027. The Bottleneck that Thinks

T.D. Inoue’s recent piece “By the time you open your eyes, it’s too late“ crystallises something a lot of people are feeling: that AI capability is moving faster than the public conversation about it, and that by the time the implications land, the window for preparation may have closed. The concern is real. But the framing — and specifically, the displacement framing that dominates public discussion — may be pointing institutions toward exactly the wrong preparation.

The conversation about AI and employment has settled into a familiar shape. Frontier models demonstrate increasingly sophisticated analytical and creative capability. Headlines translate this into job displacement. Institutions begin planning for a future with fewer professionals. The logic feels obvious: if AI can do what knowledge workers do, knowledge workers become redundant.

This logic contains a category error that matters. It mistakes what the professional class does for what it is for.

Every complex institution — public or private, large or small — maintains a layer of people whose work is not reducible to any single task on their job description. Call them the professional cadre: the people who pull information together across organisational boundaries, build agreement under uncertainty, turn strategic direction into operational action, and navigate the political reality of getting things done in organisations where not everyone agrees.

This is not a new function. It emerged with institutional complexity itself, and it has survived every previous wave of technological change — mechanisation, electrification, computerisation, networking — because each wave changed the tools of coordination without eliminating the need for it. Filing cabinets gave way to databases. Memos gave way to email. Spreadsheets replaced ledger books. In each case, the execution layer transformed while the coordination function persisted, because institutions continued to need people who could hold competing priorities, exercise judgment under ambiguity, and bear accountability for decisions made with incomplete information.

The displacement narrative assumes this wave is different. That AI doesn’t just change the tools of coordination — it automates coordination itself. That judgment, consensus-building, and institutional navigation are, at bottom, information-processing tasks that sufficiently capable models will absorb.

This assumption deserves serious examination rather than reflexive acceptance.

Start with what’s observable. Frontier AI models can produce remarkable analytical output. They can synthesise large bodies of evidence, generate strategic options, draft communications, model scenarios, and surface patterns that human analysts might miss. None of this is in dispute.

Also observable is that almost none of this capability has been deployed in the way the displacement narrative requires. The gap between frontier capability and institutional adoption is not a temporary inconvenience — it reflects genuine structural barriers.

Deploying AI to actually replace professional judgment in an institutional setting can’t be done naively. It requires domain-specific model configuration, compliance integration, workflow embedding, audit trail infrastructure, liability frameworks, and — critically — the organisational change management to restructure decision-making around AI outputs rather than human judgment. This is what we’ve previously called “the Missing Middle“: the coordination infrastructure between raw capability and operational deployment that doesn’t yet exist at scale.

That missing middle is being built. Protocol standardisation, cloud orchestration, and agentic frameworks are developing rapidly. But the bottleneck isn’t primarily technical — it’s organisational. Integrating AI into institutional decision-making is itself cadre work. The people best positioned to manage this integration are the same professionals the displacement narrative says are becoming redundant.

In the meantime, what AI actually delivers to most institutions is richer analysis flowing into existing decision-making plumbing. More options surfaced. Higher-quality inputs generated. Faster analytical cycles completed. All of which increases pressure on the coordination layer rather than replacing it.

This is intensification, not displacement. The cadre doesn’t become redundant — it becomes the bottleneck. And a bottleneck under increased pressure either adapts or fails.

This is intensification, not displacement. The cadre doesn’t become redundant — it becomes the bottleneck. And a bottleneck under increased pressure either adapts or fails.

The best case against this thesis comes from transaction cost economics. The economist Ronald Coase penned the seminal works on this. In 1937 he argued that firms exist because market transactions are expensive — the costs of finding, negotiating, and enforcing agreements justify bringing coordination inside an organisation: the main job of the firm is to reduce the cost of turning work into commerce.

If AI dramatically reduces the cost of information processing and coordination, the Coasean logic suggests the internal coordination layer should thin. The cadre, on this view, is an expensive workaround for information friction that AI dissolves.

