A Third Way

If you're a lawyer in a big firm, there's a high chance your job is going to be impacted by AI.

If you're in charge of procuring AI solutions at that law firm you will probably have considered a couple of different options.

You can buy in a SAAS product, designed for legal use, from someone like Harvey. Or you can get an enterprise licence from Anthropic and use their Claude for Legal plugins. In both cases though, you are paying for a product AND you are feeding your core IP (legal docs and the knowledge of the lawyers on your payroll) into that third party product. The input of this information makes the product you are using more performant – and therefore more valuable. But you don't see the main benefit of this increase in value. You are probably doing things more efficiently in the short term. But in the long term you might be signing your own death warrant. Once these products reach a certain level of competency, who needs you or your firm?

You probably haven't thought about building your own version of Harvey internally as you don't have the skills in-house and startups like that have raised millions to get to where they are. (Harvey has raised $1.22B). This scenario is true across almost all areas of knowledge work. Abridge has been built for doctors, Hebbia for financial analysts, Basis for accountants and Cursor for software developers.

But now it is an option.

There was a watershed moment in late 2025 when general purpose AI models became so good at coding that software development could safely move from using AI to finish lines of code to using it to write whole modules. This allowed developers, who knew what they were doing, to build things an order of magnitude more quickly. Elver was formed in December 2025 to capitalise on this. We realised very quickly that the whole software development cycle would change overnight and we built our own agentic management system (think JIRA for the AI age) to ensure we could maintain methodological rigour when it comes to design, coding, testing, security and compliance in this new way of working. We have called this system Sargasso.

At the start of a project we will spend a lot of time with clients to gather requirements, identify data sources and work out evaluation loops (think forward deployed AI Engineers) but instead of signing off wireframes and getting a designer to work through several iterations in Figma, we now build something that fits the brief very quickly. Something that works, something the client can play with. They give us feedback and then we throw the first thing away and build something better. We do this a few times until we've reached a point where the people using the product are seeing real value – then we switch to making updates and new features rather than building everything from scratch. The whole process is managed in Sargasso.

As product people, we love it. We get feedback much earlier in the cycle and it feels as though we have superpowers. Agents, embedded in our products give our clients their own superpowers but with guard rails. They get a product that looks and behaves like a custom SAAS AI product that had years of prototyping and testing. But they own it and there are no subscription costs.

But the main benefit of building AI products this way is not financial.

The main benefit is how it enables our clients to keep their institutional knowledge in-house. They don't have to share it with Anthropic or Open AI or third party products like Harvey. They get to develop an internal system that draws on all their data and all their experience, but in which they own the IP. And as the system improves they get more productive, in a virtuous cycle.

We were already doing this with a large international client (not a law firm) when Satya Nadella (CEO of Microsoft) published his essay "A frontier without an ecosystem is not stable" in June 2026. He could have been describing the thing we were already selling.

"Companies need to turn their workflows, domain knowledge, and accumulated judgment into AI systems that improve with each use."

He describes the trap we saw coming perfectly. A trap where whole sectors outsource their AI and in doing so guarantee their demise.

He has done us a huge favour. As the term 'human and token capital' move into the lexicon of LinkedIn, we are at the coalface – building.