Micro-Productivity Trap

Confidence 0.95 · 52 sources · last confirmed 2026-08-31

A failure pattern in enterprise AI deployment, named in Dutt, Rapoport, Chatterji, Weeratunga & Satcher (2026). Firms invest heavily in AI, treat it as plug-and-play SaaS with isolated use cases, and realize task-level productivity gains that fail to translate into firm-level business value — because the surrounding workflow still depends on tacit knowledge, manual handoffs, or legacy systems not built for AI.

The framing distinguishes this article’s argument from generic “ROI of AI” complaints by giving the failure mode a structural name and two named sub-modes.

Working definition

“AI can accelerate a task, but unless companies address workflow bottlenecks, productivity gains may not translate into business value.” — Dutt et al. 2026

Two specific lock-ins, frequently co-occurring:

Lock-in	Definition
Offering lock-in	AI applied to optimize the existing product/service rather than reframe what value the firm provides
Process lock-in	AI applied to automate current processes step-by-step rather than rebuild around an outcome

The escape from the trap is “reinvent the business”: an organization-wide, future-focused, outcome-oriented posture that explicitly assumes “we live in a world in which powerful AI tools exist” and rebuilds workflows on that premise.

Key claims

Empirical anchor: 10–25% EBITDA gains for the transformation mindset

Bain client work: firms that escape the trap and adopt the four-step transformation framework (below) report 10–25% EBITDA gains that continue to scale as programmes scale.

The four-step framework for escape

1. Narrow possibilities strategically — pick 4–5 critical domains. Top-4 across Bain’s portfolio: software development, customer support, knowledge worker efficiency, marketing.
2. Reimagine workflows across the organization — process redesign is “the most challenging part of AI deployment” and “creates most of the value.” Cross-functional, not silo-by-silo.
3. Engage those closest to today’s process — top-down + bottom-up; prototyping culture in all business areas, not just the tech team.
4. Measure what matters — concrete business-outcome metrics with non-AI baselines, plus continuous evals (evaluation suites for AI behaviour).

Worked cases

• Lowe’s + OpenAI partnership (Mylow / Mylow Companion, March 2025): >2× online conversion when customers engage Mylow; +200 bps customer satisfaction when associates use Mylow Companion; deployed across 1,700+ stores.
• FabricationCo (anonymized Fortune 1000 manufacturer; Bain client): 14 use cases identified, ~$30M additional profit on track; quote-generation ~15× faster; +10pp win-rate increase in 3 months.

The boardroom imperative

Lagging firms recognize AI as important but delegate it to technology groups without specific goals or metrics. The article argues this delegation pattern is a structural cause of the trap — AI transformation requires top-of-house ownership with ambitious, business-owned targets.

Second-source corroboration: data architecture as load-bearing (Nishar & Nohria 2026)

Independent of Bain/OpenAI’s framing, Deep Nishar and Nitin Nohria (HBR.org Digital, April 2026) reach the same conclusion via a different route:

“Moving quickly is not the same as moving effectively. Organizations that rush to automate without re-architecting their data and processes often find that quality suffers, edge cases accumulate, and systems become difficult to manage. The lesson is not to slow down, but to pair ambition with discipline: to treat data architecture, governance, and ownership as integral to the transformation rather than afterthoughts.”

The mechanism Nishar-Nohria name as the trap is the same one Bain/OpenAI named as process lock-in: applying generative-ai to existing workflows without restructuring the data layer or the workflow itself. Their prescription is consistent with the “reinvent the business” posture: treat data architecture as the foundation, revisit build-vs-buy per function, govern internally-developed systems, and recognize that software choices are workforce choices.

The two sources together — practitioner consulting (Bain/OpenAI) and practitioner investor + academic (Nishar/Nohria) — converge on data architecture + process redesign first, technology second as the load-bearing decision. Source overlap is low (different vantage, different audience), so this is real cross-source agreement, not an echo. Confidence on this concept lifts from single-source 0.70 to two-source 0.75.

Third-source corroboration: leadership response patterns (Carucci 2026)

Ron Carucci’s practitioner essay reframes the trap from a third vantage — leadership behavior. His four-category typology of resistance includes a category that maps directly to the operators-see-problems-leaders-dismiss pattern of process lock-in:

“Sometimes resistance isn’t about emotion at all. It’s about execution. Employees may see practical issues leaders have overlooked, such as: timelines that aren’t realistic, processes that don’t match how work actually gets done, conflicts with other ongoing initiatives. But when leaders dismiss resistance too quickly, they also dismiss the possibility that the change itself needs refinement… The very people resisting the change are the ones closest to the risks that could derail it.”

Carucci’s prescription matches the Bain/OpenAI and Nishar-Nohria diagnoses in spirit but operates at the leadership-behavior layer rather than the strategy or transformation-architecture layers:

• Get curious before you get corrective — treat pushback as potential insight; ask follow-up questions to understand the operational reality behind the concern.
• Separate the signal from the delivery — feedback may come wrapped in frustration or bluntness, but tone shouldn’t disqualify content.
• Be willing to adapt the plan — strong change efforts evolve.

Three independent sources, three vantages (consulting practice, investor + academic, executive coaching), three vocabularies — all converging on the same point: the trap is fundamentally a listening failure, not a technology failure. Confidence lifts from 0.75 (two-source) to 0.80 (three-source agreement across distinct vantages).

Fourth- and fifth-source corroboration: the operational mechanism (Ransbotham et al. 2024 + Kiron & Schrage 2026)

Two MIT SMR sources from a fourth vantage (academic / journalistic) supply both the empirical scale and the operational mechanism of the trap.

Empirical scale: 3,467-respondent MIT SMR × BCG survey finds 59% of organizations are “Limited Learners” — low on both organizational learning capability and AI-specific learning capability. Only 15% are Augmented Learners combining both. The 59% Limited-Learner majority is the empirical correlate of being stuck in the trap; the 15% Augmented-Learner cohort is what escape looks like at population scale.

Operational mechanism: Kiron & Schrage 2026 specify what process lock-in looks like in operational detail:

“Most organizations treat AI outputs as verdicts to be confirmed rather than starting points to be interrogated. The result is consumption dressed up as adoption — verification mistaken for the whole job.”

The flywheel that escapes the trap:

Step	Question	Distinction
Verification	Does this output meet the standard?	Binary, against existing criteria. Necessary but insufficient.
Evaluation	What does this output reveal?	May generate new standards. Requires domain expertise. Where most orgs fail.
Learning capture	How do we ensure this insight persists?	Version control for organizational judgment (e.g. CLAUDE.md-style files).

Why this matches the Bain/OpenAI diagnosis: Bain/OpenAI’s process lock-in (AI automates current processes without redesign) and Nishar-Nohria’s data-architecture-as-afterthought both name the same trap — applying AI to the output stage of a workflow without changing how the organization metabolizes AI’s outputs. Kiron-Schrage make the metabolism mechanism explicit: orgs that don’t evaluate (only verify), don’t capture (let insights evaporate after each session), or both, never compound.

Statistical anchor for the operational claim: orgs that build systematic feedback loops between humans and AI are 6× more likely to derive substantial financial benefits; orgs investing in learning with AI are 73% more likely to achieve significant financial impact (Ransbotham et al. 2024 cited in Kiron-Schrage 2026).

Five concrete moves (from Kiron-Schrage), each correcting a specific failure mode of the trap:

1. Preserve evaluation expertise — domain experts repositioned as evaluators, not removed because “AI can do that now.”
2. Build minimally viable verification — start the cycle even when full verification is expensive (multijudge systems, consistency checks).
3. Institute evaluation practices — “What worked? What failed? What was interestingly wrong?” The third question converts AI-output anomalies into surfaced tacit expertise.
4. Create capture systems — decision journals, prompt repositories, evaluation logs. Discipline, not cost or creativity, is the true constraint.
5. Measure the cycle, not just the output — count verifications, evaluations, learning captures, and how quickly captured learning changes practice.

Five independent sources now (Bain/OpenAI, Nishar-Nohria, Carucci, Ransbotham et al., Kiron & Schrage), four vantages (consulting practice, investor/academic, executive coaching, academic survey + practitioner column), all converging. Confidence lifts to 0.85 (five-source agreement; the new sources add the operational mechanism missing from the prior three). The trap is robustly named; what remains is measurement — how would an external auditor distinguish a verification-only org from a verify-evaluate-capture org without insider access?

Sixth- and seventh-source corroboration: agent-harness practitioner essays (Kokane 2026 + Chatterjee 2026)

Two practitioner essays on agent harness engineering, ingested concurrently, supply a fifth vantage (working AI engineers / product PMs at the runtime level) that independently arrives at the same operational mechanism the prior five sources named.

