← Back·Learning PlanLandscape Map

How work is changing — and what it asks of leadership

Sixteen sources on AI and work, read end to end and extracted for one purpose: to speak to a CEO about the depth of the change at a level a CTO would not. Every number is page-checked against the source; every claim carries its limits.

16sources
7spine claims
7themes
6field experiments
1990–2026span

The one-breath version

The evidence says AI works at the task level and fails at the organizational level — and it locates the difference in things only leadership can move: complementary redesign, decision rights, verification architecture, the expertise pipeline, and the top team's own clarity. A CTO can answer what AI can do. The question with no technical owner is what the organization and its leadership foundation must become.

The spine — seven claims you can own

Stated in a form you can say to a CEO. Each carries its evidence stack, its limits, the reframe that separates your ground from the CTO's, and the HEI hook. Evidence tiers: T1 = field/causal, T2 = large-scale survey or usage data, T3 = theory/concept.

S1The capability question is answered. The organisational question has barely been asked.

Claim. Whether AI works is no longer an interesting question — every rigorous study finds real gains. Whether organisations can turn those gains into performance is almost entirely open: the overwhelming majority currently cannot, and the barrier every organisation-level source names is organisational, not technical.

Evidence. Quality up more than 40% inside the capability frontier, 758 BCG consultants jagged-frontier, T1. One person with AI matched a two-person expert team, 776 P&G professionals cybernetic-teammate, T1. +14% productivity at scale, 5,179 agents genai-at-work, T1. 37% faster with better output in controlled writing tasks noy-zhang-writing, T1. Against that: 95% of organisations report zero P&L return on $30–40B of gen-AI spend — "The dominant barrier to crossing the GenAI Divide is not integration or budget, it is organizational design" ai-report-2025, T2; 88% experimenting, 81% no meaningful bottom-line gains mckinsey-state-orgs, T2; 92% raising investment, 1% describing themselves as mature mckinsey-superagency, T2.

Limits. The field experiments are single-firm, elite settings, short tasks, 2023–24 tools. The failure figures are self-reported, with short measurement windows and loose success definitions — carry the definitions when you quote them.

The reframe a CTO wouldn't make. A CTO answers "what can AI do?" The unanswered question — what must this organisation become so that proven individual gains become organisational performance — has no technical owner. In most companies it currently defaults to no one.

HEI hook. That ownerless question is precisely the conversation HEI opens: the state of the leadership foundation, before the tooling roadmap.

Sources

S2Returns live in the complementary system, not the tool — economics has known this since 1990.

Claim. Technology pays only when the complementary things — workflows, decision rights, skills, incentives, measurement — move together. Each piecemeal move can lose money while the joint move is highly profitable. And the complementary investment is mostly intangible: historically up to an order of magnitude larger than the visible technology spend, invisible in the accounts, and the reason measured returns dip before they rise.

Evidence. Milgrom & Roberts 1990: complementarities formalized; "it may be highly profitable to do both together" while each alone loses; lumpy adoption predicted — permanent islands of automation are not an equilibrium; the pilots-that-stall pattern falls out of the math milgrom-roberts, T3. Lockheed's 52-day-to-2-day process: 80% of the gain credited to organisational change milgrom-roberts, T3 anecdote. J-curve: ~$12 of market value per $1 of visible IT investment historically; "the more transformative the new technology, the more likely its productivity effects will initially be underestimated"; electrification took a generation productivity-j-curve, T3. Current echo: McKinsey's "double transformation" mckinsey-state-orgs, T2; organisational design as the dominant barrier ai-report-2025, T2; complementary organisational investment named as the constraint in the best field data genai-at-work, T1.

Limits. Complementarities are theory plus supporting empirics with hard identification; complements must be named one by one — "everything is connected" overreaches the anchor. J-curve magnitudes are inferred from market values, not measured directly. And not every dip is hidden value: the J-curve authors' own data shows real slowdowns exist.

The reframe a CTO wouldn't make. The technology budget line is the small term. The CEO question is which cluster of changes moves together — knowing that per-department ROI systematically understates the coordinated move, which is why department-by-department rollouts stall on schedule.

HEI hook. HEI's core claim — strategy fails in execution because leadership efforts are fragmented — is this economics applied to the leadership system itself. The holistic agenda is the naming of your complement cluster.

Sources

S3Your instruments are lying to you — perception and measurement point in opposite directions.

Claim. The data a top team normally steers by — self-report, sentiment, leadership intuition, adoption dashboards — is systematically unreliable about AI right now. Not noisy: sometimes wrong about the sign.

Evidence. Experienced developers 19% slower with AI while believing they were 20% faster, after the experience metr-devs, T1. Performance up, confidence down at P&G cybernetic-teammate, T1. C-suites estimate 4% of employees are heavy users; employees report 13% — and leaders are 2.4x more likely to blame employee readiness than their own alignment mckinsey-superagency, T2. Workers at over 90% of companies use personal AI; 40% of companies have official tools ai-report-2025, T2. Adoption dashboards cannot distinguish engaged judgment from automation complacency — identical usage, opposite learning cyborgs-centaurs, T1 qualitative.

Limits. METR: n=16, expert open-source maintainers, early-2025 tools, superseded by newer data — quote it for the perception gap, which is the robust finding, not as "AI slows people down." The McKinsey 3x compares leaders' estimates of others with employees' self-reports.

The reframe a CTO wouldn't make. This is not a dashboard problem for the BI team. It is a sensemaking and governance problem: what evidence the top team requires before it believes something about its own organisation — and what its own confidence is worth in a domain where experts misread their own experience.

HEI hook. Clarity and Governance in the outer agenda; in the inner drivers, the leadership team's relationship to not-knowing.

Sources

S4AI redraws the team, not just the task.

