How we estimate the two modeled numbers
Two figures in the Index are estimates, not measurements: the 30-40% of the workforce plausibly retrainable to build with AI, and the ~$30M to retrain 10,000 residents. This page shows exactly how each is built, the inputs, the assumptions and the honest confidence. Everything else in the Index is a dated, sourced figure. These two are reasoned, and here is the reasoning.
Rule of the house: read the band, not the point. Both numbers are ranges for a reason. Change an assumption below and the number moves, on purpose. Nothing here is a quote or a promise.
The retrainable share (30-40%)
The claim is not that 30-40% of Haifa can become software engineers. It is that 30-40% of תושבי חיפה המועסקים have the cognitive baseline, the technical habits and the English to be taught the applied tier of building with AI (automations, agents, AI-enabled tools), given a short, employer-linked program. It is an estimate of a plausible pool, not a forecast of enrollment.
The inputs (each dated and sourced in the Index)
| Sub-metric | Value | What it establishes |
|---|---|---|
| Already in high-tech (District) | 17% | The immediate builders. ~1.5x the national rate of 11.5%. Essentially all English-capable. |
| In knowledge / desk-based occupations (ISCO 1-4) | 60.0% | Managers, academic professionals, technicians and clerical support. Academic professionals alone are 35.4% of this; the full desk-based pool is what the derivation uses. |
| Hold an academic degree (residents 15+, CBS LFS 2023) | ~41% | The cognitive baseline for reskilling. This is the latest city-level figure published: CBS releases locality-level breakdowns with a ~2.5-year lag, and the next update is due in the 2026 yearbook. |
| Israeli 3-20 week reskilling placement (benchmark) | 70-85% | Evidence the applied tier is reachable fast, in Israel, for a screened cohort. |
The derivation, from Haifa's actual occupational mix
The band is not asserted. It is computed from the real occupational distribution of employed Haifa residents, published by the Central Bureau of Statistics through the Haifa Statistical Yearbook (Labour Force Survey, 2023). Shares below are renormalised to the known-occupation base of 133,000 workers, because roughly 6% of the total is recorded as occupation unknown.
| Occupation | Share of employed residents | Retrainable to the applied AI tier? |
|---|---|---|
| Academic professionals | 35.4% | Yes. Degree-holding, desk-based, the core pool |
| Technicians, practical engineers, associate professionals | 14.3% | Yes. Technical baseline already present |
| Clerical support | 7.0% | Yes, and most urgently. Highest AI-automation exposure of any group |
| Managers | 6.8% | Yes. Domain judgment plus AI leverage |
| Service and sales | 21.0% | Partly, case by case |
| Skilled manufacturing and construction | 10.3% | Largely not, in the near term |
| Elementary occupations | 5.2% | Not in this model |
The arithmetic. The four desk-based, degree-adjacent groups (ISCO major groups 1 to 4) total 60.0% of employed Haifa residents as published, against 57.5% nationally. Renormalised to exclude the ~5.6% recorded as occupation unknown, the same pool is 63.5%. Not all of that pool will retrain: motivation, life stage, age and the reserve-duty drag below all bite. We assume roughly half to two thirds of this pool is realistically reachable (a participation-and-feasibility retention factor of 0.47 to 0.63), which yields 30% to 40%, the published band. Note where the uncertainty actually sits: the 63.5% input is measured, so the entire width of the band comes from that assumed retention factor. Change it and the band moves, which is the honest property of a model.
What this number is not, stated plainly
Every published instrument in this field, the OpenAI GPTs-are-GPTs exposure scores, the Felten AI Occupational Exposure index, and the ILO's ISCO-08 gradients, measures exposure: the risk that AI performs your tasks. Retrainability is a different claim. It is about the worker's capability, not the job's threat. So we do not dress an exposure index up as a retrainability finding. This band is our own construct, built from published occupational data using a documented method, and we say so.
For calibration, we also ran the standard exposure computation, weighting ILO 2025 ISCO-08 GenAI scores by Haifa's occupational mix. It puts Haifa at roughly 32% of employment in an exposed occupation against about 31% nationally. That is a real but modest edge, and it validates the pipeline against the ILO's 34% benchmark for high-income countries. It does not support a claim that Haifa dramatically out-indexes Israel on exposure. Haifa's case rests on composition, the combination of degrees, knowledge occupations, hi-tech concentration and cost, not on exposure.
