WORKING PAPER · VERSION 1.0 · JULY 2026 · A live inquiry, released in versions as evidence accrues. Comments and case evidence welcome: info@impactthinking.co.uk
Senior judgment was never taught. It was caught — absorbed through two channels no organisation ever managed because no organisation ever had to: proximity to experienced practitioners, and the long apprenticeship of foundational work done by one’s own hand. Within five years, both channels have been cut. Hybrid work removed the proximity; generative AI is removing the reps. Each change is individually defensible and jointly unexamined — and because tacit capability registers on no dashboard, every current indicator will read healthy while the pipeline that produces the next generation of senior judgment quietly stops producing. This working paper sets out the mechanism, the early signals we are tracking across client organisations, and the open questions on which we are gathering evidence.
Ask any senior leader where their judgment came from and the answers are strikingly consistent — and strikingly un-curricular. They sat near someone. They watched how a particular partner handled a particular client in a particular room. They did the grunt work — the first drafts, the models built from scratch, the analysis that a senior then pulled apart in front of them — a thousand times, until something that could never be written down had moved from the senior’s hands into theirs.
The theoretical account of why it works this way is old and solid. Polanyi named the substrate: we know more than we can tell — the most valuable professional knowledge is tacit, incapable of full articulation, and therefore incapable of transmission by articulation.1 Dreyfus and Dreyfus mapped the ladder it climbs: the passage from competence to expertise is precisely the passage beyond rules, into an embodied, situational grasp that is built only through volumes of lived cases with real stakes and real feedback.2 Lave and Wenger supplied the social mechanism: professionals are formed by legitimate peripheral participation — doing real, initially small work at the edge of a community of practice, absorbing its judgment by participation in it.3 Three literatures, one conclusion: expertise transmits through proximity and reps, or it does not transmit.
Neither channel was ever designed, budgeted, or measured. Both are now being unbundled — by separate decisions, made on separate grounds, whose interaction nobody owns.
The proximity cut. Distributed work has many virtues; ambient apprenticeship is not among them. The overheard negotiation, the watched recovery from a bad meeting, the two-minute unscheduled “why did you do it that way?” — the entire incidental curriculum ran on co-presence, and scheduled video contact does not carry it: what juniors absorbed was rarely the content of meetings but the conduct around them. The structural evidence points the same way — Microsoft’s study of over 60,000 employees moving to firm-wide remote work found collaboration networks becoming more static and siloed, with fewer of the bridging ties through which knowledge, especially new knowledge, travels.4 The tie type that thinned is exactly the tie type apprenticeship runs on.
The reps cut. The second channel is being removed by a force with better press: generative AI is absorbing precisely the work that built junior judgment — the first draft, the base analysis, the initial model, the document review. The efficiency case is real and is being banked. What is not being priced is that this work was never only output; it was training load. Brynjolfsson and colleagues’ study of AI assistance in customer support found the largest performance gains accruing to the least experienced workers5 — celebrated as democratisation, and fairly so; but performance support and capability formation are different quantities, and a novice performing at competence with assistance is not thereby becoming competent. Dell’Acqua and colleagues’ field experiment with consultants adds the warning label: substantial gains inside the AI’s frontier of competence, degraded performance beyond it — skilled professionals, in the authors’ phrase, falling asleep at the wheel.6 Judgment is, by definition, what operates beyond the frontier.
The defining property of this problem is its invisibility to current measurement, for three compounding reasons. Tacit capability has no metric — what Polanyi showed cannot be articulated cannot be KPI’d. The lead time exceeds every planning horizon — today’s juniors are the senior bench of the late 2030s, and no quarterly cycle prices a 2037 capability shortfall. And the early symptoms are indistinguishable from good news: junior output is up, quality-with-assistance is up, cost per deliverable is down. The first undeniable signal — a cohort arriving at senior roles able to operate brilliantly inside the frontier and unable to stand anywhere beyond it — arrives roughly a decade after the cause, unattributable and unfixable on any useful timescale.
Talent directors, in our conversations, already sense the leading edge and lack the language for it: “technically excellent, but green in a way I can’t train.” The mechanism above is, we believe, the name for that greenness.