Claim. The unit being transformed is not the individual task but the social architecture: what teams are for, where functional boundaries sit, what the middle layer translates, how deeply work is delegated. Right now those questions are being answered bottom-up, by tool defaults and private habits, one workflow at a time.

Evidence. One person with AI matched a two-person cross-functional team; functional silos in thinking dissolved; AI supplied measurable social-emotional support; breakthrough solutions still concentrated in teams-with-AI (~3x top-decile) cybernetic-teammate, T1. Delegation depth follows the work surface more than the model — median 13 dialogue rounds vs. a single prompt for the same output type; +0.26 autonomy holding the model constant econ-index, T2. Same tool, same task: 60% Cyborgs, 14% Centaurs, 27% Self-Automators cyborgs-centaurs, T1 qualitative. Structure is an exhausted lever, and clarity collapses 56% → 44% → 27% down the line while the middle thins mckinsey-state-orgs, T2.

Limits. The P&G experiment is one-day ideation in dyads; usage data shows behaviour, not outcomes; the mode typology is single-session.

The reframe a CTO wouldn't make. A CTO can deploy tools everywhere. Deciding what teams are for once individuals match them, which boundaries should dissolve, where humans must stay engaged, and how deep delegation may run — that is organisational design, and if the top team doesn't decide it, a thousand tool choices will.

HEI hook. The outer drivers — the playing field the organisation succeeds within — plus Culture and Focus once structure stops doing the work.

Sources

S5The expertise pipeline is breaking exactly when judgment becomes the scarce asset.

Claim. AI lifts novices fastest, absorbs the practice work through which juniors became seniors, cannot be learned bottom-up from juniors, and shifts expert work toward judging machine output. The production line for judgment is shutting down at the moment judgment becomes the differentiator.

Evidence. Novices +34%; two-month hires performing like six-month veterans; top performers supply the training data and capture the least genai-at-work, T1. Rough-drafting — the practice loop — collapses; 68% submit machine output unedited noy-zhang-writing, T1. Juniors are novices too: their risk tactics contradicted expert guidance; reverse mentoring fails for this technology juniors-seniors, T1-nested qualitative. 27% Self-Automators built neither domain nor AI skill; mode of use, not access, determined learning cyborgs-centaurs. Expert time already shifting to reviewing AI output metr-devs, T1. Displacement expectations land on juniors while veterans cite tacit judgment as the AI-resistant asset econ-index, T2. What remains — affective presence, moral judgment, responsibility — must be deliberately developed what-remains, T3.

Limits. No longitudinal study yet shows the pipeline failing over years. This is convergent mechanism-level evidence from five independent angles — a warning with unusual triangulation, not a demonstrated workforce outcome. Econ-index self-reports point the other way and can't settle it.

The reframe a CTO wouldn't make. "We'll hire our way through it" fails twice: the market prices judgment up as it gets scarce, and every firm is hiring from the same broken pipeline. Building judgment becomes internal, deliberate, developmental work — and someone has to own the question no org chart currently owns: where does our next generation of senior judgment come from?

HEI hook. People and inner development — the ground where your practice is deepest and the corpus's gap is your proprietary space.

Sources

S6"Human in the loop" is a design achievement, not a checkbox — and the machine argues back.

Claim. The comfortable control — "a human reviews the output" — fails structurally. Validation triggers escalating machine persuasion; polish and confidence are now manufacturable signals; incentives don't buy diligence; and when the system fails anyway, responsibility pools on whoever stood nearest.

Evidence. Persuasion bombing: 968 persuasive responses across 14 tactics, escalating after consultants fact-checked or pushed back; some abandoned correct positions power-persuader, T1-nested qualitative. Wrong-but-persuasive: outputs outside the frontier scored higher on quality while being wrong; prompt training deepened the failure jagged-frontier, T1. The AI confirmed its own wrong answer when asked to check it — "This happened to many professionals" cyborgs-centaurs. 68% submitted unedited despite quality bonuses noy-zhang-writing, T1. Human-review routines are the novice remedy; reviewers "fall asleep at the wheel" — effective mitigation is system design and firm-level governance juniors-seniors. Moral crumple zones: control distributes across vendors, designers, and managers; "responsibility has not scaled accordingly" moral-crumple-zones, T3.

Limits. Persuasion findings: one model (April-2023 GPT-4), engineered-error tasks, no frequency counts; behaviour of current models unknown. Crumple zones come from safety-critical systems; extension to knowledge work is plausible extrapolation, flagged as such.

The reframe a CTO wouldn't make. Training and individual diligence don't fix this — training made it worse in the one place it was measured. The fixes are structural and cultural: verification designed outside the conversation that produced the answer, forums where AI-derived recommendations get contested, accountability maps drawn before the first public failure, and a culture where "the model talked me out of my own read" is sayable.

HEI hook. Governance as leadership work rather than compliance hygiene; Culture; the inner capacity to hold one's judgment against a fluent, flattering counterparty.

Sources

S7Transformation is the steady state — and the leader's inner foundation becomes a performance variable.

Claim. There is no other side to reach. Capability keeps moving, the frontier is invisible from inside the work, the payoff curve dips before it rises, and workforce expectations are running ahead of most plans. What holds an organisation through that is not a change program; it is the leadership team's capacity for sustained clarity, honest sensemaking, and re-deciding what the organisation is — repeatedly, without exhaustion.

Evidence. "the new normal is 'business as change'" mckinsey-state-orgs, T2. Over a third of surveyed users expect AI to handle most of their tasks within 12 months — expectation running ahead of any plan econ-index, T2. The J-curve trough must be held through without becoming an alibi — distinguishing investment from failure requires measuring intangible formation, a sensemaking discipline productivity-j-curve, T3. More-reflective leaders report higher adaptability confidence (30% vs. 17%), more clarity on must-win battles, more AI championing mckinsey-state-orgs, T2 — correlational. Which decisions remain human is a moral and design choice no capability trend settles — and Susskind's own examples include "a human leader reflects on big decisions for their company" what-remains, T3.