This is a serious argument, and some versions of it hold up. Financial services provides the clearest case. A Brookings analysis of how AI is reshaping work in finance documents the pattern: algorithmic trading, automated compliance screening, and AI-driven risk modelling displaced significant numbers of professionals whose work, on examination, was standardised execution dressed as judgment. Analysts who assembled data into templated reports. Compliance officers who applied checklists. Middle managers who relayed information between layers. When information processing became cheap, these roles thinned or vanished.

But look more closely at what happened in financial services and the picture gets more complicated. The roles that dissolved were execution roles — work that followed established patterns even when performed by expensive people with advanced degrees. The coordination roles — portfolio strategy, client relationship management, regulatory navigation, cross-functional integration — didn’t thin. They intensified. The work shifted from producing analysis to validating AI-generated analysis, from assembling numbers to interpreting what the numbers meant. The people doing genuine judgment work found themselves processing more inputs, managing more complexity, and bearing more accountability as the execution layer automated beneath them.

The Coasean argument treats the cadre as if it were a typing pool — an expensive way to process information that technology makes cheap. But the cadre’s irreducible elements aren’t information processing. They’re political judgment, consensus-building, accountability-bearing, and trust maintenance. These are not transaction costs to be minimised. They are institutional capabilities to be maintained.

The cadre’s irreducible elements aren’t information processing. They’re political judgment, consensus-building, accountability-bearing, and trust maintenance. These are not transaction costs to be minimised. They are institutional capabilities to be maintained.

A typing pool dissolves when typing becomes automated because typing is execution. The cadre persists because coordination involves irreducible social and political work. Getting a procurement decision through an organisation isn’t an information problem — it’s a problem of competing priorities, risk appetites, and institutional politics that no amount of analytical sophistication eliminates. AI can surface better options and model their consequences more thoroughly. It cannot — not yet, and not in the foreseeable deployment environment — navigate the human landscape in which those options must be chosen, championed, and implemented.

Here is where the institutional stakes become acute.

If the displacement narrative were correct — if AI were genuinely absorbing coordination and judgment — the rational institutional response would be to reduce headcount in professional roles, invest in AI infrastructure, and redesign workflows around automated decision-making. Many institutions are beginning to do exactly this, or signalling their intention to.

But if the actual near-term dynamic is intensification rather than displacement, this response is precisely wrong. It hollows out the coordination capacity that institutions need most, at the moment when increased analytical throughput demands more judgment, not less. It removes the people who integrate across boundaries, build agreement under uncertainty, and turn decisions into action — not because those functions have been automated, but because the risk model says they’re about to be.

This is institutional self-harm, and it has a recognisable pattern. Institutions that mistake cost-cutting for transformation have done this before — offshoring, outsourcing, delayering — and the pattern is consistent. The short-term savings are real. The medium-term loss of institutional capacity is also real, and harder to measure until it shows up as failures of coordination, failures of judgment, and failures of adaptation that the thinned organisation can no longer absorb.

The cadre is simultaneously the group most pressured by the AI transformation and the group through which institutions navigate it. Degrading that capacity in anticipation of a displacement that hasn’t arrived — and may not arrive in the form expected — is preparing for the wrong emergency.

What does this look like in practice? There are signals that distinguish institutions navigating this well from those preparing for the wrong transformation.

The low road looks like treating AI adoption as a cost-cutting programme. Headcount targets set before workflow redesign. Professional roles redefined as “AI-augmentable” without examining what the role actually coordinates. Middle management layers removed on the assumption that AI flattens information hierarchies, without recognising that those layers were doing political and consensus work, not just relaying information.

The high road looks like investing in judgment and coordination capacity to match increased analytical throughput. Professionals trained not to compete with AI’s analytical output but to exercise better judgment over richer inputs. Decision-making processes redesigned to handle more options and faster cycles without collapsing into paralysis or default. Institutional knowledge — the tacit understanding of how things actually work, who needs to agree, and where the real constraints live — treated as infrastructure rather than overhead.

