Companies keep pouring AI coding tools into their engineering teams, yet hardly any can tell whether the money actually improves their products. That blind spot gave rise to Milestone, an Israeli startup co-founded by Liad Elidan, who now helps enterprises tie AI usage directly to business outcomes.
Elidan made his case on the Jerusalem Post Business and Innovation podcast. His company works with organizations employing 200 or more engineers, connecting AI adoption to engineering output, code quality, governance, and bottom-line metrics.
The core problem, he argued, involves speed without direction. AI lets developers complete far more work in less time, but no one verifies whether that work matters or meets quality standards. “The gains, even doing it blindly, are so high that you better do it,” he said. “But you’ll need governance for the software development lifecycle.”
Milestone tracks cycle time, links individual tickets to AI models and token costs, and measures how long code survives in production before someone rewrites it.
Meanwhile, engineers left to their own devices gravitate toward the most powerful and expensive AI model, even for trivial tasks. “It’s a huge waste,” Elidan said. “The best engineers are the laziest. If I can choose a slow car or a fast car and I need to drive from Tel Aviv to Jerusalem, I’m going to take the fast one, if it’s free.”
He pointed to industry anecdotes of a single engineer racking up roughly $20 million in token costs within one month, discovered only after the fact. He declined to name the company.
As a result, code review has become a choke point. Senior engineers who once reviewed 100 lines daily now face 1,000, and attention per line drops. “If you’re in the ER and you have ten patients and now you have 30, you’re probably going to give less attention to each patient,” Elidan explained. Some organizations now push code to production when an AI reviewer signs off, with zero human oversight.
Human engineers still hold one advantage over AI agents. Agents operate inside narrow environments, while people can move across an organization’s systems and knowledge. That distinction is driving companies to deploy “forward deployed engineers” who wire agents into real workflows.
On cost, Elidan estimated a senior engineer running five agents might cost 25% more while delivering triple the output. But if the token bill matches several salaries and the output cannot be measured, the equation breaks.
Elidan compared the transition to replacing horse-drawn carts with cars. Some skills will fade, though he rejected the word “extinction.” By 2030, he predicted, one person could direct a swarm of agents across marketing, design, and code, becoming “a one-person small company.”
His outlook carries a warning wrapped in optimism. “I think it’s good to be paranoid,” he said. “The paranoid survive. You just need to evolve and adapt.”
