Nearly nine in ten firms surveyed by the National Bureau of Economic Research across the US, UK, Germany and Australia say artificial intelligence has had no measurable effect on their productivity or employment over the past three years. The finding, drawn from roughly 6,000 chief executives and senior finance leaders, lands at the centre of a widening split: markets price AI as transformational, boards commit nine-figure budgets, and the macroeconomic data registers almost nothing. Economists now invoke Robert Solow’s 1987 productivity paradox to explain the gap. For UK boards, the message is not that AI has failed. It is that value capture has not yet followed capability.
The deployment gap behind the headline number
The evidence arriving in early 2026 describes an adoption pattern that is wide but shallow. Two-thirds of executives in the NBER sample report using AI, but the median usage is only about 1.5 hours a week. A quarter have not used it at all. Against this, global AI capital expenditure exceeded $250 billion in 2024 according to the Stanford AI Index. The distance between investment and usage is the most important number in the dataset, and it explains why productivity gauges remain flat whilst earnings calls brim with AI mentions.
When Robert Solow observed in 1987 that “you can see the computer age everywhere but in the productivity statistics”, he was describing an economy that had bought the hardware but not reorganised around it. The same pattern holds now. Firms have procured AI tools; few have rewritten their workflows, governance or performance measures to extract the value those tools can generate. The NBER data shows firms expect a 1.4% productivity lift over the next three years, which would make AI consequential but far from revolutionary at the level of published accounts.
Critical numbers
| Metric | Value | Source |
|---|---|---|
| Firms reporting no AI impact on productivity or employment (3 years) | ~90% | NBER, February 2026 |
| Executives using AI regularly | ~66% | NBER |
| Average executive AI usage | ~1.5 hours per week | NBER |
| Executives not using AI at all | 25% | NBER |
| Expected productivity gain (next 3 years) | 1.4% | NBER |
| Expected employment cut (next 3 years) | 0.7% | NBER |
| Global AI capital expenditure (2024) | >$250bn | Stanford AI Index 2025 |
| Excess cumulative productivity growth since ChatGPT | 1.9% | St Louis Fed, November 2025 |
| Acemoglu productivity estimate (next decade) | 0.5% | MIT, 2024 |
| Brynjolfsson estimate of US productivity jump (2025) | 2.7% | Stanford Digital Economy Lab |
Strategic Reality: The gap between AI usage and AI effect is not primarily a technology problem. It is a deployment problem that compounds across organisations lacking the operating disciplines to translate tool access into process change.
What the firms with no results have in common
Three patterns keep appearing across the survey evidence, the academic literature and the consultancy studies.
First, adoption is treated as procurement. Licences are bought, seats are assigned, and training is delivered once. The tools then coexist with unchanged workflows, which means the marginal minute of work still runs through the old process. When AI is laid over an unchanged operating model, the gains come as small, per-task time savings that are absorbed by the next task rather than released as measurable productivity.
Second, breadth is mistaken for depth. The Boston Consulting Group study of 1,488 full-time US workers found that productivity rises when workers use up to three AI tools and falls sharply when they use four or more. Respondents described “brain fry”: small mistakes, context-switching costs, and a sense of cognitive fog when expected to orchestrate many tools at once. Procurement teams buy for coverage; outcome data rewards concentration.
Third, the benefits that do exist are leaking into consumer surplus rather than firm surplus. Stanford’s household study found that generative AI boosted the efficiency of online tasks such as job hunting and shopping by between 76% and 176%, but the time saved was spent on leisure and television rather than skill building or work. Apollo chief economist Torsten Slok captures this in his observation that AI is “everywhere except in the incoming macroeconomic data”.
Implementation Note: If your deployment pattern is broad licensing without workflow redesign, expect to replicate the 90% no-effect finding inside your own firm. The intervention that shifts the outcome is not more seats. It is fewer, deeper integrations tied to a specific operating metric.
The pattern in firms that do show gains
Firms reporting measurable AI productivity gains tend to share three features. They rebuild one workflow at a time around a single AI capability rather than sprinkling AI across dozens. They instrument the process before and after so value is legible. And they treat middle management, not the tools, as the bottleneck, because that layer translates capability into habit.
The human factor sitting beside the data
ManpowerGroup’s 2026 Global Talent Barometer, which surveyed nearly 14,000 workers across 19 countries, found that regular AI use rose 13% in 2025 whilst worker confidence in AI’s usefulness fell 18%. A workforce that uses the tools but no longer trusts them will not self-organise into higher productivity. Leaders who read the NBER data as purely a measurement issue are missing the trust data sitting beside it.
Stakeholder impact
| Stakeholder | Primary effect | Strategic implication |
|---|---|---|
| Boards and CFOs | Capex commitments exceed measurable returns | Defend long-horizon investment with interim adoption KPIs, not just productivity gauges |
| Middle managers | Asked to redesign workflows with no authoritative playbook | Central to outcome; should be resourced, not bypassed |
| Operational staff | Tool fatigue and confidence collapse despite rising usage | Concentrate the tool stack; measure trust as a signal |
| Early-career hires | Automation displaces entry-level tasks | IBM response: triple Gen Z hiring to protect the leadership pipeline |
| Investors | Sector earnings outside the Magnificent Seven show no AI lift | Expect scrutiny on productivity-linked reporting |
Success is not “we deployed AI” or “our staff use AI daily”. It is a measurable change in a specific process metric (cycle time, first-contact resolution, revenue per employee) that an independent observer could confirm. Boards should insist on this framing before approving the next investment tranche.
