Bain tests software takeover targets by vibecoding AI replicas

Having done a stint in consulting, most internal software teams are useless, that's why these companies hire outsiders. They are sick of having to deal with stubborn teams who think they know everything, and refuse to change. You can see that mentality in these threads.

- A TDD loop where the Indian QA team played the role of the tests. The engineers would yeet some broken code at the end of the day, the poor QA testers would click through all the broken interfaces, and then the engineers would fix it the next day.

- A release process that was so slow and hellish that everyone just went to the DBA to have him add a stored procedure to implement their feature. He could get it in for you the next day.

- A frontend framework discussed in hushed tones, being built by a mysterious monk-like engineer, which was going to be the client's big secret weapon. In reality it was a terrible version of React built on top of jQuery.

- A core in-group of backend devs (most of these guys had advanced degrees for some reason), who would stay late every Friday, going through heroics to do a release of the client's email-templating app. There would be then be lots of back-slapping and congratulations the next Monday for these geniuses who were keeping the business afloat.

- "when in doubt, set timeout". They didn't know about callbacks

Usually consultants are brought in when upper management can tell that there is something very wrong and they can't fix it within the chain of command of their full-time staff.

I'm aware — I've used LLMs to find vulnerabilities, myself. But it doesn't follow that because AI can find them that AI can find the optimal fix, because fixing vulnerabilities often involves tradeoffs.

Also, can we please have a civil discussion?

But since you bring it up, those of us that have been around a while have used many code generation tools that took care of tedious work like model creation, forms, validation, class creation, CRUD, component creation, project skeletons, starter templates, visual design tools that generate code, etc... etc... basically most of the stuff that LLMs are good at. Simple things that are tedious and don't require much thought - the sorts of things that statistical pattern matchers like LLMs are good at.

The rest of it, the stuff that requires thoughtful architecture and reasoning is where LLMs fall flat and mostly end up costing time. The research supports this.

So yep, my face is indeed quite straight.

If your argument is that LLMs are more effective and less costly than typical low-quality outsourced code, I generally agree with that, depending on the details.

For prototype/POC stuff, absolutely.

For long term deployable product stuff that needs to be maintained, supported with SLAs etc... eh, like everything, it depends on the people/team.

I've been leading dev teams for around 25 years, and I tend to think in terms of total cost of code.

Cognitive debt is a real thing. The cost of a developer that leaves the team, and the struggle to have someone take over their work is real.

With LLMs you have all this hidden cost, plus no "theory of mind", which makes the cognitive debt much worse.

Combine that with atrophied skills, subtle bugs, architecture facepalms, security issues, etc... and the total cost of LLM generated code is a lot higher than just a subscription price.