Chatterjee 2026’s fourth layer — Compounding — is operationally identical to Kiron-Schrage’s verification → evaluation → learning capture flywheel:

Kiron-Schrage step	Chatterjee equivalent
Verification — does this output meet the standard?	Constraints layer — pre/post-tool hooks score outputs against criteria
Evaluation — what does this output reveal?	Contracts layer — formal evaluable specifications, score breakdowns, “your last synthesis scored a C, here is why”
Learning capture — how do we ensure this insight persists?	Compounding layer — telemetry as training data for the harness; nightly meta-learning loop proposes harness adjustments; human-reviewed approvals become workspace overrides

The construct is now visible both at the organizational learning level (Kiron-Schrage’s CLAUDE.md as version control for organizational judgment) and at the runtime engineering level (Chatterjee’s structured telemetry → harness tuning → workspace overrides). Same mechanism, two scales — strong cross-source convergence.

Kokane 2026 adds the sceptical-vantage corollary: most of the work that escapes the trap is mature systems engineering (retries, state machines, idempotency, observability) applied to a new substrate. The 10% genuinely novel — non-determinism at the execution layer + context as a degrading resource — is precisely what cannot be solved by checklist; it is where domain expertise and human evaluation are non-substitutable.

Chatterjee’s Friday-in-March story as the wiki’s clearest worked example of the trap’s failure mode at the engineering layer: an agent emptied a customer’s workspace because the harness lacked an intent-validation layer. “The model was not the problem. The problem lived in the layer around the model.” Operationally identical to “verification masquerading as evaluation” but at a different stack layer — the same diagnostic principle (failures are in the layer around the model, not the model itself) generalises.

Seven independent sources now, five vantages (consulting / investor + academic / executive coaching / academic survey + practitioner column / runtime engineering practitioner essays), all converging. Confidence lifts to 0.90 — among the wiki’s most strongly corroborated concepts. What remains genuinely open is quantitative measurement: nobody has yet measured the slope of harness-tuning compounding (how much contract-score uplift per unit of telemetry-driven adjustment, over what timeframe). The trap is robustly named and its escape mechanism is robustly described; what is missing is empirical confirmation that the escape mechanism delivers on its promise at scale.

Eighth-source corroboration: the AWS-vendor-altitude pilot purgatory anchor ( AWS London Exec Forum 2026 + MIT NANDA 95%)

Jonathan Allen at the AWS London Executive Forum 2026 anchors the trap with the MIT NANDA enterprise-investment report headline (April 2026, not yet ingested): “Despite 30 to 40 billion in enterprise investment to GenAI, this report uncovers a surprising result, and that 95% of organizations are getting zero return.” Allen names three operational failure modes mapping cleanly onto this concept’s process-lock-in / offering-lock-in taxonomy:

• Lack of workflow alignment — “AI bolted on is going to fail”. The exact phrase the wiki has held since Dutt et al. introduced process lock-in.
• Wrong targets — “companies aiming at sales and marketing high visibility instead of back-office operations where there’s higher ROI are going to struggle.”
• Missing integration — “pilots stay as demoed, never connected to actual business decision flows. Pilot purgatory.”

Allen’s 5%-pattern counter-prescription (pick one point, execute well, right team composition, integrated into business decision flows from day one) is the AWS-vendor-altitude restatement of the four-step transformation-mindset framework from Dutt et al.. Nick Francis’s Brooklyn Solutions testimony in the same talk ratifies the workflow-level focus via the frequency × variability use-case discovery workshop method — Brooklyn sit with SMEs from every department to find “things that are mundane, high-frequency, not that variable, that happen day in, day out”. The Brooklyn case is also a successful micro-productivity-trap-escape worked example — phase-1 task gains (document summarisation, key-point extraction) compounded into phase-3 agentic execution within the application via iterative data-context architecture, not bolt-on.

Ninth-source corroboration: anthropomorphizing lock-in as a third named sub-mode ( BCG, HBR May 2026)

Dutt et al. named two sub-modes of the trap (offering lock-in and process lock-in); Kropp et al. supply a third — anthropomorphizing lock-in. In their N=1,261 randomized experiment, the AI-as-employee framing (vs AI-as-tool, holding everything else constant) caused 9pp drop in personal accountability (8pp shift to AI), 44% more escalation requests, 18% fewer errors caught, and 13% higher identity uncertainty — without meaningfully increasing adoption intent. The mechanism is structurally a third sub-mode in the Dutt et al. taxonomy: an apparently-cultural choice (how to refer to AI in the org) causes operationally consequential value leakage at the firm level. BCG and Bain converging at near-identical timing on the same trap-and-mechanism construct, at advisor altitude, with overlapping prescription sets (workflow redesign + explicit accountability + capability-building for human managers of agents).

Tenth-source corroboration: the academic-altitude RCT showing untailored augmentation gradually decreases market share (Krakowski et al. 2025)

Where the prior corroborating sources operate at consultant / vendor / practitioner altitudes, Krakowski et al. supply the wiki’s first strictly-academic RCT-grade empirical anchor on the trap. N=72 Nordic pharma sales experts, DiD design, three conditions (legacy IT control / untailored AI / tailored AI per Kirton 1976 cognitive style). The headline result is itself a worked instantiation of the trap: untailored AI deployment causes market share to gradually decline vs a stable legacy-IT baseline — task-level technology adoption happened (the system was deployed), but firm-level value decreased over 9+ quarters post-intervention. The mechanism (per qualitative interviews + login mediation): sales experts withdraw from the system (utilization decreases over time in the untailored condition), driven by role conflicts and ambiguities. The Finnish/Norwegian innovator quotes (“felt like in prison… killed my internal drive!” / “felt like a less good sales representative”) name an internal-motivation-collapse fifth sub-mode alongside Dutt et al.’s offering-lock-in / process-lock-in, Kropp et al.’s anthropomorphizing-lock-in, and the broader interaction-design-misfit lock-in that the four-parameter (work-procedure / decision-authority / training / incentives) framework names operationally.

Eleventh-source corroboration: the MGI <40%-of-90% statistic at structural-modeling altitude ( MGI 2025)

MGI’s Agents, Robots, and Us (November 2025) supplies the wiki’s most quantitatively-precise empirical anchor on the trap. Chapter 3 opens with the diagnostic:

“Nearly 90 percent of companies say they have invested in [AI], but fewer than 40 percent report measurable gains. The gap may reflect the fact that many projects are still in pilot or trial phases or that organizations are applying AI to discrete tasks rather than redesigning entire workflows.”

The <40%-of-90% statistic is the workflow-redesign-disciplines-against-task-automation empirical anchor at MGI’s structural-modeling altitude — more quantitatively precise than AWS’s MIT NANDA 95% pilot-fail (different denominator: 95% of pilots fail to scale into production; MGI’s denominator is companies that have invested, with <40% reporting measurable gains — a broader and arguably less stringent measure). Three substantive contributions to this concept:

(a) The diagnostic framing — workflow vs task as the difference between failure and value capture: “In banking, for example, this would be the difference between offering employees access to a chatbot for ad hoc use and deploying custom agents alongside people in a reimagined process to approve, process, and manage loans more efficiently and deliver better customer service. Unlocking larger productivity gains from AI will require reimagining workflows along the lines of the latter, rather than taking a task-based approach.” The framing converges with Bain’s offering-lock-in and process-lock-in but at the workflow-design layer rather than the strategic-business-design layer.

(b) 190+ workflows mapped as the operational unit-of-analysis. MGI’s sidebar “An early view of workflows across the US economy” maps 190+ workflows across 16 business functions (~100 cross-cutting workflows in commercial/functional/operational domains + ~90 sector-specific workflows in knowledge/frontline/production services). This is the wiki’s most granular publicly-available workflow taxonomy. The implication: the trap’s exit strategy is workflow-by-workflow identification + redesign, not firm-wide AI strategy. ~60% of potential gains concentrate in sector-specific workflows ($1.7T of the $2.9T claim), 40% in cross-cutting ($1.2T) — the load-bearing prioritisation guidance is sector-specific workflows first.

(c) Four operational case studies operationalising the trap’s escape:

• B2B sales redesign (5 agents — prioritisation / outreach / customer response / scheduling / handoff) → 7-12% revenue increase + 30-50% time saved across sales roles.
• Utility customer operations (4 agents — inbound call / intent ID / scheduling / self-service) → handle ~40% of all calls, resolve >80% without human involvement; avg cost per call cut ~50%; CSAT +6pp.
• Biopharma medical writing (6 agents — planning / data mapping / drafting / validation / reviewing / submission) → touch time for first human-reviewed drafts -60%, errors -50%.
• Regional bank IT modernization (3+ agents — modernisation planning / assessment / functionality / coding / QA / testing) → code accuracy up to 70%; required human hours -50%.

