Does AI Adoption Improve Productivity? Effects Over the First Three Years

Bank of Korea Issue Note No. 2026-12 (June 8, 2026). The first vault source to put a macro econometric number on the gap between worker-level AI adoption and firm-level productivity. The dataset is a Korean household survey of 5,512 employed persons aged 15–64 conducted May–June 2025 by the BOK Research Department's Labor Market Research Team.

The paper names a phenomenon — the AI Productivity Disconnect — and explains it as the textbook lag phase of a general-purpose technology (Solow Paradox / J-curve).

Headline numbers

51.8% of Korean workers (as of 2025) use generative AI for work. Diffusion is ~8× faster than the internet's at the same age.
AI adoption reduces work time by 3.8% (≈1.5 hours per 40-hour week).
If those time savings were fully reinvested into output, the potential productivity gain would be ~1.0% (GDP grew 3.9% from Q4 2022 ChatGPT launch to Q2 2025; the estimate makes AI a 1.0pp upper-bound contribution).
Correlation between worker-level time savings and worker-level output increase: zero. Holds in regressions with individual controls + fixed effects. Confirmed at the industry level (Box 1: AI adoption rate ↔ labor productivity correlation ≈ −0.10 for the 2022 Q4 – 2025 Q4 window).
Enterprise AI adoption rate: 9.6% (2024 Survey of Business Activities). Worker AI adoption rate: 51.8%. The 5× gap is the structural picture: workers adopted; firms didn't redesign.

The named phenomenon: AI Productivity Disconnect

"While AI is currently improving efficiency at the individual task level, it has not yet fully transitioned to the 'productivity' stage that involves increases in total output."

The disconnect is heterogeneous in a way that reveals the binding constraints:

Heavy right tail. Most workers see ~0% time change; a small group sees very large savings. Time savings are concentrated in cognitive non-routine tasks — student counselling (32.2% time saved on that task), curriculum development (24.6%), assignment & exam evaluation (23.3%), student assessment (17.2%) — and absent in coordination / physical work.
Time savings ≠ output. Across the full sample the correlation is 0. But subgroup interaction terms show output gains do materialize for:
- Self-employed workers (+1.0pp output vs. wage workers) — performance directly ties to income.
- Professionals (+0.7pp vs. office workers) — high autonomy and output visibility.
- Top 50% AI usage hours (+0.5pp vs. bottom 50%) — past the friction phase.
- Youth 15–39 (+0.6pp vs. 50–64) — higher digital adaptability.
No effect from gender, education, or job tenure — AI's productivity translation depends on employment status, job characteristics, intensity, and incentive structure, not demographics.

Why time savings stall before output

The paper's four mechanisms:

AI diffusion remains at the task level. Only 4.4% of tasks show >20% time savings. AI substitutes for specific tasks inside an unchanged workflow — costs and verification uncertainty still bite at the multi-task level (Bai et al. 2026).
Workflow rigidity. Pereira et al. (2026), 51 corporate cases: the hardest AI-adoption challenges are invisible — organizational culture, employee behaviour, data quality, workflow adjustment. Many successful cases reached value only after prior failures and iteration.
Production bottlenecks (Agrawal et al. 2026). If a downstream stage — decision-making, collaboration, approval — is unchanged, saved time shows up as idle/waiting time, not output.
Incentive misalignment. If extra performance isn't rewarded, workers cash time savings as leisure or stop at "good enough." Chen et al. 2025 — generative AI raised illustrator quality per unit of time but quickly lowered the marginal return to additional effort, so participants reduced work time (and some let final quality slip).

The policy framework: standardized vs open tasks

The paper's prescription is the Standardized vs Open Tasks distinction:

Standardized tasks (clear outputs, evaluable criteria: summaries, bookkeeping, regulation review) — redesign the workflow around AI. Modularize, standardize inputs, predefine verification + exception handling. Critically, establish reallocation mechanisms so saved time flows to high-value activities. The point is not adoption; it's that without reallocation, savings don't become output.
Open tasks (no predetermined form: strategy, R&D, business planning) — AI as augmentation, not substitute. Humans interpret, judge, decide. The bottleneck is human capacity to integrate AI-expanded possibilities, not the AI itself.
Redesign learning opportunities. Standardized tasks were historically how junior workers earned domain knowledge. If AI absorbs them, the long-run pipeline of people capable of open-task work collapses. Counter-design: observe → assist → lead progression on open tasks from day one; senior time saved by AI reinvested in mentoring/pair work, not skimmed off the top.
Continuous monitoring. Track AI use intensity, task restructuring, job switching as the leading indicators of the productivity transition — not adoption rates.

