Naftali Bennett was a software founder before he was a prime minister, which is why he knows the gap in pace he talks about from both sides. He built and sold companies, then ran a country, and he has been out of government for four years. His first company, he says, had a hundred and forty employees. That exact same product, by his own estimate, could be built today with seven or eight people.
That double vantage point is the whole reason this conversation was worth having. “AI is not another technology,” he told me on the Autonomous Business Podcast. “It is the infrastructure for everything.”
The question that keeps returning across the half hour is how a state operates at the pace of something moving faster than it is. Bennett has an answer, and he has exactly one precedent. “Government work at startup pace, that was COVID.”
Across the episode he laid out a five point plan for turning the startup nation into an AI nation, explained the working method he believes a state has to adopt, conceded one campaign he thinks Israel is no longer in, and said exactly what he wants to be able to point at by the end of 2027.
Full disclosure: I am a member of the AI advisory committee he assembled. That is also how I know it is not decoration.
Constant Friction
His framework begins by refusing the category everyone puts the subject into. It is not sectoral and it is not a passing trend. “The best parallel I find for AI is electricity,” he said. Electricity did not improve one industry. It became the floor beneath all of them. “It is a horizontal resource that will serve all the verticals.”
The second half of the framework is that none of it waits for our decision. “With us or without us, the pace of progress here is unprecedented,” he said, adding that the pace produces enormous challenges and enormous opportunities in equal measure. He is enthusiastic and he is not naive about it. He expects hard problems at the global level and the national level, alongside vast opportunity. “First of all you need a doctrine.”
So what do you do against an unknown accelerating at that rate, when we are, in his words, already brushing right up against the level where AI surpasses humans in every cognitive domain. His answer is two words, and he repeated them. “My answer is constant friction. Simply constant friction.”
What he means by it is borrowed directly from the world he came from. You cannot forecast ten years ahead, and pretending you can is the failure itself. “A good startup does not plan ten years ahead, and not five years ahead either,” he said. “It runs forward, comes out with a product, touches customers, comes back with the feedback in very short iterations, corrects while moving.”
The argument is that the state now has to operate in exactly the same way. Enter it in the present and not in the future, intensively, and produce short iterations at the level of the state in order to keep steering the event at all times. That is an unusual thing to hear from someone who has already sat in the chair, because the chair is built precisely to prevent it.
That, in his view, is also the explanation for why the subject is almost absent from public debate. I asked him how that is possible in a country whose entire economic story is technology. The analogy did the work. “It is as if we are four or five years after the discovery of electricity,” he said, “and no politician wants to touch the subject, because who knows what will happen.” He reads it the other way around. “This thing, in my eyes, is genuinely strategic. And it is such a big opportunity. Which is why I am completely on it.”
What Worked During COVID
The self-evident objection is the one I put to him, and any founder would raise it. A government cannot run iterations. It has too many parts that depend on one another, and its tolerance for error is close to zero.
He did not argue. “The risk bar is much higher,” he said. “We cannot do stupid things and make mistakes that are too big.” He describes the public system as “very hierarchical, very slow,” which is not a controversial position coming from a former prime minister, only an unusual one to hear said out loud.
And yet, he says, it has already been done once. During his term a pandemic war room operated under a projector who reported directly to him, and the cycle was daily. He would join the meetings, a decision would be made, and it went out into the field that same week. Not a committee, not a position paper, not an approval route through four ministries.
So the mechanism he proposes is a copy of that structure: a national AI war room with a direct line to the prime minister, one “that cuts through the bureaucracy,” existing specifically to sustain the same continuous contact and friction he described. The difference between the two is that COVID was an event with an end, and this is not. A war room built for waves of a pandemic closes when the pandemic is over. A war room built for infrastructure has to stay.
He wants to load the retraining problem onto that same body, and he named names: the Haredi male population, Arab women. Precisely the populations that AI will either bring into the labor market or leave behind.