These four cases form the wiki’s strongest practitioner-validated escape patterns from the trap at MGI structural altitude. The cross-source convergence — Bain (workflow lens), McKinsey/Sternfels (org-change-dominates), Allen/AWS (workflow-not-bolted-on), MGI/Yee (190-workflow taxonomy + 4 worked cases) — places the workflow-vs-task framing on the wiki’s strongest cross-consultancy structural footing.

Cross-source positioning (descriptive only, per cross-source neutrality)

The micro-productivity-trap framing sits among several wiki vocabularies addressing the same broad territory of “AI adoption breadth ≠ transformation depth”:

• Anand-Wu’s 2×2 — where to deploy GenAI (cost-of-errors × type-of-knowledge).

• Tin-Man — adaptive vs predictability-optimized organizational design.

• MIT CISR Four Stages + Four S — staged maturity (28%/34%/31%/7% distribution; only 7% Stage 4) and Strategy / Systems / Synchronization / Stewardship challenges.

• Cisco — chatbot → agent → multi-agent progression and 5-foundations readiness framework (only 13% AI-ready).

• Bansal & Birkinshaw systems thinking — flows and relationships over products/services.

• Warner & Wäger — digital sensing/seizing/transforming nine microfoundations under Teece.

• McKinsey “Rewired” 2nd ed (2026) — 6 capabilities (business-led roadmap, talent, operating model, technology, data, adoption-and-scaling); 20% EBITDA uplift / $3:$1 / 1–2yr breakeven across ~20 deep-dive AI-leader companies; 70% talent-density shifts.

• AI Index 2026 — empirical context: 88% organizational AI adoption but AI agent deployment in single digits per business function.

• Nishar & Nohria 2026 — firm-boundary 4-model lens (Build / Compose / Collaborate / Buy Outcomes); the trap manifests when a firm rushes to “Build” or rushes to automate any layer without first treating data architecture, governance, and ownership as the foundation.

• Carucci 2026 — human-reaction lens on the trap; the Flaws-in-change resistance signal is operators surfacing exactly what process lock-in misses; leaders’ three traps (personalize / moralize / rush) systematically suppress it.

• Ransbotham et al. 2024 + Kiron & Schrage 2026 — organizational-learning lens on the trap; 59% of orgs are “Limited Learners” stuck in the trap; the verification → evaluation → learning capture flywheel is the operational machinery to escape it. ROI reframed as return on iteration.

• Chatterjee 2026 + Kokane 2026 — agent-harness engineering lens on the trap; the agent harness is the runtime infrastructure where the trap is escaped or perpetuated. Friday-in-March story makes the failure-mode concrete; “build constraints before you build cleverness” is the operational discipline.

• McKinsey) — consulting-vendor-self-narrative lens on the trap; the second consulting firm (after Bain via Dutt-Chatterji 2026) to name the same diagnosis from inside the engagement: “Half, if not more, of the secret sauce is organizational change as opposed to technology implementation.” Sternfels’ CFO-vs-CIO truth-room dilemma captures the trap from the client’s seat: “My CFO is in my ear that we’re spending a lot of money on technology, but we’re not yet seeing enterprise-level value. … CIO’s saying: this is one of those moments. And if we’re not in the lead, we’re going to get disrupted.” Two of the world’s largest consulting firms converging on identical diagnoses is structural evidence that the trap is real and pervasive — not just a Bain framing or a McKinsey framing. Operational escape mechanism Sternfels names: flatter organizations cutting middle layers + workflow consolidation across departmental boundaries (his mortgage-process worked example: origination + credit-scoring + collection + after-service collapse from 4–5 departments to one AI-enabled flow once the boundaries are removed).

• Böckeler 2026 (Thoughtworks at QCon London) — practitioner-consultancy developer-experience lens on the trap, completing a three-firm convergence (Bain + McKinsey + Thoughtworks) from three structurally distinct vantages: client engagement engineering (Bain/OpenAI), global-managing-partner self-narrative (McKinsey), and cross-client developer-tooling observer (Thoughtworks). Böckeler’s two contributions: (1) a token-economics mechanism for why the trap deepens in 2025–26 — the agentic-coding inner loop (research → plan → review-plan → implement → run tests → fix tests → check lint → fix lint → check browser → code-review subagent → react → summarise) burns $380/day-equivalent tokens for what may be two lines of code, so flat-rate developer tooling has collapsed into metered pricing; (2) the Goldilocks speed counter-framing — speed pressure without harness investment causes corner-cutting and outages, observable in the report that Amazon reacted to AI-code-related outages by adding senior-engineer review gateways, defeating the speed gain. The trap’s escape mechanism Böckeler implies: invest in the agent-harness (feed-forward + feedback, CPU + GPU) before dialling autonomy up. Convergent with Sternfels’ organizational-change-over-technology framing but at a different stack level — Sternfels at the firm-level operating model, Böckeler at the engineering-team feedback-loop level.

• Startup School) — control-systems vocabulary for the trap. Hu names the trap-escape mechanism in control-systems terms that are vendor-neutral and discipline-portable: “Open loops are inherently lossy. A closed loop, on the other hand, is self-regulating. It continuously monitors its output and adjusts its process to better meet the stated goal.” The wiki’s first formal-systems framing of the trap: orgs that run open-loop cannot escape the trap; orgs that run closed-loop continuously monitor + adjust + improve. Operational instantiation Hu names: make the company queryable (record meetings, minimize DMs, embed agents throughout channels, build custom dashboards for everything), use agents to read the artifacts (Linear tickets, Slack channels, customer feedback, standup recordings, sales calls), let the agent propose sprint plans based on what actually shipped vs. customer needs. Reported result across multiple YC companies: “teams that do this cut their engineering sprint time in half and get close to 10× more done”. Also names the productivity-thinking-is-itself-the-trap insight: “Most people talk about AI in terms of productivity … This framing misses the shift we’re currently seeing, which is less about productivity boosts than entirely new capabilities.” Convergent with Kiron-Schrage 2026’s verification → evaluation → learning capture flywheel and with Jassy 2025’s flatten-management initiative, framed in control-systems vocabulary.

• Jassy 2025 (Amazon CEO on HBR IdeaCast) — target-firm CEO operator-of-trap-escape worked example. Where Bain / McKinsey / Thoughtworks / NYT describe the trap from observer vantages, Jassy is a target-firm CEO actively deploying the trap-escape prescription at Amazon scale. Concrete operational datapoints: +15% IC-to-manager ratio target (already beaten in Q1 2025), 1,000+ no-bureaucracy emails received + 375 processes changed in response, 5-day RTO mandate with the explicit hypothesis that “invention is stronger when people are together”. Jassy’s taxonomic refinement of the trap: “there is a difference between process and bureaucracy” — most companies of scale need process (legitimate, scaling discipline); the trap-relevant pathology is bureaucracy (layers that don’t add creative value). The “two-way-door vs one-way-door” decision taxonomy is Jassy’s named mechanism for delegating decisions to ICs. Importantly: this is the earliest-published wiki source on the trap (May 2025) — predating both Bain (May 2026) and McKinsey (Feb 2026) — and is consistent with Jassy recognising the trap experientially before the consulting-firm vocabulary existed. Note Jassy’s epistemic humility: “it’s very hard to measure how well you’re inventing” — he names his RTO defence as essentially un-measurable, which weakens the per-mechanism evidence even as it strengthens the broader operator-implementing-the-prescription signal. The trap is now visible from five distinct observer vantages across five sources (Bain engagement-engineering / McKinsey GMP self-narrative / Thoughtworks consultancy / NYT journalism / target-firm CEO) — and the May-2025 anchor extends the trap’s temporal footprint backward by ~9 months.

• 75-developer field-report) — journalist-observer empirical lens. Two cleanly-quantified datapoints crystallise the trap from the developer-side: small two-person startups report up to ~20× faster delivery (full-day features → ~30 min) with ~100% AI-written code; Google reports ~10% overall speedup despite ~40-50% of code being AI-written. The 2× order-of-magnitude gap between micro and macro speedup at the same per-task AI penetration level is the trap’s clearest single empirical statement in the wiki. Mechanism named: organisational metabolism — change-management overhead, legacy interactions, customer commitments rate-limit how much per-task speed survives into firm-level value. Two further contributions: (1) the 1980s productivity-paradox parallel — computers didn’t increase corporate productivity until firms reorganised work around them; Thompson predicts AI-coding’s industrial impact will be “longer than we think” for the same reason; (2) the hire-back-within-6-months anecdote — Thompson reports that companies who lay off staff thinking AI will replace them often have to rehire within six months because the replacement didn’t work. The trap is now visible at four distinct empirical scales (firm, function, team, individual-developer) and from four distinct observer vantages (Bain engagement-engineering / McKinsey GMP self-narrative / Thoughtworks consultancy / NYT journalism) — among the wiki’s most strongly corroborated patterns. Confidence lifts to the cap (0.95).

