For decades, when a company wanted to make a complex internal process faster or cheaper, they hired specialists, ran months of analysis, and made incremental improvements. AlphaEvolve proposes a different approach: give the problem to an AI system, let it generate and test thousands of possible solutions automatically, and surface the best one in days rather than months.
The way it works is straightforward once you strip away the technical language. You give the system a description of your problem and a way to measure whether a solution is good or bad. The AI then generates hundreds of variations of a possible solution, tests each one against your own measurement rules, keeps the best performers, and uses those as the starting point for the next round. This loop runs continuously until the system finds something significantly better than what you started with.
Google has been running this internally for over a year. It found a circuit design for its custom AI chips so unusual that human engineers would not have considered it, and that design is now embedded in Google's next generation of chips. It also cut the amount of unnecessary data Google's databases write to storage by 20%, which matters because every unnecessary write costs money and slows things down at scale. Work that previously required months of specialist engineering effort took the system two days.
Now Google is offering this to outside companies through Google Cloud, currently in an early access programme where you need to contact a Google Cloud representative to get in. The commercial results published so far are striking. FM Logistic, a Polish warehousing and logistics company operating facilities spanning eight football fields, used it to optimise delivery routing and cut over 15,000 kilometres of distance per year. Klarna, the buy-now-pay-later company, doubled the training speed of a core AI model while also improving its accuracy. Schrödinger, which builds software for drug discovery and materials research, got roughly a four-times speedup in running molecular simulations, compressing what used to take months into days.
WPP, the advertising group, achieved a 10% accuracy improvement in campaign targeting models. That number sounds modest, but in advertising where margins are thin and volume is enormous, 10% accuracy gains translate directly into revenue.
The wider pattern here is worth paying attention to. Meta, entirely independently, has built a near-identical system called KernelEvolve, which improved the speed of its advertising model processing by 60% in hours of work, a task its engineers estimated would take weeks. Two of the world's largest technology companies have arrived at the same conclusion separately: using AI to automatically improve your own internal systems is now one of the highest-return investments available.
This matters for non-technology businesses too, because the types of problems AlphaEvolve solves are not unique to tech. Routing problems exist in logistics, insurance, retail distribution, and manufacturing. Scheduling problems exist in every industry. Model accuracy problems affect anyone using data-driven forecasting, from insurers pricing risk to retailers managing inventory. The system works on any problem that can be described clearly and measured objectively.
The honest limitation is that AlphaEvolve cannot help with problems where the answer requires human judgment to evaluate. It needs a clear, automatic scoring method. If your problem is ambiguous or requires qualitative assessment, this tool is not the right fit. But if your problem involves finding the most efficient route, the fastest process, or the most accurate prediction from structured data, this category of AI is now commercially available and producing double-digit improvements for the first companies using it.
The companies that understand this early will have a meaningful head start. Not because of the technology itself, but because knowing which of your internal processes are candidates for this kind of optimisation is itself a strategic advantage.