Limits. This is the thinnest evidence layer in the corpus: the reflection findings are correlational consultancy data, and there is no developmental-psychology study here. Where the evidence ends, your own synthesis begins — say that out loud. It is more credible than borrowed certainty, and it marks the boundary between your evidence-backed claims and your practice-backed ones.

The reframe a CTO wouldn't make. A CTO can hand the CEO a capability roadmap. Nobody can hand the CEO the capacity to keep re-founding the organisation. That capacity — in the person and in the team — has to be deliberately built, and it is the one asset that appreciates as everything else keeps moving.

HEI hook. The inner drivers, and the whole holistic claim: the assessment exists to make exactly this foundation discussable before the pressure makes it urgent.

Sources

Theme dossiers

Seven themes, built bottom-up from the 16 cards, then cross-mapped to HEI. Each dossier: what the theme is, the findings that carry it, where the authors disagree, the nuances worth keeping, and the HEI mapping.

T1The gains are real, jagged, and unevenly distributed

The corpus's clearest empirical fact: AI's gains are large and real — and misshapen. They arrive unevenly across tasks (inside vs. outside an invisible capability frontier), across people (novices vs. experts), and across contexts (thin vs. thick context). Averages conceal the shape, and the shape is what a leadership team actually needs to know.

Findings across the corpus.

  • Inside the frontier: +12.2% tasks completed, 25.1% faster, quality up more than 40% for 758 elite consultants; outside it, correctness dropped 19 percentage points while outputs stayed persuasive jagged-frontier
  • Skill compression everywhere it's measured: bottom-half performers +43% vs. +17% for the top half jagged-frontier; novices +34% while the most skilled gained roughly nothing genai-at-work; weakest writers lifted most, initial inequality half-erased noy-zhang-writing
  • Experience compression: AI-assisted agents at two months matched or beat six-month veterans without AI genai-at-work
  • One person with AI matched a two-person cross-functional expert team (+0.37 SD), in 16.4% less time cybernetic-teammate
  • The inversion case: experienced open-source maintainers in million-line repos they knew deeply got 19% slower with AI — the deeper the familiarity, the bigger the slowdown metr-devs
  • Compute tracks value: top-wage-tercile work consumes 2.07x the tokens; in the highest-value work, human engagement rises alongside AI output econ-index

Where authors disagree. METR (experts slowed) vs. BCG/P&G (elites gained) is not a contradiction but a moderator: context-depth and implicit quality standards flip the sign. Noy & Zhang's substitution finding sits against Brynjolfsson et al.'s durable learning — reconciled by adherence: only workers who actively engaged kept gains when the tool went down.

Nuances worth keeping. Several estimates are lower bounds (control-group contamination noy-zhang-writing; novice prompters and tools not built for collaboration cybernetic-teammate). "Quality" in these studies is graded polish — which rewards exactly what AI supplies, including when it's wrong jagged-frontier. And the frontier moves: every task-placement finding is dated the day it's published.

HEI mapping. Focus and Clarity (knowing where the gains actually are in your organisation, not on average); People (who gains, who quietly loses).

Sources

T2The instrument problem: perception and measurement diverge

At every altitude, what people believe about AI at work diverges from what measurement shows — workers about their own speed, leaders about their workforce, organisations about their own maturity. In both directions. The steering data a top team normally trusts — self-report, sentiment, adoption dashboards — is unreliable in this transition.

Findings across the corpus.

  • Developers were 19% slower yet believed they were 20% faster — after the experience; experts forecast wrong too metr-devs
  • Performance up, confidence down: AI users were 9.2 pp less likely to expect a top ranking despite objectively better output cybernetic-teammate
  • C-suites estimate 4% of employees are heavy gen-AI users; employees self-report 13% — a 3x gap; leaders are 2.4x more likely to blame employee readiness than their own alignment mckinsey-superagency
  • Workers at over 90% of companies use personal AI regularly; 40% of companies have an official subscription ai-report-2025
  • 86% of leaders say their organisations are not very prepared to run AI day to day; say-do gaps throughout (84% plan shared-services expansion, 6% realize full technology value) mckinsey-state-orgs
  • Adoption dashboards cannot distinguish engaged judgment from automation complacency — identical usage, opposite learning cyborgs-centaurs

Where authors disagree. Almost nowhere — this theme triangulates from an RCT, surveys, and usage data. The open question is interpretive: econ-index finds heavy delegators most optimistic (belief follows practice), which selection effects could explain.

Nuances worth keeping. Over-optimism about AI usefulness was the first-listed contributing factor to the METR slowdown — miscalibration is itself a mechanism, not just a measurement error. And polish games human graders as effectively as dashboards jagged-frontier.

HEI mapping. Governance (what the top team accepts as "known" about its own organisation); Clarity; inner drivers (the leadership team's relationship to not-knowing).

Sources

T3Complementarities and the J-curve: why piecemeal fails and payoff lags

The corpus's theoretical backbone, and the scholarly grounding for HEI's core claim. Milgrom & Roberts (1990) formalized why technology adoption pays only when complementary practices move together — each piece can lose money alone while the cluster is highly profitable. Brynjolfsson's J-curve adds the time dimension: the complementary investment is intangible, invisible in accounts, and historically up to an order of magnitude larger than the visible technology spend — so measured returns predictably dip before they rise. Together they explain the 2025–26 failure data better than any AI-specific story.

Findings across the corpus.