It is also why the advisory committee exists. “I really believe, in any new field, in bringing the most brilliant people, the most diverse, sitting them down, getting smarter, and coming out at the end with a work plan,” he said. “Not letting clerks in Jerusalem prepare the plans.”
Five Points, and One of Them Is Aimed at Startups
The plan is deliberately concrete, and four of its five points are directed at the citizen rather than at the industry.
The first is bureaucracy. Every Israeli citizen will have an AI agent serving as their representative before the entire system, so that nobody drives to a licensing office again. His example is everyday on purpose: the agent notices that the car registration is about to expire, that a fee is due, and by the way, that a tax refund is waiting for you. All of it without you chasing anything.
The second is education. An additional private tutor for every child, synced to the material the child actually needs to know from the curriculum. He was careful to say that this is not instead of teachers but alongside them. What changes is the resolution. “It knows the child,” he said. “It knows how David likes to consume information, it knows exactly where he is struggling.” Instead of waiting for a test, there is a reading in real time.
The third is moving state systems onto AI as fast as possible, with hard service level agreements that commit the state to numbers: response time to the citizen, the pace of issuing a permit, a license or an approval.
The fourth is the one that touches everyone building here. “The State of Israel will be the testing ground for Israeli hi-tech,” he said. The government becomes the first customer instead of the last one, putting itself at the disposal of startups in transportation, in medicine, in education, in a controlled way. “We will be very, very inviting and embracing.” In his view the loop runs in both directions, because government ministries forced to work at that pace learn the pace, and learn something rarer: “that it is permitted to be wrong in certain areas.”
The fifth is infrastructure, at every level, and that is where the conversation went next.
He was willing to attach dates, which is not a given. By the end of 2027 he wants the private tutors in significant deployment. The bureaucratic agent he estimates will take longer, because it requires an enormous number of connections into state systems, and mainly because it requires a level of reliability the technology does not yet supply. “One of the problems with AI today that holds things back is that they are excellent but they are wrong from time to time,” he said. “And there are things where we cannot accept mistakes.”
We Will Not Build the Models. The Ground Beneath Them, We Can
Israel built the applications layer of the previous internet. ICQ, Waze, the list is long. The current layer runs on somebody else’s language models, which means Israeli startups build on infrastructure owned by another country. I asked him what happens if that dependency ever turns into leverage.
In his answer he separates the civilian dimension from the defense one and says the second is something he needs to study more. On the first he did not stumble. “You are right. The campaign over the LLMs, at least for now it looks like we are not there.”
What he would do instead is go one layer down. “Large Israeli data centers, robust, definitely in private ownership,” he said, running on open source weights. The reasoning is that the open ecosystem has turned out to be strong, alive and surprisingly good, and it is free in exactly the dimension that matters to a country without frontier lab budgets. “You do not carry their training costs, you only use them for inference,” he said. The training, where nearly all the cost sits, has already been paid for by somebody else. What remains is the cost of running them. A model like that can be adapted to local needs, compressed into a smaller and cheaper version, or stripped for parts.
The point of putting the iron on Israeli ground is sovereignty in the plainest sense. In the Negev, in Modi’in, “it does not matter where,” he said. “Then nobody can stop us.” He is worried about the visible version of that risk already, and points to export restrictions on NVIDIA chips as the kind of thing that arrives with no advance notice.
Then he listed what a project like that actually requires, and it is the least glamorous paragraph in his plan. “The land, the building, the chips, the energy, the water, the knowledge.” On the last item, in his view, Israel holds an advantage it has not worked: the generations of engineers and physicists who arrived here from the Soviet Union.
What he did not do is bluff. When the conversation touched the defense side of the same question, he stopped himself. “I do not want to answer. Not because I know and am hiding it, but because I need to learn it more.” That is a rarer sentence than any commitment.
Arrow Is Arrow. He Also Wants Things That Cost Sixty Dollars
The same gap in pace shows up in the defense industries, and there he speaks from familiarity. There is one subject he has been pressing on the defense establishment for five years, including the four in which he was a private citizen: swarms of cheap autonomous drones, running on internal AI rather than a remote control, which is why jamming frequencies does not help against them. “We have to prepare for it, the whole world has to prepare for it,” he said, stressing that he means both the defensive side and the offensive side, and adding of his own accord: “of course not against civilians.”
