When 95% of high-tech employees use AI regularly and 1% of construction firms do, the question is not whether AI raises productivity — but whose?

There is a number that ought to keep Israeli policymakers awake at night, and it is not a casualty figure or a debt-to-GDP ratio. It is 95% — the share of Israeli high-tech employees who now use AI tools regularly, 78% daily, per a November 2025 survey by the Israel Innovation Authority and the Myers-JDC-Brookdale Institute. The global comparator, from Microsoft and LinkedIn, sat at 75%. Israel is not catching up to the AI frontier. Israel is the frontier.

And yet, walk a few kilometres from the Tel Aviv glass towers to a construction site in Bnei Brak or a hotel kitchen in Eilat, and the numbers collapse. The Central Bureau of Statistics reports that just 1% of Israeli construction firms made paid use of AI in June 2025. In trade and commerce, 11%. In services and hospitality, 16%. The high-tech sector reports 55% paid plus 5% free. The gap between Israel’s leading sector and its lagging sector is not a gap. It is a chasm.

This is the global productivity paradox in microcosm — but Israel exhibits it in the most extreme form on the planet. An NBER survey of nearly 6,000 senior executives across the US, UK, Germany and Australia found 69% of businesses actively use AI, yet 89% report no detectable impact on productivity and 90% no impact on employment over three years. PwC’s 2026 AI Performance Study of 1,217 executives across 25 sectors found 74% of AI’s economic value is captured by just 20% of organisations. The distribution is not flattening. It is sharpening.

Three financial-economic frameworks make sense of what is happening, and why Israel sits at the epicentre of all three.

First, real options. Adopting AI at scale is not a single decision. It is a sequence of irreversible investments — data infrastructure, workflow redesign, governance, training — each of which only pays off if the preceding ones were made. The 20% of firms capturing 74% of the value are not luckier or smarter. They exercised their options earlier, when premia were lower, and hold the complementary assets that make further exercise profitable. The remaining 80% are holding out-of-the-money options on a depreciating clock. Every quarter they wait, the strike price effectively rises as leaders pull ahead and the talent pool shrinks. The rational response is not patience. It is exercise — even at apparently uneconomic terms — because the alternative is option expiry.

Second, fat tails. Controlled trials are starting to map where AI’s productivity gains land within the workforce distribution. A Harvard Business School and Boston Consulting Group field experiment found AI lifted consultants’ task quality by roughly 40% and speed by about 25%, with the largest gains accruing to below-average performers. Customer-service studies show a similar pattern: roughly 34% gains for novice agents, negligible or negative for experts. Findings on experienced developers are more contested. Together the results suggest AI is acting as a skill leveller — raising the floor while doing relatively little for the top. In an economy where high-tech alone contributed NIS 352 billion, or 18.3% of GDP, and roughly half of total growth in 2025, compressing the lower tail is welcome — but cannot substitute for raising the top.

Third, the macro-micro paradox. The BIS and European Investment Bank, drawing on 12,000+ European firms, measure a +4% short-run labour productivity gain from AI adoption with no adverse employment effect — concentrated in medium and large firms. The Atlanta Fed and Richmond Fed’s survey of nearly 750 CFOs puts the mean AI-driven gain at 1.8% in 2025. And yet Daron Acemoglu’s task-based model implies just 0.07 percentage points of TFP growth per year over a decade, and the IMF’s medium-term estimate sits around 1% of cumulative TFP. The dispersion exists because firm-level and national-accounts productivity measure different things. Firm-level gains include rents extracted from competitors who have not yet adopted. National accounts net those out. When laggards catch up, much of the apparent gain at leading firms disappears into lower margins and consumer surplus, neither of which shows up as TFP.

For Israel, this means the 18.3% of GDP coming from high-tech is partly an illusion of leadership — one that will erode as Indian, American and Emirati competitors close the adoption gap. The Israel Innovation Authority’s 2026 State of High-Tech Report shows the warning signs already: for the first time in over a decade, R&D employment in Israel fell, with roughly 3,500 fewer R&D roles. By March 2026, only 62% of employees at private Israeli high-tech companies were based in Israel, down from 69% in 2019. AI does not just compress productivity distributions. It compresses geographic ones, because a software moat that AI dissolves no longer requires a Tel Aviv postcode.

