Two recent studies examine how AI investment reshapes local borrowing costs, but with different conclusions. In the US, Andreadis et al. (2025) exploit county-level AI job exposure and find that increases in AI job postings reduce municipal bond yields, particularly for longer-maturity and lower-rated bonds. Yield declines are explained by improvements in local productivity, skilled in-migration and rising property values – all of which expand the tax base and strengthen governments’ ability to service debt. Yet for five OECD countries – Belgium, Canada, Germany, Spain, and Sweden – the sign reverses. Using comparable Lightcast vacancy data to measure local AI job intensity and linking it to subnational bond outcomes, we find that expansions in AI job intensity are associated with higher, not lower, yields (Dougherty and Makridis 2026). 

Figure 1 Bivariate relationship between bond yields and the AI share of jobs

Source: London Stock Exchange Group (LSEG) and Lightcast.
Notes: The figures show the correlation between the share of job postings that are AI related and bond yields across municipalities. The left panel uses observations for U.S. municipalities; the correlation is -0.15. The right panel uses observations for 152 municipalities in Belgium (9), Canada (43), Germany (32), Spain (32) and Sweden (36); the correlation is 0.50. In both panels, the samples are winsorised at the bottom and top percentile.

The divergence is not a question of which result is ‘correct.’ It suggests that bond markets are pricing the transition to an AI economy – with risks, timing, and institutional context – rather than its long-term potential. Understanding the distinction requires looking at the institutional architecture that sits between AI diffusion and fiscal capacity, and at what labour-market data reveal about how the transition may unfold.

Why the same technology produces opposite market signals

The US municipal bond market is unusually deep, tax-advantaged, and issuer-differentiated. Investors can, and do, price local fundamentals at fine geographic resolution: when AI investment raises the productivity and property-value trajectory of a county, that signal reaches borrowing costs relatively quickly. The OECD settings we study feature thinner subnational bond markets, greater reliance on bank lending or pooled funding agencies, and equalisation formulas that partially offset incremental local revenue gains. In a prior VoxEU column, Dougherty et al. (2025) also showed that subnational bond markets in many OECD countries exhibit low dispersion in borrowing costs across jurisdictions despite clear differences in fiscal performance, a feature that limits the ability of local AI gains to register in yields. In these settings, the near-term uncertainty and transition costs associated with AI diffusion may loom larger in investor pricing than the long-run productivity potential.

Several institutional features may help explain these patterns. First, fiscal retention: subnational governments that do not retain the marginal revenue gains from local AI-driven growth cannot improve their balance sheets commensurately, even as transition costs – retraining, social services, digital governance – remain local. Second, labour-market rigidities: where adjustment is slower, firms may shift toward automation and outsourcing more abruptly, concentrating transition risk rather than spreading it. Third, market depth and price discovery: in shallow markets, pricing places greater weight on shared risk and uncertainty than on incremental local gains. The US result likely reflects these institutional features working in the municipality’s favour; the OECD result reflects the opposite configuration.

What the labour market data are showing

The bond-market divergence becomes more interpretable when viewed through labour-market data, which shape the short-run fiscal base that bond markets ultimately price. AI has so far complemented labour more often than it has substituted for it, consistent with an augmentation-dominant phase of adoption (Acemoglu and Restrepo 2019). Yet the distribution of gains and adjustment costs is uneven, and local institutions determine which predominate.

Johnston and Makridis (2025) track industries and states across the US from 2017 to 2024 using administrative data for nearly all employers. Industries one standard deviation more exposed to AI recorded productivity gains of about 10%, employment gains of 3.9%, and wage gains of 4.8% relative to comparable industries. The pattern is consistent with demand expansion: lower task costs raise output enough to increase labour demand. Gains were concentrated in states with more efficient labour markets, pointing to the importance of local institutions. 

The evidence is admittedly mixed. Firm-level data from Europe show productivity gains without negative employment effects (Aldasoro et al. 2026), pointing to capital deepening and augmentation as the dominant mechanisms in the early stages of adoption. The absence of displacement, however, does not imply a smooth or evenly distributed transition across workers, firms, or regions.

Early labour-market signals appear at the entry margin. Brynjolfsson et al. (2025) document relative employment declines for early-career workers in AI-exposed occupations, termed a ‘Canary in a Coal Mine’ effect. Related work finds that AI adoption is ‘seniority-biased’, disproportionately affecting junior roles. By contrast, Danish administrative data examined by Humlum and Vestergaard (2025) show precisely estimated near-zero average effects on earnings and hours in the short run, consistent with heterogeneous and gradual adoption. 

