower and checkable: remove a layer without rehousing its refusal function and you have reduced the organisation's capacity to say no, whatever the strategy deck says about human judgement.

If the audit questions above are answered well by firms that flattened aggressively, the concern in this essay is misplaced and should be withdrawn.

The World Economic Forum published the cleanest version of this argument in January 2025: leaders should adopt a "double bottom line" that weighs societal and individual empowerment alongside profit, because AI "could either uplift humanity to new heights or exacerbate old inequities." The piece is honest about the pressure it is written against — a majority of CEOs name AI a top priority and roughly seventy percent are already investing at scale, with profitability, efficiency, and performance the stated drivers. It cites the IMF figure that forty percent of global jobs are exposed to AI, and answers it with reskilling, public-private partnership, and regulation on bias, privacy, and transparency.

Every one of those recommendations is defensible. The structural problem is that only one of the two bottom lines has a reporting standard. Profit is audited quarterly, in a unit that everyone agrees on, by a party who is not the company. The second line — well-being, empowerment, community resilience — has no comparable unit, no cadence, and usually no external measurer. When one term in a two-term objective is measured continuously and the other is measured rhetorically, the optimiser does not split the difference. It follows the gradient it can actually see. That is the same failure the companion essays price in reward-function terms: an unmeasured objective is not a weak constraint, it is an absent one.

The forward-looking employment numbers make the point sharper. The essay reaches for a projection of 97 million new AI-related jobs — a figure whose horizon has now passed, which is exactly the kind of claim the sixth audit question exists to retire. The IMF exposure figure is a measured estimate of jobs at risk; the job-creation figure was a forecast. Netting one against the other, as strategy decks routinely do, is not analysis. It is an accounting move that lets a displacement cost be offset by an unaudited receivable.

None of this makes the double bottom line wrong. It makes it incomplete in a specific, fixable way: give the second line a unit, a cadence, and an outside measurer, or expect it to lose every time it competes with the first. The concrete instances the piece cites — a regulatory sandbox, an open-weights foundation, a skills partnership with a target and a date — are the parts of the argument that already have those properties, which is why they are the parts worth copying.