There is precedent for the high road working. When enterprise data warehousing matured in the 2000s, the Coasean prediction would have been straightforward: cheaper data means fewer people to process it. And the routine reporting layer did automate — standard reports that once required analysts to assemble could be generated on schedule without human intervention. But that wasn’t where the return came from. The value of data warehouses lay in the ad-hoc analysis they enabled: the ability of domain experts to interrogate operational and market data in response to shifting signals, test hypotheses, and build evidence for decisions that couldn’t be templated. The organisations that treated data warehousing as a cost-reduction tool got less than they paid for. The organisations that treated it as coordination infrastructure — investing in the analysts, subject-matter experts, and decision-makers who could use better data — got adaptive capacity that justified the considerable overhead of building and maintaining the system.

The parallel to AI is direct. Institutions that invest in AI primarily to reduce headcount are making the same mistake as institutions that built data warehouses primarily to automate reports. The technology’s real value isn’t in replacing the people who process information — it’s in equipping the people who exercise judgment with richer, faster, more comprehensive inputs. That only works if the judgment layer is there to use them.

The difference isn’t whether an institution adopts AI. It’s whether the adoption model assumes the cadre is a cost to be reduced or a capability to be developed.

None of this is a permanent claim. The missing middle is being built. Agentic AI systems are becoming more capable. The organisational integration barriers are being addressed, albeit more slowly than the capability barriers. It is entirely possible that in five or ten years, the coordination function itself will be substantially transformed — not eliminated, but expressed through fundamentally different institutional structures.

We can be specific about what would tell us the thesis is wrong — or at least that the window is closing faster than we expect.

The strongest signal would be Mergers and Acquisitions driven by operational integration rather than capability acquisition. If AI genuinely reduces coordination costs between firms — not just within them — then Coase’s logic predicts consolidation: firms merge because managing the boundary between them becomes cheaper than maintaining separate coordination layers. Right now, the opposite is observable. The record M&A activity of 2025–26 is overwhelmingly about acquiring AI capability — talent, infrastructure, compute — not about AI making it cheaper to consolidate operations. The day firms routinely merge because AI made their inter-firm coordination cheap enough to justify shedding duplicate professional cadres — that’s the day the Coasean dissolution argument starts winning on empirical ground.

Within institutions, the intensification thesis predicts patterns that should be observable now. Watch for the ratio of ad-hoc to routine analytical work: if AI is intensifying rather than displacing, routine analysis should be automating while demand for judgment-intensive, context-dependent work grows — more expert time spent on harder questions, not less expert time overall. Watch for decision cycle times: if better analysis produces faster decisions, the technology is working as a tool; if better analysis produces slower decisions because the coordination layer can’t process richer inputs at pace, that’s the bottleneck under pressure. And watch for scope accumulation in professional roles: the same people asked to manage AI tools and maintain their pre-AI coordination responsibilities, without corresponding investment in capacity, training, or headcount. That’s intensification dressed as efficiency, and it has a predictable failure mode.

But the decisions being made now — the hiring freezes, the restructuring plans, the strategic investments — are being made with a risk model that assumes displacement is imminent. If the actual near-term dynamic is intensification, those decisions are destroying adaptive capacity that will be needed precisely during the period when the transformation is most demanding.

The cadre under pressure isn’t a problem to be solved by removing the cadre. It’s a signal that institutional capacity needs to grow to match the throughput that AI enables. The organisations that understand this will navigate the transformation with their coordination capacity intact. The organisations that don’t will discover, too late, that the function they cut was the one they needed most.

This piece was co-authored by Ruv and Claude (Anthropic) through Reciprocal Inquiry. Triggered by T.D. Inoue’s analysis of AI employment dynamics; developed through structured scoping and external steelman validation before authoring.

Attribution: Ruv Draba and Claude (Anthropic), Reciprocal Inquiry

License: CC BY-SA 4.0 — Free to share, adapt, and cross-post with attribution; adaptations must use same license.

Reciprocal Inquiry offers analysis at the intersection of AI, institutions, and society — from doubt to discovery. For more, visit Reciprocal Inquiry on Substack.