The framing the paper hangs everything on

"The productivity disconnect currently observed can be viewed as a typical lag phenomenon (J-curve, Solow Paradox) in the early phase of general-purpose technology adoption."

i.e. AI is in the efficiency stage and has not yet transitioned to the productivity stage. Whether it does depends on policy response and organizational transformation, not on the technology itself.

Connects to your work

This is the first vault source to give the CIO Agenda 2026 (CXOTalk) "88% use AI, <6% get value" claim a macro-empirical anchor. Crawford & Sacolick name the symptom from the CIO seat; this BOK note shows the same gap in a representative national household survey, and names the structural reasons.

A few load-bearing implications for the user's enterprise IT leadership work:

The gap is not a productivity delay; it's a workflow-design failure waiting to be done. The 9.6% enterprise / 51.8% worker adoption gap is the diagnostic. Manila IT's AI-productivity bet only pays out where someone owns redesign at the workflow tier — not pilot rollouts at the task tier.
The Tasks to Responsibilities Shift is what closes the disconnect. The paper's "AI is stuck at task level" is the mirror image of Felix Ryberg's "you should be designing loops that own a responsibility, not handing out tasks." The exception groups in the BOK data (self-employed, professionals, intensive users) are the ones who already operate at the responsibility tier, even without naming it.
Code Is Free is empirically confirmed at the task level and empirically unconfirmed at the output level. Lopopolo, Cherny, Karpathy are right that the per-task economics flipped. The BOK note shows that at the population scale, that doesn't propagate to output until the org is restructured.
The apprenticeship-pipeline concern in Designing IT Roles for an AI Era (Talent Strategy POV) is a real long-run risk, not just a precaution. The BOK paper makes this its policy section: if standardized tasks are the historical learning rung, AI absorbing them collapses the pipeline for open-task talent.
The Cobra Effect / incentive trap is documented in this data. Section II.3 finding: low-incentive workers ("good enough" stopping; Chen et al. 2025 illustrators) don't reinvest savings. This is the empirical case that bonus / reallocation design must move in lockstep with AI rollout.

Cross-references

AI Productivity Disconnect — the named concept this paper anchors.
Solow Paradox — the historical antecedent (1987: "you can see the computer age everywhere except in the productivity statistics"); the lens the paper invokes for why productivity follows technology with a multi-year lag.
Standardized vs Open Tasks — the paper's policy framework for organizational redesign.
Bank of Korea — the publishing institution.
Tasks to Responsibilities Shift — the negative-space framing; the paper's "task-level diffusion" is Ryberg's "still in Era 2".
Knowledge Work Factory Redesign — OpenAI's rhetorical "knowledge work is waiting for its factory redesign" is what the BOK note empirically demonstrates.
CIO Agenda 2026 (CXOTalk) — the practitioner-side observation now backed by macro data.
Code Is Free — task-level efficiency yes; output-level no — until org structure catches up.
Cobra Effect — the incentive trap mechanism the paper documents.
FOBO (Fear of Becoming Obsolete) — the trust/incentive modulator that determines whether workers reinvest AI-freed time.
Skill Change Index (SCI) — McKinsey's labor-market companion; same direction, different methodology.

Sources

Raw: Daily Learning 2026-06-26 22-55 #3472.md (Telegram capture) → linked PDF: raw/assets/8ed6b7d68c4f40e8878b852eccbb8dab---dac3c24b-3a7e-46f8-8cbb-03dc1c1956a4.pdf
BOK Issue Note No. 2026-12, "Does AI Adoption Improve Productivity? Effects Over the First Three Years," Donghyun Suh / Samil Oh / Jongwon Yoon, Bank of Korea Research Department, Labor Market Research Team, June 8, 2026.
The paper's own anchor on a predecessor BOK study: "Rapid Adoption of Artificial Intelligence and Its Productivity Effects: The Case of Korea" (BOK Issue Note No. 2025-22), for sample design and the underlying production-function estimate behind the 1.0% potential figure.