How Bain Turned Vibecoding Into a Build vs Buy Weapon

There is an old question that every company asks before it writes a check for software. Do we build this thing ourselves, or do we buy the company that already built it? For most of the last two decades the answer leaned hard toward buy. Good enterprise software was assumed to be expensive, slow, and painful to replicate, which is exactly why acquirers paid steep multiples for it.

Bain & Company just put a price on that assumption, and the price is dropping fast.

According to a Financial Times report published on June 22, 2026, relayed by The Decoder and Private Equity Wire, Bain's consultants have been quietly vibecoding rough replicas of the very companies their private equity clients are thinking about buying. Not the full product. Not the customer base. Just enough of the core functionality, built by describing what they want to an AI coding tool, to answer one uncomfortable question before anyone signs: how much of this would we actually need to buy?

From a specialist team to ordinary consultants

This did not start as a stunt. Bain says the practice began back in 2023 with a dedicated group of software engineers. What changed is who holds the keyboard now. The firm reports that hundreds of these prototypes have been built, and the work has moved from that specialist team out to rank and file consultants who are not engineers by trade. They lean on tools like Anthropic's Claude Code and agents in the same family to reconstruct a target's product from the outside in.

Rebecca Burack, who runs Bain's global private equity practice, framed the value of an interactive clone as the difference between seeing something "in 2D versus 3D." A spec sheet tells you what a product claims to do. A working replica tells you whether the clever part was ever clever.

The point of the exercise is narrow and a little brutal. A traditional technical audit checks whether the code works. The clone test checks whether the code matters. If a small team can stand up a believable version of your platform in a few days, then the thing you were selling as a moat was really a head start, and head starts expire.

It is already killing deals

This is not a thought experiment that lives in a slide deck. Two Silicon Valley private equity executives told the FT they had slowed their dealmaking and were now stress testing AI risk in every target they looked at. One put it bluntly: if AI exposure shows up in the risk column, they walk. In the words of the other camp, "If it's in the question box, we're not going to touch it."

A second investor went further and described a specific case. Bain's vibecoded recreation of an analytics platform fed directly into their firm's decision to drop out of the bidding. The replica did not have to be production grade. It only had to be good enough to make a nine figure premium look optional.

The public markets are pricing the same fear. Enterprise software heavyweights including Salesforce and ServiceNow have shed more than a third of their value this year as investors reprice how durable software businesses really are. The private side moved faster and harder. KPMG data cited by the FT shows the total value of private equity led technology, media, and telecom deals fell 69 percent in the first quarter of 2026 against the final quarter of 2025. That is not a wobble. That is a market deciding it no longer trusts the old story about what software is worth.

What the clone cannot copy

Here is where the panic deserves a cold shower. The clone test is good at exposing one specific lie, which is the idea that a polished interface equals a defensible business. It is much worse at valuing everything the interface sits on top of, the unglamorous engineering that only surfaces in production.

Maor Shlomo learned this from the inside. He built Base44, an AI app building startup, with a tiny team and no outside funding, then sold it to Wix in June 2025 for roughly $80 million. He has since argued that vibecoding tools like his were genuinely easy to clone at the surface, and that the hard, valuable parts were the unglamorous plumbing underneath. Databases that stay consistent under load. Authentication that does not leak. Analytics that people trust. The visible magic, in other words, is usually the cheapest thing to copy.

Lovable makes the same case from the winners' bracket. The Swedish coding startup crossed $100 million in annual recurring revenue roughly eight months after launch and raised a $200 million Series A at a $1.8 billion valuation in 2025, according to reporting from TechCrunch. Investors were not paying for a codebase that an afternoon of prompting could reproduce. They were paying for distribution, retention, and the speed of a team that kept shipping.

There is a folk version of this point that engineers keep repeating online. You can vibecode something that looks like Epic's MyChart in a weekend. You cannot vibecode your way into the hospital contracts, the regulatory approvals, the switching costs, and the years of integrated patient data that make MyChart unkillable. The feature set was never the moat. It just looked like one when copying it was expensive.

The new shape of build vs buy

So the old binary is gone, and a third option has muscled into the room. You can build, you can buy, or you can spend a few days cloning the thing first to find out which of those two is even worth doing.

For acquirers this is a genuine gift. CNBC ran a version of the same test on Monday.com back in February 2026 to gauge software stock exposure, and the technique has clearly graduated from healthcare and vertical niches into the broad enterprise category. A buyer can now interrogate a moat before committing capital instead of discovering its absence in year two of integration.

For founders the message is colder. Convenience sold as defensibility has a shelf life now, and the test is cheap enough that someone will run it on you. The companies that survive the weekend clone test will be the ones whose value never lived in the screen to begin with. Proprietary data nobody else can legally gather. Workflows wired so deep that ripping them out costs more than tolerating them. Networks and habits that took a decade to form and cannot be prompted into existence.

The good news, if you can call it that, is that the test is honest. It does not punish good software. It punishes software that was only ever pretending to be hard.