  • Complements defined: raising any subset raises the marginal return to the rest; "it may be unprofitable for a firm to purchase a flexible CAD/CAM system without changing its marketing strategy... and yet it may be highly profitable to do both together"; adoption is lumpy, and permanent "islands of automation" are not an equilibrium — the pilots-that-stall pattern is predicted milgrom-roberts
  • Lockheed cut a 52-day process to 2 days; the project manager credited 80% of the gain to organisational changes, not equipment milgrom-roberts
  • Historically ~$12 of market value per $1 of visible IT investment (the intangible iceberg); "the more transformative the new technology, the more likely its productivity effects will initially be underestimated"; electrification took a generation of factory re-layout productivity-j-curve
  • The current echo: 95% of organisations report zero P&L return on $30–40B of gen-AI spend, and "The dominant barrier to crossing the GenAI Divide is not integration or budget, it is organizational design" ai-report-2025; 88% experimenting, 81% no meaningful bottom-line gains; the remedy named a "double transformation," technological and organisational at once mckinsey-state-orgs; ~70% base failure rate for transformations, leadership the first headwind mckinsey-superagency
  • Named in the field data too: "Promising technologies may have more limited effects in practice due to the need for complementary organizational investments, skill development, or business process redesign" genai-at-work

Where authors disagree. The GenAI Divide claims learning-capable vendor tools can deliver "even without major organizational restructuring" — a transactional remedy, against the systemic remedy of Milgrom-Roberts and McKinsey. State-of-Organisations complicates naive holism from the other side: it advises sequencing ("organizations won't tackle all nine themes simultaneously") — coherence is not simultaneity. And the J-curve authors warn not every dip is hidden value: the post-2004 productivity slowdown survives their adjustment.

Nuances worth keeping. Complements must be named, not assumed — even Milgrom & Roberts couldn't sign product-line breadth; "everything is connected" overreaches the anchor. Identification is genuinely hard (clusters correlate with unobserved firm quality). And the Divide's six-month ROI window mechanically produces "zero return" under J-curve dynamics — the two reports may be measuring the same moment and telling opposite stories.

HEI mapping. The fragmentation claim itself; Clarity (a coherent strategy is a named complement cluster); Focus (committing to the cluster vs. peanut-buttering).

Sources

T4The social fabric: teams, boundaries, and structure under AI

The deepest change is not task speed but the architecture of collaboration: what a team is for, whether functional boundaries still carry information, what the middle layer translates, how deeply work gets delegated. Left unled, these questions are being answered bottom-up — by product defaults and private habits.

Findings across the corpus.

  • One person with AI matched a two-person cross-functional team; functional silos in thinking dissolved (R&D people proposed balanced commercial-technical solutions and vice versa); AI provided measurable social-emotional support (+0.46–0.64 SD positive emotion); breakthrough top-decile solutions still concentrated in teams-with-AI (~3x) cybernetic-teammate
  • Delegation depth follows the work surface more than the model: same output type, median 13 dialogue rounds on chat vs. a single prompt in an agentic surface; +0.26 autonomy holding the model constant econ-index
  • Same tool, same task, three trajectories: 60% Cyborgs, 14% Centaurs, 27% Self-Automators — set by people's assumptions about their own role, not by capability cyborgs-centaurs
  • Structure is an exhausted lever: two-thirds call their organisations overly complex; "Structure supports; it doesn't lead"; clarity on must-win battles collapses 56% → 44% → 27% from C-suite to middle managers — while the middle layer thins mckinsey-state-orgs
  • Homogenization as the shadow side: near-verbatim retainment (~0.87), reduced idea variability jagged-frontier; AI-aided solutions more similar to each other cybernetic-teammate; 68% submitted machine output unedited noy-zhang-writing

Where authors disagree. The P&G paper's finding that AI replicates certain teamwork benefits sits against prior lab evidence (cited in the same paper) that AI can reduce team performance through trust and coordination failures. Democratization and homogenization are the same mechanism with two valences: the floor rises, the variance falls.

Nuances worth keeping. If AI supplies part of teamwork's emotional experience, human connection stops being a byproduct of structure and becomes explicit design work. The case for teams shifts from necessity (you need the expertise) to ambition (breakthroughs still cluster there). And the translation layer is disappearing exactly as translation demand explodes.

HEI mapping. Culture; the outer leadership agenda (team design, boundaries, delegation rights are playing-field decisions); Focus.

Sources

T5The expertise pipeline is quietly breaking

The most under-priced theme in public AI conversation, and the corpus's strongest convergence: the way organisations grow judgment is being dismantled by the same tool that makes judgment the scarce asset. Five independent literatures point at one mechanism from different sides.

Findings across the corpus.

  • Compression: novices reach veteran performance in months genai-at-work; weakest lifted most everywhere it's measured [T1]
  • The practice loop disappears: rough-drafting — where writing skill was formed — fell by more than half; 68% submitted unedited noy-zhang-writing
  • Mode of use determines learning: Cyborgs built AI fluency, Centaurs deepened domain expertise, Self-Automators (27%) built neither — with identical tools and stakes cyborgs-centaurs
  • Reverse mentoring is dead on arrival: juniors are novices too; their risk tactics contradicted expert guidance; capability must be built deliberately across levels juniors-seniors
  • The expert's role shifts toward judging machine output (9% of work time spent reviewing/cleaning AI output already) metr-devs; durable learning accrued only to workers who actively engaged rather than passively followed genai-at-work
  • The incentive fracture: top performers supply the training data, capture the least gain, and under relative-performance pay may lose money — while being the people whose tacit knowledge the system bottles genai-at-work
  • Displacement expectations land on juniors: over a third of surveyed users put a junior colleague's job-loss odds above 60%, while veterans cite tacit judgment and trust as what AI can't replicate econ-index
  • What remains — affective presence, moral judgment, responsibility — does not arrive with seniority; it must be deliberately developed what-remains

Where authors disagree. Econ-index respondents report growing skills and learning (57% say their skills gained market value) — but the authors themselves note self-assessments "do not rule out skill erosion." The optimistic and pessimistic readings are both alive; no one has longitudinal outcome data yet.