• Developer Intelligence team) — engineering-team DORA-grounded vantage on the trap. The talk explicitly cites the DORA (DevOps Research and Assessment) research programme — which Forsgren has led since Accelerate (2018) — for the trap’s most direct engineering-team statement: “there’s some evidence that our DORA team found by studying multiple companies, that increasing AI adoption can lead to individual productivity benefits while at the same time decreasing team-level benefits.” This is the engineering-team correlate of the firm-level trap — individual gains, team-level losses, same divergence mechanism. The talk’s three-shift response for engineering leaders maps directly to escape-from-trap practices: (1) redefine productivity measurement — “stop measuring your teams by PR throughput or lines of code accepted” (the engineering analogue of CFO/CIO metric realignment); (2) actively protect productive struggle — protected learning time so engineers build the shared mental models AI-generated code requires (the engineering-team analogue of Carucci’s get-curious posture); (3) foster radical psychological safety — agentic-workflow experiments must fail without punishment (operationally consistent with Carucci’s resistance-as-data framing). The talk also makes the DORA amplifier-and-mirror framing explicit: “AI is an amplifier and a mirror. It magnifies the existing strengths and it holds up a mirror to those weaknesses.” This is the trap’s clearest one-line diagnostic — AI doesn’t create dysfunction, it exposes and accelerates dysfunction already present. Convergent with all prior nine framings but operates at the engineering-team layer the prior nine did not address directly. The wiki’s first vantage that grounds the trap in a longitudinal published research programme (DORA’s State of DevOps reports, 2013–present) rather than consulting engagements or single-case observations.

• Chamath 2026 (Stanford AI Club) — operator-anecdotal lens on enterprise-scale trap. The wiki’s most concrete single instance of the trap at the legacy-modernization layer: a $100B/year Fortune-1000 customer brings COBOL-code retirees back from retirement to explain what their existing code does, because the current engineering team cannot reconstruct the institutional context. Chamath: “None of you geniuses were able to figure that out. And nobody before you was able to figure it out. And it turns out that these kinds of problems are the thing that’s stopping ROI on trillions of dollars of investment.” The mechanism named: hidden tribal knowledge as the ROI ceiling — task-level AI productivity does not aggregate to firm-level value when the symbolic context of legacy systems remains inaccessible to the AI agents. Convergent with Thompson’s organisational-metabolism mechanism but operating at the institutional-knowledge-archaeology layer that Thompson does not address. Also a sharper-than-prior framing of the trap’s macro implication: “trillions of dollars … if we don’t figure this out, people will hit this trough of disillusionment and say this was a joke” — explicitly linking the trap’s escape to the AI capex trajectory’s social-license-to-continue.

• Warren 2026 (YC Startup School) — YC-altitude vendor-side anti-trap discipline for AI-native services companies. Warren names two trap-escape prescriptions directly relevant to the micro-productivity-trap diagnosis: (1) “Variance is the existential problem … customers will fire you for variance faster than they will fire you for being a bit slower or a bit more expensive than the incumbents. They need to trust the output. Inconsistency destroys trust, which causes churn” — the wiki’s clearest single-source statement that variance-not-throughput is the customer-firing trigger in AI-services. The trap manifests at the vendor-product layer when teams optimise for per-task speed/cost (the micro axis) and ignore output-uniformity-over-time (the firm-value axis customer churn actually responds to). (2) “It’s okay to do things that don’t scale at the very beginning, but eventually you really do need to scale. Automating the process is the product.” — the operational sequencing rule for the augmentation-to-automation transition. Warren also names the early-demand trap as a vendor-side analogue to the buyer-side micro-productivity trap: “It’s easy to sign up a lot of pilot customers when you’re just starting out and have nothing. But it can quickly overwhelm your ability to serve them and you won’t be able to build the product to scale. You’ll be stuck using humans. It is a literal trap.” The wiki’s first vendor-side founder-prescription on the trap, complementing the buyer-side framings (Bain, McKinsey, Thoughtworks, NYT, target-firm CEOs) — convergent with Storoni’s gear-3-reactive-quantity-trap framing applied at vendor-product-design rather than individual-cognition layer. Plus the AI-operating-leverage 30%-services-margins → 50%+ trajectory claim is the escape-velocity quantification of the trap at YC altitude: traditional services firms stay around 30% margins because they can’t escape their own humans-scale-linearly structure; AI-services-vendors can escape it via the harness layer compounding into margin lift over time.

• Giles 2026 (WP Intelligence) — executive-readership news-survey lens with the explicit chatbot-vs-deeper-agent productivity-gain split anchor. Kiva Allgood / WEF San Francisco Center for Advanced Manufacturing (quoted at a WEF manufacturing/supply-chain executive meeting): “Things like chatbots that can be queried for simple answers can give you 2 to 3 percent productivity gains. But if you go deeper with agents and other forms of automation, you’ll see 30 to 60 percent improvements.” The wiki’s clearest single-source numerical statement of the trap’s mechanism at industrial-automation altitude: shallow chatbot deployment captures ~2-3% of the available productivity gain; deeper agent + workflow-integration deployment captures 30-60% — an order-of-magnitude gap that is the trap. Two single-firm anchors give the deeper-end its empirical floor: Siemens reports a 69% labor productivity increase at a flagship Germany electronics factory using AI; Danfoss (global cooling/heating manufacturer) used agents to cut customer-order processing from days to minutes. Both fit the whole-workflow-orchestration-over-deeper-stack pattern Allgood describes. Giles also pairs the numerical anchor with the operational prescription — recommendation 1: Review workflows first, deploy agents second + UiPath Garaba’s worked example (onboarding looked simple, ~60 subprocesses, not all suited to automation). Convergent with Bain/Dutt-Chatterji (workflow-not-task), Sternfels/McKinsey (org-change-dominates), Allen/AWS (AI-bolted-on-fails), MGI/Yee (190-workflow taxonomy), Forsgren-Macvean (DORA team-vs-individual divergence). The Allgood 2-3-vs-30-60 numbers are now the wiki’s most-quotable single-line empirical anchor on the trap’s depth-gradient. Confidence stays at 0.95 (cap); this is mechanism granularity + quotable numerical anchor, not new evidence.

• Storoni 2026 (HBR IdeaCast) — neuroscience-mechanism lens on individual-cognition trap. The wiki’s first brain-level account of why the trap’s organisational pattern (high throughput, low value) emerges. Storoni names the gear-3 reactive state — the high-norepinephrine cognitive state workers enter under continuous Slack/email pressure and deadline-driven workflows — as the cognitive failure mode that feels productive but produces low-quality output. The prefrontal cortex (judgment, planning, nuance-detection) cannot fully engage in gear-3; speed goes up, accuracy goes down, second/third-order consequences get missed. “It’s a trade-off between speed and accuracy … you will make mistakes, errors, and you will miss nuances, and you will miss subtle aspects of anything you’re going through.” The mechanism is the individual-level neurological correlate of the firm-level trap: workers in gear-3 produce the throughput pattern that aggregates to organisational micro-productivity, and that throughput is exactly what AI substitutes for better than humans do — meaning gear-3 workers are systematically placing themselves on the automation side of the automation-vs-augmentation cut by their work pattern, not by their job category. Storoni’s escape mechanism converges with Reitz-Higgins, Sternfels, Kiron-Schrage, and Hu in spirit but at the individual layer: protect gear-2 windows via chronobiology-aware schedules, walking breaks, intrinsic-motivation structures (Google/3M 20%-time as the named exemplar). Convergent with Forsgren & Macvean’s DORA finding that “increasing AI adoption can lead to individual productivity benefits while at the same time decreasing team-level benefits” — Storoni provides the cognitive mechanism for the same divergence pattern that DORA observes at the team-scale. The trap is now visible at six distinct stack layers (individual neurology / engineering team / business function / firm / industry consulting practice / macro capex) and from eleven vantages — among the wiki’s most strongly corroborated patterns. Confidence stays at the 0.95 cap; the new vantage adds mechanism granularity rather than empirical confirmation.

Each names the same broad gap with a different vocabulary; the wiki’s organizational-frameworks-for-ai-adoption synthesis maps the cluster (the synthesis was filed when 6 frameworks were ingested; Nishar-Nohria is the 7th, Carucci the 8th, Sternfels the 9th, and Chamath the 10th, at layers the synthesis didn’t surface).

