AI, productivity, and work: Evidence from US firms
cepr.org · 28 July 2026
A survey of senior corporate decision-makers conducted by Baslandze et al. (2026) reports that AI adoption is already widespread but shallow across US firms. Executives report positive productivity effects and expect larger gains in high-skill services and finance. Measured productivity gains lag perceived improvements. Gains appear to come mainly through innovation and demand-oriented channels rather than through cost cutting. Capital deepening is low, with most AI adopters renting intangible capital from upstream providers. Near-term employment effects are small on net, with larger reductions concentrated among larger firms and a shift away from routine clerical work towards skilled technical roles.
This is the observable fact the cited source supports, stated without interpretation.
This tests COMMERCIAL_ALIGNMENT and LAW_3. If firms are adopting autonomy without redesigning their value capture structures, the agentic profit paradox and revenue durability should come under pressure over time. This survey shows that productivity gains are real but concentrated in innovation and demand expansion, not cost-cutting, which is consistent with the hypothesis that activity and value capture are becoming misaligned. If, over the next 2 to 3 years, firms fail to shift staffing composition toward governance, orchestration and verification roles, or if they continue to rent upstream capability without moving toward internal judgement infrastructure, the paper's predictions would gain strength. [Strength capped from 4 to 3: source tier 3.]
This independent measurement of firm-level AI adoption shows a pattern consistent with HBI dynamics (reduction in direct human interaction by autonomous systems acting on behalf of users) and the early phase of value reaccumulation. Firms are not primarily achieving savings through cost displacement; instead, they are generating value through innovation and demand expansion. The concentration of employment reductions among larger firms and the shift toward skilled roles suggests that the locus of economic defensibility may be moving away from routine transaction execution and toward orchestration and coordination. Capital deepening is low because firms are not deepening their own autonomous capability; they are renting it from upstream providers, which is consistent with CSD exposure and the paper's prediction that value concentrates where scarce judgement governs autonomous throughput (the upstream providers). The observed productivity lag behind perception is important: it suggests that the economic gains are real but not yet as large as executives expect, leaving room for the paper's revenue durability and commercial alignment tests to play out.
This independent measurement of firm-level AI adoption shows a pattern consistent with HBI dynamics (reduction in direct human interaction by autonomous systems acting on behalf of users) and the early phase of value reaccumulation. Firms are not primarily achieving savings through cost displacement; instead, they are generating value through innovation and demand expansion. The concentration of employment reductions among larger firms and the shift toward skilled roles suggests that the locus of economic defensibility may be moving away from routine transaction execution and toward orchestration and coordination. Capital deepening is low because firms are not deepening their own autonomous capability; they are renting it from upstream providers, which is consistent with CSD exposure and the paper's prediction that value concentrates where scarce judgement governs autonomous throughput (the upstream providers). The observed productivity lag behind perception is important: it suggests that the economic gains are real but not yet as large as executives expect, leaving room for the paper's revenue durability and commercial alignment tests to play out.
This is the framework's reading of the fact above. The source does not endorse it and is not cited as doing so.
The survey measures the state of AI adoption and productivity circa mid-2026. It shows that the mechanism of value reaccumulation (LAW_3, in particular) is plausible: gains are coming through innovation and demand, not cost-cutting, and value is concentrating upstream among AI providers rather than being captured by adopting firms. However, the observation covers a snapshot in time. The paper's test for LAW_3 requires that value does not move substantially toward orchestration, verification, governance, absorption and judgement within 3 to 5 years of credible deployment in a major sector. This survey does not yet show movement toward such roles within firms; it shows movement away from routine clerical work. Longitudinal tracking of whether firms begin to hire for orchestration, governance and verification roles in future surveys would test LAW_3 directly.
Current evidence is recorded immediately. The paper's longer empirical horizon determines when persistence becomes a stronger framework-level test.
The lag between perceived and measured productivity gains may indicate that firms overestimate AI impact and will eventually cut costs once the technology matures and is better integrated. The low capital deepening might reflect only the early stage of adoption, with deepening to follow as firms move beyond pilots to scaled deployment. The shift toward skilled technical roles could be temporary reallocation before further automation; it does not prove that orchestration and judgement become the durable locus of value. Renting intangible capital is consistent with standard outsourcing and does not necessarily imply that value is moving to a small set of scarce providers; it may reflect transaction costs and integration challenges that improve over time.
Recorded for every observation assessed at strength 4 or 5. A record that cannot state the strongest competing reading of its own evidence is not published.
Where several outlets report the same development, they are recorded as additional sources on this one observation and are not counted as separate evidence.