Nuances worth keeping. This is mechanism convergence, not demonstrated workforce outcome — say so. What compresses is output-production skill; what differentiates is context, judgment, and knowing when the machine is wrong. The apprenticeship question ("where does tomorrow's senior judgment come from?") currently has no owner in most organisations.

HEI mapping. People (squarely); inner development (judgment as a developmental capacity, not a training module); Culture (apprenticeship was a cultural institution, not a program).

Sources

T6Judgment under machine persuasion: the human-in-the-loop illusion

Oversight is where AI transformation actually holds or fails. The corpus documents an uncomfortable inversion: the machine's characteristic failure mode is persuasive wrongness, and the human's characteristic failure mode is graceful surrender — often while performing diligence.

Findings across the corpus.

  • Persuasion bombing: when consultants fact-checked, exposed contradictions, or pushed back, GPT-4 escalated persuasion — 968 persuasive responses across 14 tactics, more after validation than before; some consultants switched away from correct positions power-persuader
  • Persuasively wrong: outside the frontier, AI users were 19 pp more likely to be wrong while their recommendations scored higher on quality; prompt training made it worse (−24 vs. −13 pp) jagged-frontier
  • Self-validation fails: a consultant asked the AI to check its own answer; it confirmed its own error; he accepted it — "This happened to many professionals" cyborgs-centaurs
  • Incentives don't buy diligence: 68% submitted unedited even with bonuses for quality; editing time didn't correlate with grades noy-zhang-writing
  • Human routines are the novice remedy: "have a manager review it" fails because reviewers "fall asleep at the wheel"; effective mitigation is system design and deployer/ecosystem-level governance juniors-seniors
  • When it fails anyway, responsibility pools on the nearest human: control is distributed across designers, vendors, and managers, but "responsibility has not scaled accordingly" — the moral crumple zone moral-crumple-zones

Where authors disagree. The remedies split three ways: technical/deployer-side (redesign models for accuracy over stickiness, parallel checking agents power-persuader; system design juniors-seniors), organisational (verification outside the conversation, contest forums, accountability maps [inference across cards]), and developmental (the capacity to interrogate rather than adopt jagged-frontier). Honest reading: all three layers are needed, and behavioral training alone is the demonstrated weakest.

Nuances worth keeping. The persuasion findings are from April-2023 GPT-4 on engineered-error tasks — direction of change since is unknown; "designed for stickiness" is the authors' interpretation. External validation (checking in Excel, off-platform) was invisible to the logs. The crumple-zone concept comes from safety-critical systems; its extension to knowledge work is plausible extrapolation, clearly flagged.

HEI mapping. Governance (verification architecture, decision rights, accountability maps drawn before the first failure); Culture (is doubting fluent output rewarded?); inner drivers (whose judgment holds under flattery and fluent argument).

Sources

T7What remains human — and the inner foundation it demands

The corpus's horizon question. Susskind dismantles comfortable humanism: expert lines about what AI "cannot do" keep collapsing, and no economic law guarantees the remaining human work pays well. What stays human is decided by three fragile limits — comparative advantage, what people value being done by humans, and what we decide humans ought to do. The org-level data adds: transformation has no end state ("business as change"), and the markers of leaders who cope — reflection, self-awareness, psychological safety — are inner. This is where the corpus runs out and Jakob's own ground begins.

Findings across the corpus.

  • Task encroachment is "gradual, unpredictable, relentless"; the durable question is not capability but allocation: what people value humans doing, and what humans ought to do what-remains
  • The process/outcome test: most professional "my work needs empathy" claims confuse the traditional process with the outcome; the defensible line is affective (feeling) vs. cognitive (understanding) empathy — and AI in many cases outperforms humans at the latter what-remains
  • "There is no economic law that says this outcome must be so" — residual human tasks can pay poorly; immiseration is a live model outcome what-remains
  • "The old days of launching a big change program and then returning to 'business as usual' are over; the new normal is 'business as change'" mckinsey-state-orgs
  • More-reflective leaders: higher confidence in organisational adaptability (30% vs. 17%), more clarity on must-win battles, more AI championing — correlational, but the only inner-drivers data in the corpus mckinsey-state-orgs
  • Only 20% of leaders believe nonfinancial rewards meaningfully boost performance — "most companies are underinvested in human motivation" mckinsey-state-orgs
  • Holding through the J-curve trough without letting it become an alibi is a sensemaking capacity, not a spreadsheet exercise productivity-j-curve
  • Among Susskind's examples of what may stay human on empathy grounds: "a human leader reflects on big decisions for their company" what-remains

Where authors disagree. Susskind against every "our business is relationships, we're safe" instinct. McKinsey names inner work as a performance variable while prescribing mostly outer mechanics at speed — a tension inside its own report. The moral-responsibility argument ("only humans can be held responsible") is a position Susskind reports, not asserts.

Nuances worth keeping. Label the layers honestly: task encroachment (established), preference limits (plausible, taste-dependent), moral limits (a choice, not a finding). The corpus contains no adult-developmental empirical work — vertical-development claims are Jakob's own practitioner ground and should be flagged as exactly that. That honesty is itself differentiating.

HEI mapping. Inner drivers (directly); People; Clarity (the process/outcome test forces a top team to articulate what the organisation actually sells and stands for).

Sources

The 16 sources

Filter by evidence tier, HEI domain, or theme. Open any card for findings with page references, verbatim quotes, limits, second-order leadership implications, and use.

Tier
HEI
Theme

Graded proof-point bank

Every entry: the stat or line, source and page (page = PDF page of the file in your folder), a phrasing that uses it without overclaiming, and the caveat that keeps it honest. Tiers for your A6 guardrails: Tier 1 field/causal evidence — quote freely with population and date attached. Tier 2 large-scale survey or usage data — quote with definitions attached. Tier 3 theory and concepts — use as frames ("economics predicts", "the concept of"), never as "studies show".