The agentic-DevOps vendor-keynote articulation of the productivity paradox ( Microsoft Dec 2025)

A GitHub/Microsoft developer-tooling-vendor keynote lands the micro-productivity trap squarely at the individual-developer scale, citing the METR study: “We have the expectation that AI is going to help us go faster. But sometimes you have to go slower to go faster.” Randell names the exact illusion the trap turns on — developers measure slower on average while feeling faster, because “they see more files generated… it’s that feeling of paper-pushing where I’m really busy.” His five-takeaways close is the trap’s prescription in vendor language: “focus on value, not how many lines of code we can generate with AI.” The keynote also supplies the adoption-discipline corollary the trap implies: don’t chase every new frontier model (each switch costs learning time), and “AI is powerful but not magic.” This is the developer-tooling-vendor vantage on the same throughput-≠-value pattern the wiki tracks from Bain (firm level), DORA/Forsgren (team level), and Storoni (neurological level) — notable precisely because it comes from a vendor whose incentive is to sell the acceleration, yet who warns against mistaking motion for value.

Demand 3×, not 20% — radical targets as the escape ( HBR June 2026)

Marco Argenti (CIO, Goldman Sachs) gives the trap its sharpest single-line escape prescription. The trap, in his framing: “Applying AI to streamline old processes, doing more of the same and doing it faster, can yield temporary relief, but will end up massively missing the mark in the long run.” — i.e. optimisation is the trap; transformation is the escape. The lever he proposes is the magnitude of the target itself:

“If you want your developers to change habits, ask them to be 3x more productive, not 20%… If you want to streamline your procure-to-pay process, aim for a 90% reduction of manual touchpoints, not 20%. If you get even halfway there, you will know that your team has at least gone through the motions of radical rethinking, not just optimization.”

The mechanism: a 20% target is reachable by speeding up the existing process (offering/process lock-in intact); a 3× / 90% target is only reachable by redesigning the work — so the radical target forces the process-redesign the trap’s escape requires, and even partial attainment proves real rethinking occurred. This is the wiki’s clearest managerial-lever statement of the escape, complementing Bain’s four-step transformation (firm level) and the workflow-not-task unit-of-value framing: set targets that are impossible to hit by optimisation alone.

Automate first, then reimagine — the CIO-Symposium restatement ( CIO Symposium)

Two of the eleven 2026 MIT Sloan CIO Symposium leaders state the trap’s escape in their own words, at practitioner altitude:

• George Westerman — “automate first, then put the humans in the right places… rather than applying tools to the steps of the existing workflow, [if] we truly get to the agentic world, [we] think about what outcomes do we want, and rebuild those processes that will get us there.” The bolt-tools-onto-existing-steps move is precisely the trap; rebuild-for-outcomes is the escape.
• Monica Caldas — “you’re not just automating and having it go faster. You actually want to have a different outcome… be very thoughtful about the workflow that you’re reimagining,” with clear OKRs and a deliberate maturity arc. The explicit speed ≠ value warning is the trap named from inside an IT-ops deployment.

Both restate, from the CIO chair, the page’s core: task-level acceleration without workflow redesign yields motion, not value. They converge with Argenti’s 3×-not-20% and the Bain/McKinsey reimagine-the-workflow prescription.

Refocus automation on redesign, not cost reduction — the BCG CEO directive ( Emerson, Kropp et al. 2026)

BCG states the trap’s escape as an explicit CEO recommendation: “Refocus automation on redesign, not just cost reduction. Agentic AI isn’t a blunt instrument.” The mechanism BCG names is the measurement failure at the heart of the trap: cost actions (headcount freezes/cuts) are “visible and straightforward, with explicit OpEx impact,” but “when AI drives productivity rather than cuts, ROI becomes harder to define and defend” — so leaders default to the legible cost lever and miss the larger redesign value. The prescription: new domain-specific KPIs that link productivity to outcomes (revenue per FTE, more product shipped, stronger customer impact), and task turnover within a role as a measure of how fast roles evolve toward higher-value work. BCG pairs this with the blunt warning that anchors the trap on the labor side: “Those who cut their workforce beyond AI’s ability to replace it will see productivity drop, institutional knowledge disappear, and critical talent walk away.” This is the consulting-firm CEO-directive form of the page’s reimagine-don’t-optimize thesis — and a sharper account of why firms fall in: the cost lever is measurable, the redesign upside isn’t (yet).

Innovation as a measured obligation — the DBS operator-altitude escape ( DBS June 2026)

Bidyut Dumra’s DBS interview supplies a target-firm operator-altitude worked example of the trap’s escape, alongside Amazon — a leader actively running the escape prescription at 39,000-employee scale rather than describing it from an observer vantage. The DBS mechanism reads as a triple of trap-escape moves the page has named separately:

• Workflow-not-task redesign (Bain/Dutt-Chatterji’s core): Managing Through Journeys rebuilds the operating model horizontally around customer intent, not around bolting AI onto existing functional steps — “a customer is beyond a process, it’s an intent.” Dumra’s own framing of the AI version is the banking instance MGI uses: “the difference between offering employees a chatbot for ad hoc use and deploying custom agents alongside people in a reimagined process.”
• Radical targets that force redesign (Argenti’s 3×-not-20%): DBS lets Horizon-3 bets launch without a business case (written retrospectively a year later) precisely because “if I can write a business case and I know exactly what’s going to happen, I’m not really pushing the needle” — a structural refusal to let optimisation-grade targets substitute for transformation-grade ones.
• Innovation as a metric, not a delegated project (the BCG new-KPIs-that-link-productivity-to-outcomes directive + Jassy’s structural changes): 20% of every scorecard down to each team member is “transformation.” The trap’s root cause as the page names it — “lagging firms recognise AI as important but delegate it to technology groups without specific goals or metrics” — is exactly what the DBS scorecard rule is engineered to prevent. Dumra’s information-gap → action-gap thesis (“the information gap is gone; now you’re just left with the action gap”) is the trap restated as a doing problem, with the KPI cascade as the forcing function for doing.

The individual-level efficiency-tax — why employees hide their gains ( HBR June 2026)

The trap has a micro-foundation at the level of the individual worker. When efficiency gains are treated as spare capacity to be filled rather than a dividend to be reinvested — “if I automate A and B, they’re not just gonna let me focus on C; they’re gonna make me do D, E, F” — employees rationally stop disclosing their best AI workflows (ai-knowledge-hiding). The firm-level symptom (task gains that never reach the P&L) is partly produced by this hiding: privately-held workflows can’t aggregate into shared capability. Anicich & Brouwers’ prescription — “stop taxing efficiency gains” with an explicit norm for how saved time is reinvested (deeper analysis, higher-value work, development, recovery) — is the labor-side complement to BCG’s redesign-not-cost directive and the page’s reimagine-don’t-optimize thesis: the trap closes not only by redesigning workflows but by changing the implicit deal so that surfacing a productivity gain doesn’t simply earn the worker more work.

The task-level-vs-firm-level productivity spread in one chart ( Rabobank, June 2026)

RaboResearch (June 2026) collates ~18 AI productivity studies into a single ranked figure that lays the trap bare in one view: task-level / occupational experiments cluster high; firm-level studies cluster near zero. The top of the chart is task-experiment territory — advertising +73% (Ju & Aral 2025), software development +55.8% (Peng et al. 2023), consulting +25.1% (Dell’Acqua et al. 2025), customer service +14% (Brynjolfsson, Li & Raymond 2024). The bottom is firm-level territory — +1.4% (Bloom, Barrero, Davis et al. 2026), 0% TFP (Babina, Fedyk, He & Hodson 2023), ~0% (Otis et al. 2023 entrepreneurship), and most pointedly Atlassian Research 2025: 96% of firms report no ROI — with one study (Niederhoffer, Teevan & Jaffe 2025) showing knowledge work slower (−2 units/incident).

This is the wiki’s first source to visualise the gap as a single distribution rather than assert it: the same span the page builds from individual case studies (Bain’s 10–25% EBITDA leakage, the capability-reliability gap) is here a chart whose left tail (task experiments) and right tail (firm aggregates) are the two ends of the trap. RaboResearch states the bridge explicitly: the experimental gains are “onzeker of deze effecten standhouden buiten experimenten” — implementation, integration and skills determine whether task-level productivity reaches the P&L. The figure does not name the escape (that is the page’s redesign thesis), but it is the cleanest single-frame diagnostic of the trap the wiki holds. (One caveat to carry: the METR row is plotted +18%, where METR’s own July-2025 RCT found experienced developers ~19% slower — a sign-discrepancy worth not taking at face value.)

The electricity precedent, restated as popular history (Wood, WorkLab, June 2026)

Aneesh Raman’s book opens its companies chapter with a historical instance of the trap that predates every source on this page: factories that adopted electricity but installed it exactly where the steam engine had been, changing nothing else about the floor plan — “productivity didn’t really budge, and everyone started questioning the technology.” The lesson Raman draws, in the book’s own words via Conor Grennan: “this is not about just folding it in. This is about transforming your entire way of work.”

