Diagnosis is cheap. Every serious observer of Israeli technology has, at some point, written the essay explaining that Israel risks falling behind in AI. The harder and rarer document is an agenda: a short list of concrete things the state should do, specific enough that you can later check whether they happened. In December 2024 the author published five such recommendations — talent pipelines, compute access, public-sector adoption, research coordination, and export-oriented company formation. A year and a half later, the test of a policy agenda is not whether it was clever but whether it aged. This one has. The recommendations still describe the binding constraints, the logic behind each has strengthened rather than weakened, and the gap between what was proposed and what has been executed is itself the most instructive data point.
Why these five, and not others
The agenda was built on a simple discipline: recommend only things that are inputs, not outcomes. “Israel should have a world-class AI sector” is an outcome and therefore not a policy. Talent pipelines, compute, government adoption, coordination, and company formation are inputs — levers a state actually holds. The second discipline was to exclude anything that required Israel to win a spending contest. Nothing on the list asks the country to out-build hyperscalers or subsidize a national frontier model into existence. Each item instead multiplies something Israel already over-produces: trained people, operational problems, and founders.
Notice what the frame implies. If these five inputs are the binding constraints on the national AI sector, they are also the environmental variables for every individual Israeli AI company. A policy agenda, read carefully, doubles as a diligence instrument. That dual reading is the point of this research edition.
Talent and compute: the inputs that compound
The talent recommendation was never about producing more graduates — Israel's universities and military units already do that at world-class rates. It was about the pipeline's leaks: researchers who leave for hyperscaler labs because staying means losing access to frontier-scale work, and mid-career engineers with no structured path from cyber and signals-intelligence backgrounds into machine learning. The recommendation was retention through mission and instrumentation — research positions attached to hard national problems, with the data and compute to attack them seriously. Since publication, the underlying pressure has only intensified: as frontier work grows more capital-intensive, the gap between what an Israeli academic lab can offer and what a foreign lab can offer widens by default. Passivity here is a decision.
Compute was the second recommendation, and the one where the reasoning has aged most visibly. National-scale negotiation for guaranteed research and startup compute looked prudent in 2024; it now looks like table stakes, as every serious mid-sized country has reached the same conclusion. The strategic point was never that Israel should own all its compute — it is that access terms should be a matter of national arrangement rather than a per-company scramble, and that the dependency should be mapped and managed like any other strategic exposure. That is the same logic this site applies to cloud and semiconductor dependencies in the Dependency Atlas: you cannot manage an exposure you have not modeled.
The state as customer, the state as coordinator
Public-sector adoption was the third recommendation and remains the most underrated. Government is Israel's largest holder of consequential problems — logistics, health administration, benefits processing, border sensing — and its willingness to buy and deploy AI systems determines whether domestic companies get a demanding first customer or must sell abroad from day one. The recommendation was procurement reform in the specific sense that matters for AI: acceptance criteria based on measured performance, contract structures that tolerate weekly model updates, and deployment sandboxes where a system can prove itself on real data. Adoption is where policy stops being rhetoric; a ministry that fields an AI system has, by definition, solved the data-access, evaluation, and accountability questions that white papers only describe.
Research coordination, the fourth item, addressed Israel's peculiar institutional gap: excellent universities, excellent startups, and almost nothing that connects them at the applied layer. Small countries that punch above their weight in deep technology tend to have mission-driven applied institutes that own the space between paper and prototype. The recommendation was not another committee — Israel produces committees efficiently — but ownership: named institutions with budgets, deployment authority, and responsibility for specific problem domains. The cost of not doing this is invisible and enormous: every year, research results that could have become fielded capability instead become the seed insight of a company founded somewhere else.
Export orientation: companies as the delivery mechanism
The fifth recommendation ties the others together: the ultimate delivery mechanism for Israeli AI capability is export-oriented companies. Israel's domestic market is too small to sustain an AI sector; the sector exists to the extent that it sells into the United States, Europe, and allied markets. Policy should therefore treat company formation as the output stage of the national pipeline — talent, compute, government proving grounds, and applied research all feeding firms built from inception for allied export. This is where the agenda meets the thesis of this site's alliance proposition: Israeli dual-use companies are strongest when they are structured as allied capability, not as local champions.
The aging test is instructive here too. Since late 2024, nothing about the strategic environment has weakened the export logic — allied demand for trusted AI capability has grown, and scrutiny of un-trusted suppliers has grown with it. What the agenda could not fully anticipate was how fast the application layer would commoditize, which raises the bar for what an exportable Israeli AI company must own: not model access, but data, evaluation credibility, or an operational channel.
What this means for investors
Read as a diligence instrument, the five-part agenda produces one master question for any Israeli AI company in the startup database: does this company actually draw on the national inputs, or does it merely use AI language? A company staffed from the elite talent pipeline, with compute arrangements it can articulate, a demanding Israeli government or enterprise customer hardening its product, ties into the research base, and an export motion aimed at allied markets is levered to everything that makes Israel structurally interesting. A company with none of those properties is an AI-branded software firm that happens to have an Israeli address — investable, perhaps, but not on this thesis.
The screen is checkable in ordinary diligence. Ask where the technical team trained and what they deployed before. Ask what breaks if compute prices double or access tightens. Ask who the most demanding current customer is and what that customer measures. Ask which research relationships are contractual rather than conversational. The structured version of these questions lives in the site's diligence checklists, and companies that pass them cluster visibly when you sort the database by sector and channel. The agenda was written for policymakers. Its sharpest use so far has been sorting companies.
Bottom line
The December 2024 agenda has aged well precisely because it recommended inputs rather than outcomes: talent pipelines, compute access, public-sector adoption, research coordination, and export-oriented company formation. Those are still the five levers, the logic behind each has strengthened, and the slow pace of execution has cost Israel time without invalidating a single item. For policymakers, the agenda remains actionable as written. For investors, it doubles as the most reliable screen available for Israeli AI companies: fund the ones plugged into the national inputs, and treat the rest as vocabulary. Agendas that survive eighteen months of AI-speed change are rare. This one's persistence is the strongest evidence that it identified the real constraints.
Where this argument started
A shorter version of this argument first appeared as “Five concrete recommendations for Israeli AI” in The Times of Israel (December 2024). This research edition expands the argument with database context, diligence framing, and internal links for readers who want to act on it.