There are only two things left to build
By now, you might have seen ArtCraft, the open source Adobe alternative that a team of four people vibed out with Claude Opus 5.5 (never mind that the insides are mostly AI slop that struggles to do anything complex… that will be solved, in time, probably). The fact that AI is now "good enough" to reverse engineer seven of Adobe's flagship tools is the signal that we have reached a landmark moment.
The AI optimists have been saying for a few years that the value of all software will trend to zero. I recall a conversation with Jacob Beckerman of Macro.com in 2024 where he touted this view, and I recall thinking at the time, "not for a while". Turned out that "a while" was only two short years.
So if AI can vibe out the Adobe Creative Suite, what's actually left to build? To answer that, you have to look at what the AI is really doing.
The intelligent goldfish
The AI harnesses have become so good that it feels like we have either forgotten or ignored how the current generation of LLMs work.
Every large language model is a headless and stateless word predictor. It has no innate awareness of what happened a minute ago or what occurred in a prior conversation. Everything that looks like memory or learning uses the same trick: something is fed into the context, from the top, every single time. Memory is context. Skills are context. The whole conversation is context.
Take the simplest example for lawyers. Ask 20 questions about a contract in one chat, and the model processes that contract 20 times. Caching makes it cheaper, but the work still gets done again. And sure, the answers might get sharper by question 15, but that's not the model learning. That's your first 14 questions sitting in the context, getting fed back in with everything else.
Agents use the same trick with better stagecraft. They look autonomous, like they're running a routine. They are not. The model spits out one word after another until something tells it to stop. Then an orchestrator sits and waits for something external to happen: a tool result, a file, a human. When the agent "continues", it isn't picking up where it left off, the whole thing gets injected again into the same headless engine. Anthropic's own engineers describe every new agent session as starting with no memory of what came before.
We anthropomorphise all this because the experience is so convincing. But an LLM does not work like a human. A human learns on the job. A model doesn't learn at all. It just gets handed better context.
And that's actually good news.
Intelligence isn’t your problem
You and I can’t make the model smarter. That's the labs' job, and they're going to keep doing it.
Intelligence is not a problem for us, in the legal industry, to solve. But, the context is entirely within our control, and it is where we can differentiate. Since every LLM interaction is just giving the model a set of input tokens and waiting for the output tokens, then there are only two things left for us to build:
Skills. Capturing the tacit knowledge in experts' heads and writing it down so a model can use it: how your firm reviews a share purchase agreement, what gets escalated, which positions are negotiable, and what a correct answer actually looks like. In practice, that's a skill.md file. Plain text, loaded into the context when it is needed.
Data. What we feed the model. Retrieval, cleanup, and handling: getting the right text in, keeping the wrong ones out, and making sure one client's documents never leak into another's matter. That includes access management. Either access to better data, or access to systems that create better data.
Put them together, and we have the beginnings of workflows. Everything is about giving the models better context. That's all there is left to build.
So legal tech companies don’t need to build intelligence, and they don’t need to build the harness either, because the labs are giving that away for free.
What about specialized tools?
So, what about all of the legal tech companies building tools?
ArtCraft happened.
If AI can vibe out Photoshop, it can vibe out a clause extractor or document assembly tool. Code is free (or nearly free, given the subsidized token costs).
What remains expensive is correctness. ArtCraft showed that too. The first 90% is now free, and the last 10%, where things break, is where all the value lives.
The only specialized tools still worth building are the ones whose value doesn't come from the code: but something which has been battle tested across thousands of ugly real-world cases, tools that run at scale, tools a court or regulator will accept, and tools that unlock access nobody else has. That means, the only tools that will survive are those which create better skills or better data.