This is the same bolt-on-fails diagnostic the page already holds from AI-era sources — Allen’s “AI bolted on is going to fail,” Caldas’s “automate first, then reimagine,” the Microsoft Agentic DevOps keynote’s typing-pool analogy — but pushed back one general-purpose-technology cycle earlier, and packaged for a mass-market book-promotion audience rather than a practitioner or consulting one. It does not add new evidence or lift the page’s confidence past its cap (per Lifecycle rules); its value is corroborating, via an independent historical case, that the trap is a recurring pattern across general-purpose technologies rather than an AI-specific pathology — reinforcing the page’s W&W framing that transformation primitives outlast any one technology wave.

The software-engineering-specific restatement (McKinsey panel, June 2026)

Palaniappan, Harrysson & Linderman (McKinsey) restate the trap narrowly for software engineering: a team can complete in 4 days what took 4 weeks using coding agents, but that individual-level speedup does not automatically scale to hundreds or thousands of developers — because the surrounding pipeline (code review, QA, deployment) stays a bottleneck when firms distribute a new coding assistant without redesigning the workflow around it. Their diagnosis of why firms fall into the trap matches the page’s existing consensus almost exactly: organizations captured real gains only when they rearchitected how software gets made, not when they merely licensed a tool.

This is the wiki’s first source to state the trap specifically and narrowly for the software-development function — prior software-engineering-adjacent corroborations (Forsgren & Macvean’s DORA individual-vs-team-benefit divergence, Randell & Gousset’s “feeling faster while going slower”) name the same failure mode but at the individual-developer or DORA-metric layer; this source names it at the engineering-org-scaling layer specifically. Per Lifecycle rules this is a qualitative consulting-panel claim (unquantified beyond the single 4-days/4-weeks anecdote), so it does not lift the page’s confidence past its cap — its value is narrowing the trap’s software-engineering instance to the org-scaling mechanism specifically.

The 35%-of-time-on-coding finding — a bank’s live-tooling measurement ( Danske Bank, AWS re:Invent, December 2025)

Richard Davis, CTO of Danske Bank, supplies a clean, quantified instance of the trap measured with the bank’s own live engineering tooling rather than a survey or consulting estimate. AI coding assistants cut Danske’s change lead time by roughly 50% — a substantial task-level (coding) gain. But when Davis’s team traced all the touch points across the delivery lifecycle, they found developers were spending only 35% of their time actually coding; the rest sat in Kanban-board handovers, documentation, and testing — process stages the coding-assistant rollout never touched. Developers themselves reported the mismatch directly: “progress they were seeing still didn’t equal the level of effort they’re putting in.” Danske’s response was to redesign the surrounding product development lifecycle (PDLC) itself, aiming for “one agentic workflow with zero friction,” rather than layering further coding-specific tooling onto an unchanged process — the trap’s standard prescribed escape, restated here with an unusually direct process-tracing data point (see enterprise-ai-adoption for the fuller case).

The human-understanding test — a normative escape criterion ( MIT Sloan CIO Symposium, May 2026)

Allan Tate’s MIT Sloan CIO Symposium keynote names a compact normative test that sits alongside the page’s workflow-redesign escape prescriptions: “if you increase production but you decrease human understanding… that’s a failure, because you’re surrendering to the machine. But if you increase production and at the same time you increase the human understanding of your team… that’s what we think of as success.” Where Caldas name the trap’s behavioral signature (bolting tools onto existing steps vs. rebuilding for outcomes), Tate names its cognitive signature: throughput gained at the cost of team understanding is the trap even when the workflow has been redesigned. A related claim from the same source — “AI expands the scope of work before reducing it… you have more initiatives not fewer” — is a caution against reading early-stage AI deployment as automatically trap-escaping just because activity increases; the page’s existing consensus is that activity (throughput) is precisely the metric the trap inflates while firm-level value stays flat.

”We haven’t seen AI transform companies the way we expected” — a CIO-community self-critique ( MIT Sloan CIO Symposium podcast, July 2026)

Irving Wladawsky-Berger states the trap’s individual-vs-enterprise gap as a personal frustration rather than a third-party finding: “the part where AI still has not achieved our expectations is in bringing value to the companies themselves — improving their overall productivity, improving their profits, letting them develop new products and new services. We haven’t seen AI transform companies the way we have expected.” His own examples of the individual-level gains that haven’t compounded — AI-edited blog posts, faster programming — are the same category of “very nice, individually” gains the page’s existing sources (Danske Bank, RaboResearch, McKinsey’s software-engineering panel) already document as failing to compound without workflow redesign. His prescription matches the trap’s standard escape: “start building AI-based pilots that actually work… and learn what it takes to bring them into operations and general deployment. The only way to do this is by doing it.” A distinct, CIO-community-specific addition: he judges agentic AI as lagging generative AI in organizational maturity specifically because agentic deployment raises unresolved accountability questions (“who builds the agent? who is responsible if the agent makes a mistake?”) that conversational LLM use does not — a governance gap, not just a workflow-redesign gap, standing between individual AI use and firm-level value capture.

The verification tax — a macro-level mechanism for why task gains don’t aggregate ( Bloomberg Trumponomics, August 2026)

Every source on this page so far locates the gap between task-level gains and firm-level value in organisational causes — unredesigned workflows, unchanged metrics, listening failures, hidden tribal knowledge. Carl Benedikt Frey adds a mechanism that survives even a perfectly redesigned organisation, and states it as arithmetic:

“With AI you still need verification. AI automates a lot of knowledge work, but at the end of that you need to verify the output. And so it’s the time saving minus the time for verification.”

His contrast with the prior technology wave is the load-bearing part. The computer and the internet automated downtime — “I could sit around and wait for hours, even weeks, for new material to arrive for me to start my research, and the internet gave me access to that instantaneously.” AI does not touch downtime; it automates production and leaves a verification step behind. This is the macro-economic complement to Kiron & Schrage’s organisational-learning finding that most firms treat AI outputs “as verdicts to be confirmed rather than starting points to be interrogated” — where Kiron-Schrage treat verification-without-evaluation as a failure to compound, Frey treats verification itself as an irreducible tax on the time saved. Both can be true: the tax sets the ceiling, the flywheel determines how close a firm gets to it.

Frey supplies a historical control the page has lacked. The computer revolution was, he argues, “more transformative than the AI we have today” — it connected the world’s scientists, streamlined research, and put the store of human knowledge in our pockets — and what it produced was “basically a decade-long productivity upsurge mostly confined to the United States.” Set against the page’s existing electricity precedent, this is the second general-purpose-technology comparison on the page, and the more pessimistic one: electricity eventually paid off once factories were rearranged, whereas Frey’s reading of the computer era is that the payoff was narrow in both time and geography even after the rearranging.

Two further contributions. First, an adoption-quality correction that cuts against every adoption-rate metric the page tracks: *“adoption for what? If people adopt it for email, it’s not going to drive growth in any meaningful way — but it may look good in the sense that the adoption rate is high.” Second, a mechanism for why AI’s gains dissipate at the level of the individual knowledge worker — drill deeper versus drill more holes. Given a new productivity tool, researchers took on more projects rather than going deeper on the same ones; “your attention is going to be more thinly spread across multiple projects, and as a result of that you’re actually less likely to make a breakthrough at any given time. We see that in the data. AI seems to have had the same effect.” The driver he names is incentives (publish-or-perish), which makes this the individual-incentive twin of the page’s existing efficiency-tax finding: in both cases a rational individual response to how gains are rewarded prevents them from aggregating. Per Lifecycle rules this does not lift the page past its 0.95 cap — it is a book-promotion interview whose empirical claims are asserted without citation — but it is the page’s first genuinely macro-economic vantage, and the verification-tax formulation is its most compact.

Where the chain breaks: the radiology bottleneck and a ~30% pilot-to-production ratio (BBC AI Decoded, August 2026)

The BBC’s AI Decoded panel states the trap in the plainest language the page holds, in the programme’s cold open:

“Give the AI a clearly defined task. Write code, answer customer queries. The systems are measurably faster. But most businesses are a chain. And all too often speeding up one link just piles up the work at the next.”

Dr Stephanie Hare’s radiology worked example is the page’s clearest single illustration of the mechanism, and unusually legible because the bottleneck is regulatory rather than cultural: using AI to read X-rays, MRIs and CT scans faster “doesn’t mean that you’re necessarily going to cut patient waiting lists or get faster treatment for people, because the scan is just one of many, many tasks, all of which have to be signed off. Healthcare is highly regulated by doctors the whole way through. Well, those doctors aren’t getting more time just because the scans are being read faster. So those read scans pile up, but everybody still has to then put them through the same old processes.” Her conclusion is the page’s standard prescription reached independently: “we’ll have to rethink a lot of workflows from the ground up.”