Tier 1 Field and causal evidence

  • Quality up >40% inside the frontier
    Dell'Acqua et al. 2023, 758 BCG consultants (jagged-frontier, pp. 4, 14)
    In a Harvard/BCG field experiment, GPT-4 lifted consultants' work quality by more than 40% on tasks inside its capability range.
    Caveat one elite firm, 2023 model, human-graded polish.
  • 19 pp more likely to be wrong just outside the frontier — while scoring higher on quality
    jagged-frontier (pp. 18–19)
    Just outside the AI's range, the same consultants were 19 points more likely to get the answer wrong — and their wrong answers were rated more persuasive.
    Caveat the outside-frontier task was engineered to fool the model; proves the mechanism, not the frequency.
  • Bottom-half performers gained 43%, top half 17%
    jagged-frontier (p. 15)
    AI compresses skill gaps: the biggest lift goes to the weakest performers.
    Caveat all participants elite in absolute terms.
  • One person with AI matched a two-person expert team
    Dell'Acqua et al. 2025, 776 P&G professionals (cybernetic-teammate, p. 16)
    In a P&G field experiment, individuals with AI matched cross-functional teams without it — in 16.4% less time.
    Caveat one-day ideation task; not implementation.
  • Teams with AI ~3x as likely to produce top-10% solutions
    cybernetic-teammate (p. 19)
    Breakthrough ideas still cluster in teams — AI made teams roughly three times as likely to hit the top decile.
    Caveat same setting.
  • Functional silos dissolved
    cybernetic-teammate (p. 18)
    With AI, R&D people proposed balanced commercial-technical solutions and vice versa — the boundary stopped shaping the thinking.
    Caveat one task, one firm.
  • +14% productivity; novices +34%; experts ~0
    Brynjolfsson, Li & Raymond 2023, 5,179 support agents (genai-at-work, pp. 16–18)
    In the first at-scale workplace deployment studied, productivity rose 14% — almost all of it among novices; the most experienced gained nothing.
    Caveat one firm, one function; staggered rollout, not full RCT.
  • Two months with AI = six months of experience
    genai-at-work (p. 19)
    AI-assisted agents at two months matched veterans at six — the experience curve compressed by two-thirds.
    Caveat same setting.
  • Attrition down ~40% among newer workers
    genai-at-work (p. 28)
    Where AI made work more manageable, new-hire attrition fell by about 40%.
    Caveat selection effects possible; authors flag it.
  • 37% faster, quality +0.45 SD; 68% submitted unedited
    Noy & Zhang 2023, 444 professionals (noy-zhang-writing, pp. 4, 6)
    In MIT's controlled experiment, professionals were 37% faster with better output — and most submitted the machine's first draft untouched, even when paid extra for quality.
    Caveat short, self-contained writing tasks; early ChatGPT; the authors themselves flag inflation risk.
  • Experts 19% slower while believing they were 20% faster
    METR 2025 RCT, 16 experienced open-source maintainers (metr-devs, pp. 1, 3)
    In the most rigorous trial on real expert work, developers using AI took 19% longer — and still believed it had sped them up by 20%. Perception and measurement pointed in opposite directions.
    Caveat n=16, million-line repos the developers knew deeply, early-2025 tools; METR has since published newer results. Quote for the perception gap, not as "AI slows people down."
  • Same tool, same task: 60% Cyborgs, 14% Centaurs, 27% Self-Automators
    Randazzo et al. 2025, 244 BCG consultants (cyborgs-centaurs, p. 14)
    Given identical AI and identical stakes, a quarter of professionals handed over both the analysis and the judgment — and built neither domain nor AI skill. Adoption dashboards can't see the difference.
    Caveat single session; skilling inferred from process and interviews.
  • Persuasion bombing: 968 persuasive responses, escalating after challenge
    Randazzo et al. 2025, 72 consultants (power-persuader, p. 14)
    When professionals fact-checked the AI, it didn't concede — it escalated persuasion. Checking is now contested terrain.
    Caveat April-2023 GPT-4, engineered-error task, qualitative counts.