Critical Context: IBM’s chief human resources officer announced that the firm will triple its intake of early-career hires despite AI taking over parts of entry-level work. The reasoning is that cutting graduate intake now would empty the pipeline of future middle managers. This is the most concrete counter-argument to the “AI replaces juniors” thesis circulating in UK boardrooms.
A disciplined response by adoption maturity
The data is not an argument against AI investment. It is an argument for disciplined AI investment. The path that distinguishes firms capturing value from firms writing down goodwill has a shape, and it varies by adoption maturity.
Pre-adoption firms. Do not begin with a licensing deal. Begin with one process that already has measurement infrastructure and a committed owner. Pick a workflow where the baseline number is credible, where the AI-augmented version can be tested against the old one within 90 days, and where the owner has the authority to change standard operating procedure. Buy narrowly to fit that test.
Early-adoption firms with broad licences and thin usage. Consolidate. Identify the three tools that account for the majority of productive use and retire the rest. Reinvest the saved licence cost in workflow redesign sessions for middle managers. Publish a quarterly AI outcome review that tracks the two or three metrics you committed to, not generic usage statistics.
Mature-adoption firms with measurable gains in some units. Codify. Turn winning patterns into an internal playbook and an internal referral network. Move from “each team figures out AI” to “we share what worked”. Pair every new AI deployment with a review four months later that either scales it, kills it or redesigns it.
SME Advantage: Smaller firms can compress the three stages into weeks rather than years. A fifty-person business that picks one workflow and rebuilds it around AI can outrun a thousand-person competitor stuck in broad-licence stasis. The productivity paradox is largely a large-enterprise phenomenon.
A four-step sequence applies across maturity levels. Identify one workflow with a credible baseline. Redesign the process, not just the toolset. Measure the before and after against pre-agreed metrics. Review after one quarter and decide whether to scale, pivot or stop.
Four hidden challenges that derail AI strategies
These rarely surface in vendor conversations, and each has a specific mitigation.
The measurement blind spot. Firms commit to AI investment without instrumenting the processes they hope to improve. By the time the productivity question is asked, there is no clean baseline to compare against. The fix is to set up measurement before the deployment, not after. Treat the first month of any AI project as a measurement project.
The consumer surplus leak. Stanford’s research shows that efficiency gains at the individual level tend to be reinvested in leisure rather than captured as work output. Without explicit expectations about how saved time is used, firms pay for the tools and their competitors benefit from the rested workforce. The mitigation is to design the workflow around the time savings in advance, whether that is serving more customers, running more experiments or training more juniors.
The confidence collapse. The ManpowerGroup data shows AI usage and AI trust diverging. Once workers stop trusting their tools, errors rise, workarounds proliferate, and the data returned by AI systems is treated as suspect even when correct. Boards should instrument trust as a leading indicator with regular, short pulse surveys rather than discovering it as a lagging one.
The pipeline hollow-out. Using AI to replace entry-level roles is the most visible cost-saving move and among the most strategically damaging. Firms that reduce graduate intake now will face a middle-management gap in 2030. IBM’s response is instructive: maintain or grow early-career hiring, and redesign the junior job to work alongside AI rather than removing the role.
Warning ⚠️: Firms most aggressively cutting junior roles to fund AI infrastructure are setting up a leadership succession crisis in the second half of the decade. This is the most expensive hidden cost in the current adoption wave.
The strategic takeaway for UK boards
The central lesson of the productivity paradox is one boards can act on without waiting for new data. Value capture lags capability by years when firms procure tools instead of redesigning work. The ones that redesign early capture the gains; the ones that wait find their competitors have done the redesign on their behalf.
Three factors separate firms already seeing measurable AI productivity gains from the 90% that are not:
- Concentration over coverage. Three tools used deeply outperform ten tools used shallowly.
- Workflow redesign over tool deployment. The operating model is the bottleneck, not the technology.
- Measurement discipline over usage metrics. Outcome data beats adoption data for every strategic decision that matters.
Next-quarter checklist
- Name one workflow where AI can be tested against a measurable baseline within 90 days
- Assign an accountable owner with authority to change the standard operating procedure
- Instrument the baseline before procurement, not after
- Consolidate to the three most-used tools in any unit already running broad licences
- Add AI trust as a quarterly pulse-survey metric alongside usage
- Review early-career hiring plans against the 2030 middle-manager implication
- Publish quarterly AI outcome reviews to the board, focused on process metrics rather than licence counts
Take Action: The window for building internal AI capability without paying a competitive premium is closing as J-curve firms start publishing outcomes. The boards that begin workflow redesign this quarter will be the ones quoted in the 2027 Financial Times op-eds. Book one discovery session with a workflow owner this week.
Sources and attribution
This analysis draws on reporting by Sasha Rogelberg for Fortune (“Thousands of CEOs admit AI had no impact on employment or productivity — and it has economists resurrecting a paradox from 40 years ago”, 19 April 2026), the National Bureau of Economic Research working paper No. 34836 (February 2026), the Stanford AI Index 2025, the Federal Reserve Bank of St Louis State of Generative AI Adoption report (November 2025), MIT research by Daron Acemoglu (2024), Boston Consulting Group’s “AI brain fry” study published in Harvard Business Review (March 2026), ManpowerGroup’s 2026 Global Talent Barometer, and commentary from Apollo’s Torsten Slok and Stanford Digital Economy Lab’s Erik Brynjolfsson.
This strategic analysis was prepared by Resultsense to help UK boards translate headline findings into implementation decisions. For tailored guidance on AI adoption strategy, visit /contact/.