Part of this heterogeneity reflects differences within firms, not just across countries, as technology diffuses through organisations and depends on the managers who sit between the tool and the worker. Organisational factors also matter. Evidence suggests that AI adoption and its productivity effects depend on management quality and communication, reinforcing that institutional context shapes how quickly gains materialise (Makridis 2026). These patterns are consistent with a transition in which productivity gains are real but delayed, while labour-market adjustment costs appear earlier: precisely the configuration that would lead markets to price higher near-term risk.

A transition-risk interpretation of the bond evidence

Local revenues depend on wages, employment, property values and intergovernmental transfers (OECD 2022). If early AI diffusion is associated with slowing entry-level hiring, occupational reallocation, or outsourcing of tasks that previously generated local wage income, the near-term tax base can weaken or become more volatile just as spending needs for retraining and public services rise. This timing mismatch – transition costs front-loaded, productivity dividends deferred – is precisely the configuration in which rational investors would demand higher risk premia during the transition period, even when long-run fundamentals are strong.

The US result is consistent with this logic. In the American context, the productivity dividends arrive more quickly and are more fully retained by the local jurisdiction, so markets price the long-run gain before transition costs fully accumulate in fiscal accounts. The ex-US OECD result, where gains are diluted through equalisation and markets are thinner, reflects a setting in which transition-period uncertainty dominates, before productivity improvements reach fiscal balance sheets. The same technology can have very different results due to institutional transmission. 

Implications

The evidence points to three areas where policy can shorten the transition and narrow the institutional gap between the US and OECD patterns.

Procure AI for augmentation, not replacement, and focus on organisational integration. Gallup data show that by the end of 2025, 43% of US public-sector employees report using AI at least a few times a year, up from 17% in mid-2023, yet only 21% use it frequently, compared to 25% in the private sector.  Governments have achieved breadth without depth. Public agencies should therefore treat AI deployment as a staged organisational reform with workforce complements – retraining, redesigned workflows, human-in-the-loop governance – rather than a cost-reduction exercise. Where AI augments workers rather than substitutes for them, productivity and wage gains follow. That managerial commitment is as important as the technology itself in determining whether adoption moves from occasional to embedded.

Protect entry-level on-ramps. If the strongest near-term adjustment falls on younger workers and entry-level tasks, as the ‘canary in the coal mine’ and seniority-bias evidence suggest, the policy response should preserve and expand pathways into AI-complementary roles: apprenticeships, public-private training partnerships, and AI-enabled credentialing. Intervention at this stage is less costly than remediation after displacement consolidates.

Improve fiscal pass-through and market transparency. Where equalisation formulas fully offset local AI-driven revenue gains, transition costs land on local balance sheets while productivity dividends are redistributed. Targeted adjustments, such as time-limited retention windows tied to verifiable AI-driven upgrading, can preserve long-run equity objectives while ensuring that local investment incentives survive the transition. Standardised disclosure of guarantees, issuer boundaries, and contingent liabilities would enable better investor differentiation and reduce the premia generated by thin markets, allowing yield signals to track fundamentals rather than aggregate uncertainty.

Authors’ note: The opinions expressed and arguments employed above are solely those of the authors and do not necessarily reflect the official views of the OECD or its member countries, nor of the other affiliations of the authors.

References

Acemoglu, D and P Restrepo (2019), “Automation and New Tasks: How Technology Displaces and Reinstates Labor”, Journal of Economic Perspectives 33(2): 3–30.

Aldasoro, I, L Gambacorta, R Pal, D Revoltella, C Weiss and M Wolski (2026), “How AI is affecting productivity and jobs in Europe”, VoxEU.org, 17 February.

Andreadis, L, M Chatzikonstantinou, E Kalotychou, C Louca and C A Makridis (2025), “The local effects of artificial intelligence labor investments: Evidence from the municipal bond market”, VoxEU.org, 18 April. 

Brynjolfsson, E, B Chandar and R Chen (2025), “Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence”, Stanford Digital Economy Lab Working Papers, November.

Dougherty, S and C A Makridis (2026), “Artificial intelligence and local debt: Evidence from five OECD bond markets”, OECD Working Papers on Fiscal Federalism, No. 51.

Dougherty, S, A Brochado and P de Biase (2025), “Subnational debt in turbulent times: Navigating fiscal risks and market inefficiencies”, VoxEU.org, 20 July. 

Humlum, A and E Vestergaard (2025), “Large language models, small labor market effects”, University of Chicago, Becker Friedman Institute for Economics Working Papers, No. 2025-56. 

Johnston, A C and C Makridis (2025), “AI, Output, and Employment”, SSRN Working paper.

Makridis, C A (2026), “Organizational transmission of AI: Managerial influence on generative AI adoption”, CESifo Working Papers No. 12373. 

OECD (2022), Fiscal Federalism 2022, OECD Publishing, Paris.