The supporting reasoning
The three population sub-metrics overlap heavily (most high-tech workers hold degrees), so they are not additive. We take the degree-holding plus knowledge-occupation pool (roughly a third of the workforce, anchored by the ~41% degree share and the 60.0% desk-based occupation share, of which academic professionals are 35.4%) as the group with the baseline to reach the applied tier. Within it, the 17% already in high-tech are effectively there. We estimate that most of that broader pool, discounted for motivation, life-stage and the reserve-duty drag below, is plausibly retrainable, which lands the band at 30-40% of employed Haifa residents. The base matters and is stated deliberately: the derivation runs on the occupational distribution of people who are המועסקים, so that is the population the band describes. Extending it to the whole working-age population would overstate it, because the non-employed skew lower on both degrees and knowledge occupations. The Israeli 70-85% placement benchmark for short programs is what makes "plausibly retrainable" more than a hope: screened, employer-linked cohorts do convert at those rates.
The honest drag we subtract for
The share of the workforce doing reserve duty more than doubled, from 4% to 9.4%, and reservist employees averaged ~202 days away between October 2023 and January 2026 (University of Haifa and Hilan, 345,700 employees, April 2026). The acute phase eased, with under 2% of high-tech workers called up at any one time by H2 2024 (Taub Center), but conflict with Iran resumed in February 2026 and the ceasefire lapsed in July 2026, so the exposure is live. The latent pool is therefore large but the effective near-term pool is smaller. The 30-40% is the latent estimate, which is why the band's low end, not its high end, drives cohort planning.
Confidence, and what would sharpen it
- Confidence: medium, upgraded from medium-low. The band is now derived from Haifa's published occupational distribution rather than reasoned from population aggregates, and two independent CBS tabulations (city residents, and the Haifa sub-district) can cross-check it. Medium is the honest ceiling, not modesty: the occupational data is one-digit only, "retrainable" is our construct rather than a validated instrument, and the adoption assumptions borrow from survey data collected elsewhere.
- The method has precedent. Weighting an occupation-level index by a local occupational distribution is the standard published approach, used by the OECD at metropolitan-area level in Canada and Australia, by the UK Department for Education, by the US Treasury and by the Greater London Authority, which ran exactly this construction for a single city in 2026. What appears to be genuinely new is applying it to an Israeli city: the finest published Israeli resolution is district-level, in a single appendix figure.
- The remaining honest limit: public CBS data is published at eight major occupational groups only. "Academic professionals" is a single 35% bucket spanning software developers and social workers, whose AI exposure differs enormously. Two-digit occupational detail exists in the CBS microdata but requires a licence. We do not present eight-group weighting as occupation-level precision.
- What would sharpen it further: the two-digit occupational breakdown, and decisively the actual conversion rate of a first pilot cohort. The program is designed to generate exactly that and replace this estimate with a measured one.
- What it is not: a claim that people are already AI-fluent builders. It is the opposite: it measures who could be taught to build, which is the whole reason the program exists.
The cost to retrain 10,000 (~$30M)
A blended program mixes light AI-literacy for the many with a deeper build-with-AI track for those who will ship. The total is a desk estimate from public benchmarks, not a quote. Here is the arithmetic in full.
Per-learner cost
| Track | Cost / learner | Benchmark basis |
|---|---|---|
| Light AI-literacy | ~$800 | Mass short-course and MOOC-style delivery costs |
| Deep "build with AI" | ~$3,000-12,000 | US and Israeli applied bootcamps, employer-linked |
| Blended (70/30 light/deep) + 18% wrap | ~$2,960 | The figure used below |
The build-up
| Cohort | Cost (USD) | Shekels (NIS 3.05 = $1, Jul 2026) | Share of Haifa workforce |
|---|---|---|---|
| 1,000 residents | ~$3.0M | ~NIS 9M | ~0.7% |
| 5,000 residents | ~$14.8M | ~NIS 45M | ~3.5% |
| 10,000 residents | ~$29.6M | ~NIS 90M | ~7% |
The reference check: the full 10,000-person program (~$30M) is smaller than a single existing Israel Innovation Authority funding wave (NIS 139M for 9,000 people). This is fundable at Israeli scale today, which is the point of running the number.
The assumptions, stated so you can change them
- 70/30 light-to-deep split. More deep-track learners raise the blended cost; fewer lower it.
- 18% program wrap (management, placement, measurement, overhead).
- FX NIS 3.05 = $1 (July 2026 rate; the 2026 average is 3.04). Benchmarks drawn from San Jose Work2Future, Amazon upskilling, and US and Israeli bootcamp costs, plus IIA grant ceilings.