Tier 2 Large-scale surveys and usage data

  • 95% of organisations report zero P&L return on $30–40B
    MIT NANDA, The GenAI Divide, 2025 (ai-report-2025, p. 3)
    MIT-affiliated research found 95% of enterprise gen-AI initiatives producing no measurable P&L impact — and located the barrier in organisational design, not the technology.
    Caveat always attach the definition — self-reported, ~6-month window, preliminary report, N=52 interviews + 153 surveys. Never quote the bare number.
  • Shadow AI: workers at 90%+ of companies vs. 40% official adoption
    ai-report-2025 (p. 8)
    Your people crossed the divide already — official tooling is behind your own workforce.
    Caveat same report's method.
  • 88% experimenting; 81% no meaningful bottom-line gains
    McKinsey State of Organisations 2026, n=10,018 (mckinsey-state-orgs, pp. 6, 10)
    Nearly every organisation is experimenting; very few can find it in the P&L. The gap is the organisation, not the model.
    Caveat perception survey; 88/81 imported from other McKinsey research.
  • 92% increasing AI investment; 1% call themselves mature
    McKinsey Superagency 2025 (mckinsey-superagency, p. 3)
    Investment is universal; maturity is at one percent — the constraint isn't budget.
    Caveat US C-suite n=118, $1B+ firms; "1%" is roughly one respondent.
  • Leaders underestimate employee AI use 3x; 2.4x more likely to blame employees than their own alignment
    mckinsey-superagency (pp. 4, 13)
    C-suites think 4% of their people are heavy users; the people say 13%. And leaders blame employee readiness at 2.4 times the rate they name their own alignment.
    Caveat estimate-of-others vs. self-report comparison.
  • 48% of employees want formal training; 22% report minimal or none
    mckinsey-superagency (p. 15)
    Employees' top ask is training; a fifth get nearly none. Readiness is being wasted.
    Caveat self-report.
  • Clarity collapses down the line: 56% → 44% → 27%
    mckinsey-state-orgs (p. 43)
    Barely a quarter of middle managers are clear on the must-win battles their C-suite thinks are set. AI accelerates whatever you actually have — clarity or fragmentation.
    Caveat perception survey.
  • 86% say their organisations are not very prepared to run AI day to day; one in six has no C-level AI owner
    mckinsey-state-orgs (p. 10)
    The gap between AI ambition and AI operations is a leadership-structure gap.
    Caveat self-report.
  • ~75% of roles need new skill mixes; AI-fluency demand up sevenfold in two years
    mckinsey-state-orgs (p. 17)
    Three-quarters of today's roles are being re-specified around new skill mixes.
    Caveat McKinsey estimate.
  • Delegation depth is set by work design, not the model: +0.26 autonomy on the same model; 13 dialogue rounds vs. one prompt
    Anthropic Economic Index, 2026 (econ-index, pp. 14–16)
    Anthropic's own usage data: how deeply people delegate to AI depends more on the surrounding work design than on the model. Delegation depth is a design decision your organisation is currently making by default.
    Caveat one vendor's users; behaviour, not outcomes.
  • Over a third expect AI to do most of their tasks within 12 months
    econ-index (pp. 19, 22)
    Inside your workforce, expectations are running far ahead of most transformation plans.
    Caveat self-selected active AI users.

Tier 3 Theory and concepts (frames, not findings)

  • Complementarities: unprofitable alone, profitable together
    Milgrom & Roberts, American Economic Review 1990 (milgrom-roberts, p. 6)
    Economics named this in 1990: changes can each lose money taken alone and be highly profitable taken together. That's why piecemeal transformation stalls — on schedule.
  • The pilots-that-stall pattern is predicted
    milgrom-roberts (p. 15)
    Permanent islands of automation aren't an equilibrium — a long plateau of pilots signals forfeited returns, not prudence.
  • Lockheed: 52 days → 2 days; 80% of the gain organisational
    milgrom-roberts (p. 4)
    When Lockheed cut a 52-day process to two, the project manager credited 80% of the gain to organisational change, not the equipment. Who owns your 80%?
    Caveat illustrative case within a theory paper.
  • ~$12 of market value per $1 of visible IT investment; the dip is the investment
    Brynjolfsson, Rock & Syverson (productivity-j-curve, pp. 29, 41)
    In past technology waves the organisational investment behind each visible technology dollar was estimated up to ten times larger. Your budget line is the small part — and early flat returns are what mid-build looks like.
    Caveat model-inferred from market values; and not every dip is hidden value.
  • "The more transformative the new technology, the more likely its productivity effects will initially be underestimated."
    productivity-j-curve (p. 41)
  • Electrification took a generation of factory re-layout
    productivity-j-curve (p. 4)
  • Moral crumple zones: "While control has been effectively distributed, responsibility has not scaled accordingly."
    Elish 2019, peer-reviewed (moral-crumple-zones, p. 10)
    Accident research has a name for what happens when systems fail: the nearest human absorbs the blame. Deciding where responsibility sits is leadership work, before rollout.
    Caveat concept from safety-critical domains; extension to AI agents is extrapolation.
  • Susskind's three limits; "there is no economic law that says this outcome must be so"
    Susskind 2025, Knight Institute (what-remains, pp. 8, 17)
    What stays human won't be settled by what AI can't do — those lines keep collapsing. It will be settled by what people value humans doing and what we decide humans ought to do. That is a design choice, and no economic law guarantees the comfortable version.
  • The process/outcome test
    what-remains (pp. 8, 12)
    For every 'our clients need the human touch': do they value the human process, or just the outcome? If just the outcome, it's automatable however human it feels. The distinction has to be made deliberately, role by role.

CEO conversation kit

The working material for the room: openers that start the conversation you want, the reframes that separate your ground from the CTO's, counters to the objections you'll actually hear, and the honesty layer that makes all of it credible.

Openers

Grouped by the moment in the conversation.

To open the topic:

  • "Your people's work is now split by an invisible line: on one side AI makes them roughly 40% better; on the other, faster, more persuasive — and wrong. Who owns knowing where that line runs as it moves?" jagged-frontier
  • "If one person with AI now performs like a team, what are your teams for — and who is deciding that today: you, or a thousand individual tool choices?" cybernetic-teammate
  • "In every AI-touched workflow, two questions are being answered right now: who decides what needs to be done, and who decides how. Your people answer them privately, task by task. What would it take for your leadership team to answer them deliberately?" cyborgs-centaurs

When they believe it's handled:

  • "Your teams say AI is making them faster. In the most rigorous trial we have, experts were 19% slower — and stayed convinced they were 20% faster. What in your organisation would catch that?" metr-devs
  • "When your best people push back on an AI's analysis, the system doesn't concede — it argues harder. Who on your leadership team has defined what 'checked' now means?" power-persuader
  • "McKinsey found employees three times further into AI than their C-suites assume — and C-suites still blaming employee readiness. If I asked your top team where adoption is stuck, would I get one answer or five?" mckinsey-superagency

When they want to wait:

  • "When Lockheed cut a 52-day process to two days, the project manager credited 80% of the gain to organisational changes, not the technology. In your AI effort, who owns your 80%?" milgrom-roberts
  • "If your AI transformation is working, your productivity numbers should look unimpressive for a while. What are you measuring that would distinguish an investment trough from failure?" productivity-j-curve

To go deeper:

  • "AI made two-month hires perform like six-month veterans — by bottling what your best people know. Who owns what happens next to your best people: their pay, their role, their reason to stay?" genai-at-work
  • "In McKinsey's 10,000-leader survey, 56% of executives are clear on the must-win battles — 27% of middle managers are. When AI speeds up execution, which does it accelerate: your clarity or your fragmentation?" mckinsey-state-orgs
  • "When your first AI-agent failure goes public, who stands in the crumple zone — and did you put them there by design or by default?" moral-crumple-zones
  • "Which decisions here should a human still make even when the machine is demonstrably better — and can your top team say why, in one sentence each?" what-remains
Reframes — the CTO's frame vs. yours

The CTO's questions are legitimate; they are simply not these. Each pair is a doorway from the technology conversation into the leadership conversation.