Three quantitative anchors from the episode. PwC’s Davos survey of ~4,500 CEOs across 95 countries: 56% have yet to see a return on their generative-AI investment, and only one in eight can claim an actual cost saving or revenue generation — a larger and more recent sampling frame than the MIT 95% figure the page carries via AWS, and one that Peter Grant explicitly prefers on those grounds (the MIT study sampled 153 companies). A ~30% pilot-to-production ratio in Grant’s own client index — of ten pilots running, roughly three reach production workflows — which is the quantified form of Allen’s “pilot purgatory.” And a 3:1 ratio of AI tool access to AI training access in British businesses, which is the training-side correlate of the page’s organisational-change-dominates finding.

James Kuht names the training failure precisely, in a way the page’s existing sources gesture at but do not isolate — the imagination gap. Training that does happen “focuses very much on how do you use the tool. Well, you write in the box and you click enter… It’s pretty easy to actually use it.” What is hard is “helping people cross this imagination gap between I know how to write in the box and I know how to use this tool to reinvent how I actually work — and use my time to deliver value, not just do things faster and push a bottleneck onto someone else.” That last clause is the trap stated from inside the individual role. Grant’s complementary diagnosis at the change-management layer: firms “have just given people the keys to the car without teaching them to drive.”

Caveat, load-bearing here: two of the three guests sell the remedy being discussed (Grant an AI maturity-assessment product, Kuht AI training), and the ~30% index is their own unaudited data. The framings are corroborated across the page from non-vendor vantages; the specific number is not.

”You’re still just automating the wrong work” — the shared-services restatement ( McKinsey Talks Operations, August 2026)

McKinsey’s Talks Operations episode on Global Business Services contributes the page’s sharpest statement of the sequencing error. Josh Peters, on a client whose AI strategy was largely a list of “10, 20 use cases in each area that we think might have the most potential”:

“If the first step that you’re taking is not — do we have the right materiality threshold in place for how many of these things we go after, and is there any low-value work that we can just eliminate — you’re still just automating the wrong work.”

This adds a step that sits before the page’s standard workflow-redesign prescription. Westerman says rebuild the process for the outcome rather than bolting tools onto existing steps; Argenti says set a target large enough that redesign is the only way to hit it. Peters says: first ask whether the work should exist at all. Eliminating low-value work is not a use case, so it never appears on a use-case list — which is precisely why use-case-list-driven programmes systematically skip it.

The episode’s second contribution is the automate-first versus right-shore-first decision, which gives the trap a concrete sequencing choice rather than a general injunction. Right-shore first migrates work into the shared-services organisation for standardisation and quick value capture, then applies AI to the now-standardised work. Automate first means “completely reimagining the process flow before migrating some of that work.” The panel refuses a general answer — “it’s not one-size-fits-all, it’s actually a little bit more horses for courses. Some parts of a process may benefit from moving to GBS first. Other parts of a process may benefit from automating first” — decided against three factors: existing capability and AI experience, the impact expected and when it is needed, and the process type. The observed split is itself informative for this page: green-field situations with low process maturity and urgency still choose right-shore first, while automate-first appears where the target is enterprise-value unlock (revenue leakage, price realisation, EBITDA basis points) rather than headcount or process cost. That maps onto the page’s core distinction — cost-lever targets are reachable by optimisation, value-unlock targets are not.

Third, the three E’s — efficiency, effectiveness, and customer experience — as the metric set that replaces “a mere efficiency debate,” which is the same measurement-realignment move BCG prescribes as new domain-specific KPIs. As with the BBC episode, the numbers here (GBS organisations ending up “20 or 30% smaller but dramatically more productive,” “roughly half have started to pilot”) are unattributed practitioner estimates from a firm selling GBS transformation; the frameworks are the durable content.

”AI is not a tool” — the trap as a budgeting-and-measurement error ( WorkLab, August 2026)

Allie K. Miller opens the WorkLab episode with a causal chain that reads as a compact aetiology of this page — the trap as a consequence of a category error made at procurement time:

“If you think of AI as a tool, you are going to budget as if it is SaaS or traditional procurement, roll it out as if it is a traditional SaaS platform, measure its impact solely based on productivity — and you’re not going to think of brand new business lines, org restructuring, process reinvention. None of that really happens if you’re just rolling out a productivity tool.”

The sequencing of that chain is its contribution. The page’s existing sources mostly locate the trap at redesign time; Miller locates its origin earlier, in how the spend is classified, and derives the redesign failure from it. Her metric prescription follows directly and converges with BCG’s new-KPIs directive: measure engineering velocity, net-new ideas tested, customer trust and employee fulfilment, but “growth is the one that I think it should be measured on, not just productivity.”

Two further contributions. First, a depth-of-use critique of adoption metrics that lands in the same place as Frey’s adoption for what: “I hope all of you know your monthly active, weekly active, daily active AI users… But what I would love people to look at is what are they actually using AI for? Because if my whole company is using AI to rewrite emails… that is not meaningful AI adoption. You might have people who believe that they’re amazing AI users… and they’re using it at 5% of the power.” Her “Gremlin mode” ladder makes the gap concrete: a CMO who judged AI writing inadequate (“it doesn’t have heart”) after a single un-iterated prompt, against an eight-loop, six-hour brief-generation-and-refinement workflow. This is the individual-practice correlate of the BBC panel’s imagination gap.

Second, the page’s clearest employee-side statistic, from Microsoft’s own Work Trend Index: 65% of AI users say they are afraid of falling behind if they don’t move quickly, but 45% say it feels safer to focus on current goals than to redesign their work with AI. Those two numbers describe a workforce that knows redesign is necessary and rationally declines to attempt it — which is the demand-side twin of the efficiency-tax finding already on this page. Miller’s countermeasure list starts by ruling out the intuitive lever (“leading by fear… never works. You’ll have mutiny” — she cites a firm whose productivity fell for two months after a three-month ultimatum) and ends at speed of iteration as “the moat of 2026 and beyond.” Note the source posture: this is vendor-published content recorded at the vendor’s own customer conference, and most of Miller’s numbers are unsourced as stated; the structural contributions (the frontier unit, the Gremlin ladder, the budgeting-error aetiology) are what this page takes.

70% people, 20% process, 10% technology — the trap quantified by an operator ( McKinsey, July 2026)

This page has carried the organisational-change-dominates thesis since Sternfels’s “half, if not more, of the secret sauce is organizational change as opposed to technology implementation.” Laura Hines-Pierce, co-CEO of a 70-year-old real-estate firm with ~$100bn AUM, puts a sharper number on it from the operator’s chair:

“We’re viewing this as like a 70% people challenge. This is not a technology [challenge] — technology is like 10% of the solution. The other 20% is processes. But 70% of this is about people and change management.”

Read it as a statement of where a leadership team directs attention rather than as a measurement — but note it is the third independent firm to land in the same range (Miller: “tech is… less than 50% of your strategy”; Sternfels: “half, if not more”), and the first from a target firm rather than an advisor.

Her rollout structure is the operational half: top-down direction plus bottom-up experimentation within business units, plus a deliberately created middle governance team whose job is narrowly defined as making a solved workflow get used consistently across the teams that need it — the exact failure this page tracks as pilots that never propagate. Champions sit “side by side with the business teams… to deeply understand what processes will actually help this team.” And her augmentation goal is stated as competitive logic rather than values: “our focus is not about cost cutting. It’s not about removing jobs… we want to use and leverage AI tools to get them out from behind their computer and out in the real world where they can create value.”

The J-curve, and “honestly, I think it’s a little lazy” (Brynjolfsson, August 2026)

Erik Brynjolfsson gives the trap its economic-theory framing and its bluntest moral one. The theory is the productivity J-curve: these technologies require complementary intangible investment, which depresses measured productivity before it lifts it — so a firm sitting in the trough is not necessarily failing, and a firm reporting no gains is not evidence the technology doesn’t work.

The moral framing lands on the measurement choice this page treats as the trap’s proximate cause. His CFO anecdote: told they needed to measure AI ROI better and would therefore audit each department for headcount reduction, he replied “well, okay, that’s one measure. It’s a very narrow measure, but it’s so pervasive, this sort of cost mentality, because let’s face it, it’s the easiest thing to measure.” On the efficiency-first instinct generally: “honestly, I think it’s a little lazy. It’s the easy way — cut costs, it’s not so hard to measure, but I don’t think it’s as sustainable as if you find new sources of value and you invest in your workforce. And you’re going to get more resistance from your workforce if you tell them this is a tool for cutting heads.” This is the same diagnosis BCG gives (“cost actions are visible and straightforward… when AI drives productivity rather than cuts, ROI becomes harder to define and defend”), stated by an economist rather than a consultancy and with an added methodological demand the page has not carried before: measure the output side, and establish causality — “difference-in-differences and instrumental variables… there’s been a credibility revolution, and one of my missions is to bring that to business.”

