I was listening to Zach Abramowitz's podcast with Will Chen, the ex-Latham lawyer who built an open source version of Harvey and Legora in two weeks using vibe coding tools. It got me thinking about something bigger than the episode itself.
What is everyone's actual issue with wrapper layers? Unless you're an investor assessing an opportunity, what is the reason to care that an application is a wrapper on top of the foundation model. Unless, of course, the cost of the wrapper is so extortionate that building yourself or configuring a foundation model is more economic.
I don't have a issues with wrapper/ harness applications, I'll explain why.
The build-it-yourself fantasy
Somewhere in legal AI right now there's a very online, very small group of lawyers who can and will open a terminal and build something. Will Chen showed what that looks like when he built a working, open source alternative to Harvey and Legora in two weeks using vibe coding tools. Genuinely impressive. Genuinely rare. It's exactly the kind of proof point that gets an entire industry talking, and it should.
However, for me, what gets proven by a project like that is how far the barrier to building has fallen, not that buying is now pointless. A two week build is a brilliant demonstration of capability. It's a different thing entirely from a system a firm runs its conflict checks through for three years, plugged into a DMS, patched every time a model updates, with someone accountable to a client if it gets something wrong. Or, in Correm's case an in-house operating system teams can trust is fit for purpose.
Those are completely different claims wearing the same outfit. “Look what one person can build now” and “you don't need to buy anything anymore” sound similar. They are not the same sentence. The first one is true and worth celebrating. The second one, at best, only applies to those who can build and are interested enough in doing so.
The market's own numbers back this up, brutally. The vast majority of AI wrapper products that get built never make meaningful revenue, and only a tiny fraction of them turn into anything durable. Which tells you the hard part was never typing the code. It was everything that happens after launch day, when the demo has to become infrastructure someone can actually depend on.
Fine, but devaluation still doesn't mean worthless
There's a strain of this argument that goes, why would you pay for something you could build yourself. I heard someone say almost exactly that on a podcast, completely straight-faced.
Sure. For that person. For the person who already has the skill, the time, and genuinely enjoys the tinkering. That's maybe less than a thousand people in the entire legal industry.
I was going to make a joke here about growing your own vegetables. Something like, sure, vibe code your legal stack, but by that logic also grow your own wheat. But that's not quite fair, because farming hasn't all of a sudden become easy to do. Coding just had its entire skill barrier deliberately kicked down. So, to fix the comparison, plenty of people grow tomatoes on a balcony and every single one of them still does a weekly shop. Being able to grow one thing yourself has never meant you stop buying anything. It just means you've got a hobby and a supplier, at the same time, for different reasons.
Nobody's out here writing about how Ocado has no reason to exist because photosynthesis is free.
So why does this argument show up, aggressively, for legal AI
This is the part I actually think is interesting, and to be honest, it's a theory, not a diagnosis. It has to be a theory as you'd never get good enough data to prove it.
Everything about this profession used to run on one currency. Hard work. Years earned. A door that didn't open until you'd put in the time, and that was sort of the whole point of it. Nobody handed you anything. You clawed your way to being trusted.
Then things like ChatGPT, Harvey or Legora shows up and the reaction isn't really “does this tool work.” It's closer to, who said you could do this to us. Who gave a piece of software the right to touch something we spent a decade earning the right to touch.
I don't think that's the whole explanation. There are real, specific complaints in here too. Hallucination risk. Pricing nobody can pin down. Being sold to by people who don't appear to have the legal experience to dominate the market. Those are legitimate concerns and I'm not waving them away. I think it just seems unfair because the industry has, arguably, never been shaken like this before. But underneath the specific complaints, I think there's a second thing humming in the background, and it's not really about moats or margins at all. It's about status. It's about a profession that built its entire identity on being hard to get into, watching the barrier to entry somewhat collapse.
Wrapper hate, in that light, isn't really an economic argument. It's a profession working out loud, in public, what it's worth now that the door isn't as heavy as it used to be.
The one part of the criticism that actually holds up
Strip away the status anxiety and there's a real question left standing, and I want to give it its due, because it's the only version of this argument I'd make myself.
If the entire moat is a thin layer sitting on top of a foundation model, then yes, that moat is genuinely fragile. But that was never really the moat, for anyone who's built one properly. The moat was always the mess underneath. Getting the actual tacit knowledge that sits in a senior partner's head, the stuff that never made it into any training set because nobody ever wrote it down, and turning it into something a system can use. Getting deep enough into how a team actually works that ripping the tool out would be more painful than keeping it.
Which raises the obvious follow-up. If that's true, and it's really just about money and effort, why hasn't a firm with real capital already built it internally and eaten the market? Kirkland just put half a billion dollars behind exactly that bet.
I think the honest answer is that having the money was never really the constraint. I think the legal industry as a whole is quite bad at productising knowledge. Otherwise when the first law firm was approached by Harvey/Legora, they would have just done it themselves, they had access tot he same foundation models. That's not a knock on our industry. It's just a different skill to lawyering, and it's the whole reason infrastructure gets built by people who spend all day thinking about nothing else.
But what happens when the model just does it for free
Now this is a real and fair question. If GPT-6 or whatever comes next can already draft, review, and issue-spot as well as Harvey out of the box, why would anyone keep paying for the wrapper on top of it. That's not a hypothetical fear. Plenty of thin AI apps in other categories have already been swallowed exactly this way, the moment the model shipped their one feature natively.
But legal isn't that category, and the reason is a portfolio problem, not a capability problem. A foundation lab has one roadmap and hundreds of verticals it could point it at. Medicine, legal, finance, insurance, customer service, coding, logistics. Going deep enough into legal specifically to actually replace the vertical layer means building conflict-of-interest logic, jurisdiction-specific privilege rules, matter management, DMS integrations, audit trails a regulator would sign off on, the enterprise trust stack, and doing all of it to a standard a General Counsel would or a firm would trust. That's not a model problem. That's a full time company. OpenAI and Anthropic aren't going to become that company for legal any more than AWS became a CRM company just because it technically could run one. The incentive at the model layer always points one level more general, not one level more vertical, because generality is what serves every other industry at the same time.
So the model getting better doesn't remove the need for the wrapper. It raises the bar the wrapper has to clear. What's actually durable is everything that isn't the model at all. The workflow it sits inside. The accountability wrapped around it. The tacit knowledge nobody wrote down. The fact that ripping it out breaks something that already works. If the model catches up on raw capability and the wrapper has nothing left but that capability, it deserves to disappear. If it's built the mess underneath properly, the model catching up barely touches it.
Where I land
The wrapper isn't the insult everyone thinks it is. It's just the part of the stack that's visible. The judgment still has to sit with someone qualified to carry it. The interface around that judgment being easy to use was never the part anyone should have been ashamed of.
Building it yourself is a genuinely great skill to have. It just isn't a business model for the other ninety nine percent of the profession, and pretending otherwise just makes the argument louder than it is true.
For the vast majority, buying still makes far more sense than building.