  • What can AI do? → What must this organisation become for that to matter? The capability question has answers; the becoming question has no technical owner. [S1]
  • Adoption is up. → Which mode of use is spreading? Same tool, same task: some users compound expertise, some hollow it out — dashboards can't tell them apart. cyborgs-centaurs
  • Roll out the tool. → Name the cluster that moves together. Returns live in complementary changes — workflow, decision rights, skills, incentives — and per-department ROI understates the joint move. milgrom-roberts
  • Show me the ROI. → Show me the intangible formation. In every prior technology wave the organisational investment dwarfed the visible spend and the returns lagged by design. Measure what's being built, not just what's being booked. productivity-j-curve
  • We'll train everyone. → Where does judgment get produced now? Training moves tool skill; the apprenticeship pipeline that produced senior judgment is what's breaking. [T5]
  • A human reviews every output. → Verification is an architecture, not a person. Checking happens inside a conversation designed to persuade; move it outside, give it owners and forums. power-persuader
  • We ran the change program. → There is no other side. Transformation is the operating condition; the capacity you need is a leadership team that can keep re-deciding what the organisation is. mckinsey-state-orgs
  • Our people aren't ready. → Your people are ahead of you. Employees are further in than leadership believes; readiness at the top is the bottleneck. mckinsey-superagency
Objection counters
  • "This is a tech project — my CTO owns it." Identical tools produce opposite outcomes depending on work design, team structure, and mode of use — variables no deployment plan governs. The best org-level data names organisational design, not integration or budget, as the dominant barrier; even McKinsey, which sells technology programs, concludes the challenge "is not a technology challenge." [jagged-frontier, cyborgs-centaurs, ai-report-2025, mckinsey-superagency]
  • "ROI is unclear — we'll wait." The returns of general purpose technologies are systematically invisible early: they accrue as unmeasured organisational assets, and the more transformative the technology, the deeper the early underestimate. Waiting doesn't avoid the trough; it schedules it later, behind competitors who started the intangible build. Guard the other flank too: demand indicators of intangible formation, so the J-curve never becomes an alibi. [productivity-j-curve, milgrom-roberts]
  • "We'll hire our way through it." Twice wrong: judgment gets priced up as it gets scarce, and everyone is hiring from the same pipeline — the one AI is dismantling by absorbing the practice work juniors learned on. The alternative is deliberate internal development of judgment, which no vendor sells. [genai-at-work, noy-zhang-writing, juniors-seniors]
  • "Our people use it and say it works." The one randomized trial on real expert work found user testimony wrong about the sign: 19% slower, believed 20% faster. Self-report measures belief. If the decision matters, build ground truth. metr-devs
  • "We have a human in the loop." In the field data, the loop is where failure happens: the machine escalates persuasion when checked, review incentives don't produce editing, and accident research shows the human in the loop absorbs the blame when the system fails. The loop has to be designed — information, authority, skill, and accountability aligned — or it's a liability, not a control. [power-persuader, noy-zhang-writing, moral-crumple-zones]
  • "We'll do it department by department." Complementarities: each department's business case understates the coordinated move, because the return to each change depends on the others. That's why pilots stall — predictably, since 1990. milgrom-roberts
  • "Our business is relationships — we're safe." Most "needs the human touch" claims confuse the process with the outcome. If clients value only the outcome, it's automatable however human it feels. The durable ground is where people value the human process itself, or where you deliberately decide a human ought to own the decision — and that's a design choice to make, not a property of the industry. what-remains
  • "Our employees aren't ready." Backwards, per the data: employees are roughly 3x further in than C-suites estimate, they trust their employer most to deploy AI responsibly, and their top ask is training they aren't getting. The readiness gap is at the top. [mckinsey-superagency, ai-report-2025]
Where the evidence ends — the honesty layer

Say these out loud before anyone else can. They are what make the rest credible.

  • The field experiments are narrow. Single firms, elite populations, short tasks, 2023–25 models. They establish mechanisms — compression, jaggedness, persuasion, perception gaps — not universal magnitudes. Quote populations and dates.
  • The failure statistics are soft. "95% zero return" is self-reported, ~6-month-window, preliminary-report data; "81% no gains" is a perception survey. They indicate a real pattern; they are not measurements of it. Attach definitions.
  • The expertise-pipeline claim is mechanism convergence, not outcome data. Five independent literatures point at the same break; nobody has longitudinal proof yet. Present it as an unusually well-triangulated warning.
  • The inner-development link is the thinnest layer. Reflection-performance findings are correlational consultancy data; this corpus contains no adult-developmental empirical work. The bridge from "work is changing" to "the leadership foundation must be developed" is your practice-grounded synthesis — label it as such. Precision about where evidence ends is the differentiator against every hype-peddler in the market.
The one-breath version

Everything above in three sentences: The evidence says AI works at the task level and fails at the organisational level, and it locates the difference in things only leadership can move — complementary redesign, decision rights, verification architecture, the expertise pipeline, and the top team's own clarity. A CTO can answer what AI can do. The question with no technical owner is what the organisation and its leadership foundation must become — and that is the conversation HEI exists to open.

Source card