If the incidental apprenticeship is gone, the replacement must be deliberate — the conditions rebuilt on purpose. Early practice across our client base suggests four design elements. Protected reps: designating a fraction of foundational work as development work, done by hand not because AI cannot, but because the human must — the way flight training preserves manual hours inside automated cockpits. Engineered proximity: apprenticeship-dense time (deal rooms, crisis reviews, live client work) explicitly scheduled as transmission, with juniors present for the conduct, not the minutes. Critique restored: the senior tear-down of junior work — the highest-bandwidth transmission event the professions ever had — reinstated on AI-assisted output, interrogating the judgment (“why this structure, why this number, what would make this wrong?”) rather than the polish. Judgment reps beyond the frontier: deliberately assigning work where the tools are weak — ambiguous, precedent-poor problems — early and often, because that is the terrain the next senior generation must be formed on, and it is exactly the terrain assistance now lets them avoid.
Three orders of magnitude frame the stakes. The first is the pre-existing scarcity: Korn Ferry’s global modelling projected a talent shortfall of more than 85 million skilled workers by 2030, with unrealised annual revenues in the trillions of dollars8 — a forecast made before either transmission channel was cut, and premised on pipelines continuing to convert juniors into seniors at historical rates. The second is the breadth of the proximity cut: in the post-pandemic settlement, roughly half of remote-capable employees work hybrid and a further quarter fully remote (Gallup’s tracking has been stable on this for several years)9 — meaning the ambient apprenticeship has thinned not for an exposed minority but for the majority of the professional workforce. The third is the speed of the reps cut: in Brynjolfsson and colleagues’ field data, AI assistance raised support-agent productivity ~14% on average and ~34% for the least experienced5 — adoption economics so favourable that the junior-work transfer is happening faster than any prior automation wave, and with it the transfer of the training load.
Set those three against each other and the exposure statement writes itself: institutions are entering a decade of projected senior scarcity while simultaneously — and invisibly — reducing the conversion rate of the pipeline that was supposed to relieve it.
For talent directors and CHROs, the near-term agenda has five moves. One: baseline the two channels now — measure by-hand foundational hours and junior–senior contact minutes against 2019, because in three years the baseline will be unrecoverable. Two: designate protected reps formally, in workforce plans and utilisation targets, so they survive the first cost review. Three: reinstate structured senior critique as a counted activity — what is not counted will be optimised away (see Working Paper 03). Four: redesign early-career paths around frontier work — ambiguous, precedent-poor assignments — earlier than historical sequencing would dare. Five: add a tacit-capability line to board people-reporting, explicitly flagged as judgment-assessed, so the invisibility is at least official.
For professional-services leadership, the exposure is existential rather than adjacent: the pyramid’s economics (juniors doing foundational work profitably while being formed by it) and its succession logic are the same mechanism, and AI unbundles both at once. Firms that treat the junior layer purely as a cost line to automate are liquidating their partner bench of 2035 at a price no current metric records.
For government and civil-service capability units, two specifics sharpen the general case: policy judgment has always depended even more heavily than commerce on apprenticeship (the private office, the bill team, the negotiation watched from the second row), and public institutions cannot buy senior judgment in from a market that is depleting on the same schedule. Capability reviews that inventory skills but not transmission conditions are auditing the stock while the flow fails.
As a live inquiry, this paper’s claims are staked on observables. Across participating organisations we are tracking: the share of junior-role hours spent on by-hand foundational work (2019 baseline vs current); the frequency of structured senior critique events per junior per quarter; unscheduled junior–senior contact minutes under different working patterns; and — the slow variable — cohort performance on precedent-poor assessment tasks, inside versus outside AI assistance. We expect the first usable cohort comparisons in versions 2 and 3 of this paper.
Honestly held, and open: Whether AI-assisted reps can be made developmental — whether critique-rich use of the tools builds judgment rather than merely borrowing it (the evidence could land either way, and Brynjolfsson’s novice gains hint at a trainable configuration). Whether the proximity channel can be genuinely reproduced at distance, or only approximated. What the minimum effective dose of protected reps is, given that no organisation will pay for the full historical apprenticeship. And whether the professions will notice in time — or whether this joins the class of slow institutional failures that are obvious only in the inquiry report.
If your organisation has evidence on any of these — cohort data, natural experiments, counter-examples — we want it. This paper will be revised against it.
IMPACT THINKING RESEARCH · BY BEN BOTES · WORKING PAPER 01 · v1.0 · JULY 2026
If your organisation is seeing the early signals — or has evidence that cuts either way — we want to compare notes.
Start a conversation → Back to the research desk