Two further contributions. Complementarity as strategy: “so many technologists are obsessed with making AI that imitates humans… that’s maybe kind of a cool concept, but it’s a terrible business strategy. Instead… have the AI focus on what AI can do well and not what humans are already doing well.” And the risk asymmetry that justifies the harder path: “to imagine something new, that’s harder. But my view is actually it’s more risky to not do it… ultimately no company, no country, no person has ever succeeded by just focusing on the same thing over time."

"No company ever gained competitive advantage just by buying a licence” (Ng, July 2026 + WEF-PwC, July 2026)

Two July 2026 sources restate the bolt-on diagnosis in unusually compact form. Andrew Ng: “businesses do not feel like it’s helping them yet. And one of the challenges is no company ever gained competitive advantage just by buying a ChatGPT or Microsoft Copilot licence. The problem is finding the right use cases, and that’s a people, change-management [problem].” His prescription is the kind of use case: “not just let’s chat with a chatbot and copy-paste, fix my email grammar, but… what are the automation scenarios that let you change the business model and not just get cost savings, but more excitingly get growth” — with Morgan Stanley’s “challenge labs” (business and technology leaders paired to find use cases, after firm-wide training and a CEO-set tone) as his named mechanism.

Peter Brown supplies the numbers side, from the WEF–PwC report: firms “getting ahead are seeing three times growth in productivity — and that’s where the work has fundamentally been redesigned.” The failure mode: “one thing they’re not doing is just layering AI on, bolting AI onto existing ways of working. Some organisations have done that in the chase of efficiencies and you might get some incremental efficiencies. You don’t see the long-term value.” That is the page’s standard finding with a 3× figure attached — though second-hand, from a report co-authored by the speaker’s own firm, and with “AI-exposed sector” undefined in the audio.

The J-curve narrated from inside a single job ( How I AI, August 2026)

Brynjolfsson gives the J-curve its economic-theory form: complementary intangible investment depresses measured productivity before it lifts it. Daniel Blum supplies what the corpus otherwise lacks — a first-person account of the trough, at N=1 and with the mechanism named.

His warning is unusually specific about the psychology of the trough, which is where firms abandon: before the system knows your “business logic,” “everything it does or says or thinks is going to be a little bit [off] and you’re going to lose trust and you’re going to be very skeptical and then you’re going to double-check everything.” Double-checking is the trap’s self-sealing property — it is rational in the trough, it consumes the very time the tool was meant to return, and it is indistinguishable from the tool simply not working. His two instructions for spending through it are the intangible investment Brynjolfsson’s theory predicts, stated operationally: contextualise ruthlessly and centralise whatever work you can in one harness, “even if it doesn’t feel so natural.”

The mechanism he claims for the upswing is a reinforcing loop rather than a step change: centralisation causes context to accumulate, accumulated context reduces friction, reduced friction makes further centralisation attractive. That is the individual-scale analogue of the workflow-redesign argument this page makes at firm scale — and it clarifies why offering lock-in and process lock-in bite so hard. A firm that layers AI onto existing processes never triggers the loop, because work keeps leaving the system and context never compounds.

Two cautions on weight. This is self-reported, single-subject, unaudited, and delivered by a speaker with a podcast audience; the “a week in a day” multiplier should be read as anecdote. And it sits inside rather than against the return-gap puzzle that Frey and BBC’s AI Decoded document — vivid individual gains coexisting with flat aggregate productivity is the pattern needing explanation, and one enthusiast’s account is a data point in it, not a resolution of it. What transfers is the mechanism and the trough, not the multiplier.

Debates and supersession

• A worker-level escape narrative, added 2026-08-31. Blum describes escaping the trap inside one job via centralisation → context accumulation → friction reduction. It does not contradict the page; it supplies the individual-scale mechanism and, more usefully, the first account of the trough’s felt shape. Open question the corpus still cannot answer: how deep and how long is the trough, measured across workers rather than asserted by one? A firm deciding whether to push through needs that number and nothing in the wiki supplies it.
• No genuine contradiction across sources; unusually strong convergence instead. Every source cited on this page — spanning Bain/OpenAI, McKinsey, Thoughtworks, NYT journalism, DORA, a target-firm CEO, the MIT Sloan CIO Symposium podcast, and (as of 2026-08-12) a broadcast panel, a shared-services consulting episode, an advisory practitioner and an economic historian — agrees on the core diagnosis (task-level gains ≠ firm-level value without workflow/organizational redesign). The debates that exist are about vocabulary and mechanism layer, not substance: control-systems framing (Hu) vs. organizational-metabolism (Thompson) vs. neurological gear-3 (Storoni) vs. governance-gap (the MIT Sloan CIO Symposium podcast’s agentic-AI accountability point) are compatible lenses on the same phenomenon, not competing claims.
• Vendor-interest caveat, unresolved — and now broader. The headline 10–25% EBITDA figures are Bain client-engagement data, not independent measurement (see Open questions below) — this caveat has stood since the page’s first ingest and no independent academic/AI-Index verification has yet closed it. The 2026-08-12 batch widens rather than narrows the problem: Grant’s ~30% pilot-to-production index is his own product’s data, the McKinsey GBS forecasts are unattributed practitioner estimates from a firm selling GBS transformation, and most of Miller’s figures are unsourced. The page’s diagnosis is corroborated far more strongly than any of its numbers.
• A newly surfaced tension about the ceiling (2026-08-12). Frey’s verification tax implies a structural limit on how much of AI’s task-level speed can ever reach firm-level value — the saving is net of verification time, always. The page’s dominant framing is that the gap is closable by redesign. These are compatible in principle (the tax sets the ceiling, redesign determines the distance to it) but no source on the page attempts to quantify the tax, so it is unknown whether it explains a small or a large share of the observed gap. That is now the page’s most consequential open empirical question.
• Open attribution question, newly surfaced (2026-07-15). The MIT Sloan CIO Symposium podcast’s discussion of entry-level hiring softness — is it AI displacement, a post-COVID hiring-bubble correction, or a deliberate shift toward retaining experienced staff for infrastructure work — is explicitly unresolved by its own sources (Wladawsky-Berger: “I don’t know if it’s related to AI or the economy”). This sits adjacent to, but does not resolve, the parallel attribution debate already tracked on ai-employment-effects.

Related concepts

• ai-knowledge-hiding — the individual-level response to the efficiency-tax: workers conceal workflows when disclosure earns more work or threatens their standing, so task gains never aggregate to firm value.
• automation-vs-augmentation — process lock-in maps to automation-without-redesign; reinvent the business to outcome-oriented augmentation.
• enterprise-ai-adoption — broader concept; this is one diagnostic lens within it.
• dynamic-capabilities — escape from the trap requires the digital sensing/seizing/transforming microfoundations.
• systems-thinking — the workflow-redesign step is systems-thinking applied to AI.

The measured version, in software (added 2026-08-30)

The trap’s usual evidence is enterprise-transformation survey data. The 30 August 2026 ingest supplies a randomised instance of the same shape, in the one function where AI adoption is furthest along. Full treatment at ai-coding-productivity-evidence.

METR’s RCT found 16 experienced developers 19% slower with AI on 246 tasks in their own mature repositories — while forecasting 24% faster beforehand and, after finishing the work, still estimating 20% faster. A ~39-point gap between the felt and the measured, held by the people who did the work. That is the micro-productivity trap’s perception mechanism, isolated experimentally: local assistance feels like a gain and is measured as a loss.

The system-level counterpart is DORA 2025: AI adoption’s relationship with delivery throughput turned positive while its relationship with stability stayed negative — “without robust control systems… an increase in change volume leads to instability.” More local output, no better delivery. And the downstream ledger: 46.41% of agent-authored fixes rejected (2026-06-11-abujadallah-rejection-of-agentic-pull-request-fixes), 22.7% of AI-introduced issues never fixed (2026-03-30-liu-debt-behind-the-ai-boom).

The practitioner statement of the same point comes from the person with the most to lose by making it — Ryan Carson, at $20k/month of agent spend: “I don’t think I get multiples of quality off of multiples of output.”

Open questions

• Empirical 10–25% EBITDA range is from Bain client work — vendor-of-deployment data, not independent measurement. Independent verification (academic field study, AI Index data) would be a useful counterweight.
• Whether entry-level hiring softness (discussed on the MIT Sloan CIO Symposium podcast) is itself trap-adjacent — i.e. whether firms retaining experienced staff for infrastructure work while under-hiring juniors is a rational trap-escape allocation or a new failure mode — is untested by any source on this page.