- Not a stipend model. This blended cost is the delivery cost. A paid-apprenticeship version, where learners receive a stipend so they can afford to take part, costs materially more per learner and is funded separately. The program page describes that design.
Change the split, the wrap or the FX and the totals move. That is the honest property of a model. What does not change is the order of magnitude: retraining a meaningful share of Haifa is a tens-of-millions program, not a hundreds-of-millions one, and it sits inside the range Israeli institutions already fund.
The composite score (75 / 100)
The headline Index score is a weighted blend of the eight dimensions, each scored 0-100 from the sourced sub-metrics in the report, against Israeli and global-hub peers and a fixed reference band. It is a reasoned, calibrated estimate with an author sensitivity range, not a measurement with a statistical confidence interval. Here is every input and the arithmetic.
The dimension scores and weights
| Dimension | Score | Weight | Why this score |
|---|---|---|---|
| יכולת הסבה מקצועית | 80 | 22% | 17% high-tech (~1.5x national 11.5%), ~41% degree-holders, 60.0% in desk-based occupations against 57.5% nationally, strong English. The most distinctive dimension, and the most modeled, so it carries the widest band. |
| זרימת הון | 85 | 13% | The highest-scoring dimension: ~$6.3B North/Galilee package, ~$4.6B high-speed rail under construction, ~$1.5B Nvidia server farm, ~€1B light rail at financial close, plus the unpriced Bay megaproject. Discounted because much is regional rather than Haifa-specific and the largest items are projections. |
| צפיפות חדשנות | 78 | 13% | Intel Core's birthplace, IBM's largest lab outside the US, Matam (~80 firms, ~15,000 staff), 1,000+ Technion startups. Held back by no neutral mega-lab and no fabs. |
| כלכלה | 72 | 13% | Among Israel's lowest unemployment (2.5% district), 17% high-tech. Discounted for flat civilian-software employment and record job-seekers. |
| כישרון | 70 | 13% | ~30,000 students across the Technion and University of Haifa, Israel #1 in AI-talent concentration. Dragged by net emigration. |
| ערך שוק | 80 | 9% | Cheapest of the major cities: average apartment NIS 1.88M vs NIS 3.03M in Tel Aviv (~38% less, CBS Q3 2025), led national price growth in 2024, +50% permits. |
| איכות חיים | 68 | 9% | Very high quality of life at ~17-25% below Tel Aviv cost and roughly half the rent, offset by pollution still rated high, though falling. |
| בריאות וסביבה | 60 | 8% | A real petrochemical health legacy, in active reversal (56% VOC cut, bay closure, a new park and hospital). The honest drag dimension. |
The arithmetic
(80 × .22) + (85 × .13) + (78 × .13) + (72 × .13) + (70 × .13) + (80 × .09) + (68 × .09) + (60 × .08)
= 17.60 + 11.05 + 10.14 + 9.36 + 9.10 + 7.20 + 6.12 + 4.80 = 75.37, rounded to 75.
Confidence and sensitivity
- Author sensitivity range: ~73-77. This is deliberately not called a confidence band. There is no sampling process here and the dimension scores are author-assigned, so no statistical confidence interval exists. It is the range across which our own reasonable disagreement about the scores moves the composite. Propagating a more honest +/-10 uncertainty on each dimension would widen it to roughly 68-82, and readers who want the composite to carry a genuine interval should treat that wider range as the honest one.
- Why Capital was added (2026-07-28): the original seven-dimension score of 74 omitted capital inflow entirely, which understated the case. Roughly $12B and more in identifiable committed or under-construction capital is landing in and around Haifa this decade. Adding it as a 13% dimension moved the composite from 74 to 75.
- Sensitivity: shifting the Retrainability weight by 10 points, or its score by 10 points, moves the composite by only ~1-2 points, so the number is stable to reasonable disagreement about the weights.
- What is next: scoring the same eight dimensions for Tel Aviv, Jerusalem, Beer Sheva and the global hubs, to turn the score into a ranked, like-for-like comparison. Until then it is Haifa's own scorecard, not a league table.
Why we publish this
An index that leans on two modeled numbers has to show the models, or it is doing the exact thing it exists to replace: an unsourced claim dressed as fact. These two estimates are honest, they carry ranges, and they are built to be replaced by measured data the moment a pilot cohort produces it. If you find an assumption you would set differently, that is the point. Tell us, and we will show you how the number moves.