The temptation, in Jerusalem as in every other capital, is to subsidise AI adoption at the leading edge — more chip fabs, more compute credits, more visa fast-tracks for AI researchers. This is fighting the last war. Frontier firms in Israel will adopt regardless; they already have. The gains not yet captured are in the firms outside high-tech that, by global benchmarks, sit one or two standard deviations below where they could be. A construction firm using AI to optimise material schedules will not generate a unicorn. It will, however, raise the productivity floor of an economy whose top is increasingly geographically mobile.

The Bank of Israel’s research division has noted that AI threatens substitution most acutely in the high-tech sector itself, because tasks there are most codifiable. The policy implication is uncomfortable: Israel’s productivity miracle is most exposed where it has been concentrated. The cushion lies in the sectors the Tel Aviv financial pages routinely ignore — logistics, construction, retail, agriculture, public services. These are the out-of-the-money options the state can help bring into the money, through procurement standards, sectoral data pools, and adoption credits tied to measurable outcomes rather than software licences purchased.

There is one further consideration that Israel, of all countries, ought to take seriously. AI productivity gains accrue most reliably to firms with the data infrastructure, governance maturity, and skilled human capital to absorb them — precisely the institutional assets that war and political volatility erode. The Bank of Israel has estimated the gross economic cost of the war at roughly $80 billion across 2023–2025; defence spending reached 8.8% of GDP in 2024, the world’s second-highest, and debt-to-GDP climbed from 61.3% to 69% in a single year. The Ministry of Finance has put the cost of a single reservist at around NIS 48,000 per month. A country whose human capital has been periodically called away from the desks where productivity is generated cannot assume 95% adoption translates into 95% uplift. The numerator is technological. The denominator is institutional. And it is the denominator that has been under strain.

Israel was not built for an easy productivity story. It was built for the hard one — where leadership at the top must be reconciled with breadth across the middle, where geographic concentration must be defended against AI-driven dispersion, and where the option value of being early must be cashed in before the laggards catch up. The data say Israel is winning the race at the frontier. The data also say the race at the frontier is no longer the one that matters.

Appendix: Policy recommendations

What the analysis implies for Israeli policymakers.

1. Reweight the AI subsidy stack toward the lagging sectors. Israeli AI policy currently runs heavily through Innovation Authority grants, defence-tech procurement, and compute credits — all of which flow toward frontier firms. The marginal return from another shekel of compute credit at a Tel Aviv unicorn is far smaller than from getting a mid-sized logistics or food-processing firm onto its first usable AI workflow. Target: at least 40% of public AI support directed at firms outside high-tech by 2028.

2. Establish sectoral data pools for SMEs. The binding constraint on AI adoption in construction, retail, agriculture, and hospitality is rarely software cost. It is the absence of proprietary data needed to train or fine-tune useful models. Government-curated, opt-in sectoral data pools — anonymised maintenance records, demand data, soil and yield data — would lower the entry cost for laggard firms without picking winners.

3. Tie small-business adoption credits to measured outcomes, not software licences. Subsidising the purchase of AI tools rewards adoption-without-use. Subsidising measurable outcomes (output per worker, throughput, error rates) rewards firms that actually integrate AI into workflow. The administrative cost is real but manageable through tax-credit clawbacks.

4. Treat geographic relocation as a productivity issue, not just a security one. The decline in Israel-based R&D is framed as a brain-drain story. It is also a productivity-capture story: when R&D shifts to Austin or Bucharest, Israel loses the agglomeration externalities that let other Israeli firms learn from it. Tax incentives tied to retention of frontline R&D roles (rather than to startup formation) would address this directly.

5. Resource the Bank of Israel and CBS to measure AI’s actual productivity effect. The macro evidence base is fragmentary. Israel is well-placed to lead on measurement — the CBS Business Tendency Survey already runs a near-real-time AI module, and Israel’s compact size makes longitudinal firm-level tracking feasible. Investing in measurement now would let policymakers distinguish capture from leakage as the data accumulate.

6. Recognise that the institutional denominator binds. Wartime mobilisation, fiscal strain, and emigration of skilled workers erode the absorptive capacity for AI gains. Policies that strengthen the denominator — predictable defence budgets that crowd in civilian investment, retention programmes for high-skill emigrants, governance reforms that reduce country risk premia — are AI productivity policies, even if not normally labelled that way. The numerator cannot pull the country forward if the denominator continues to shrink.