From there the argument turns from a security question into a procurement one. Israel has Rafael, Israel Aerospace Industries and Elbit, and he called them “remarkable industries” before naming their two hard limits: development time and cost. Everything there is priced in millions of dollars, in tens of millions, in billions. “Arrow is Arrow,” he said. “But I also want things that cost 60 dollars, and a lot of them.” At a pace he defines as super fast.
His argument is that the country already holds both conditions and has realized neither. Israel fights in built-up terrain and in open terrain, strikes at long range, and operates in cyber, which in his words makes it the country with “the widest range of warfare.” It is also the startup nation. “We have both of the perfect conditions to be the defense tech nation as well,” he said, “and it has not happened enough.”
I offered him the obvious framing: a Manhattan Project, except this time a Tel Aviv Project. He rejected it, and the correction is the same logic as the war room and the iterations. “The Manhattan Project was for one big thing, for the bomb. And here I am talking about thousands of small things.” Not one breakthrough that arrives after ten years. Many small ones that keep arriving.
The Seniors Got a Multiplier. The Juniors Lost the First Rung
The question I brought him from my own community was not about policy. Students who invested three or four years in a computer science or software engineering degree walked into a world that reset one morning, and some of them are in real anxiety.
He did not soften it. “The problem is a real problem, a genuine problem,” he said, and located it precisely at the entry stage. “The seniors have far greater productive power, but the juniors have no opportunity to become seniors if they never get to be juniors.”
His recommendation is the same thing he says about the state, scaled to a person. Do not stay away. Go in. Use all of the most advanced AI agents in every field, “and turn yourselves into actual productivity machines.” Then he pointed at where the demand sits, which is not where a computer science graduate expects it: integration. “Say I am now some big car rental company, I do not know what to do,” he said. “So if someone comes and puts it in order for me, builds me a plan and implements the plan, that is worth a great deal to me.”
The second opening he sees suits the Israeli ethos more than any clause in his plan: a mass of entrepreneurship at the level of a single person, or two or three, building companies that once required a hundred and forty. He thinks the fear is understandable and partly justified. He also thinks it is beside the point. “The only way is to go in and leverage it and ride it and not let it run you over.”
When I told him that AI has made me five times more productive and that I now work more than I ever have, he identified the paradox before I could formulate it. The Jevons paradox: efficiency is supposed to reduce demand, and in practice the computer became a hundred times better and the revenues of computer manufacturers went up, because productivity at that cost creates uses nobody imagined. It holds at the personal level too, he notes, which is either reassuring or the opposite.
What All of This Is For
He asked to close with a message aimed specifically at the people who are not in this world.
“AI is not, God forbid, something that is supposed to disconnect us from life itself,” he said. “The opposite. It is supposed to release us to travel the country, to be with family more.” Alongside it he wants the values layer preserved deliberately: love of the country, patriotism, strengthening the family unit. “Beautiful things. I am very optimistic about this.”
It is an unusual note on which to end a conversation about technology, and it is the sign of how he reads the whole problem. Everything else he described is mechanism: a war room, service level agreements, open source weights, a state that lets startups experiment on it. None of it is the goal. The goal is what gets done with the capacity that frees up, and he argues that Israel should decide that in advance rather than discover it after the fact.
And that returns to where he began. Electricity was not a better candle, and the countries that understood this first were not the ones with the good candles. They were the ones that laid infrastructure. We are, by his own reckoning, four or five years after the discovery. One war room will not close that gap, but without somebody sitting at the center and cutting, the gap does not begin to close either.
Mickey Haslavsky is the Founder and CEO of enso and a Forbes 30 Under 30 alumnus. He previously founded Rapid (acquired by Nokia) and hosts The Autonomous Business Podcast, in partnership with Forbes Israel. He is a member of the AI advisory committee assembled by Naftali Bennett.
