By Stephen Costigan, Founder, AtlasAI
I've never seen technology move this fast. ChatGPT came out at the end of 2022, and since then we've gone from a chat window that could draft an email to agents that plan, write and ship production software with very little supervision. This is the most basic AI will ever be again. I said as much to a friend from the financial technology world last week, and we spent the rest of the call on what it means for every company that sells software.
Building software is no longer the hard part
Our engineering teams build with agents every day. From that experience, I think a small team using agent-based development patterns could build a working database engine, with its own replication and caching, in about two weeks. A few years ago that was a multi-year program for a well-funded company.
That puts pressure on a lot of vendors. Database providers are in real trouble. Businesses already running on licensed databases will probably keep paying for a while, because migrations hurt, but the software itself is no longer scarce. The expensive managed services on the big cloud platforms, like caching layers, are in the same position. I don't expect companies to build their own storage or web servers, because the economics there still favor the providers. The higher-margin services on top are another matter. The providers will either find a way to keep those workloads on their platforms or lose share to customers who build their own.
Take a harder example. You could build a Bloomberg terminal today. The software isn't the obstacle. The obstacle is access to the market data syndication that makes the terminal worth paying for. Anyone can now copy the software. Nobody can copy the data.
For a law firm, the data is its documents, matters and judgment
Apply that to legal. The feature set of a legal AI tool gets easier to replicate every quarter. Search, chat, drafting, clause review and summarization are converging across every vendor. What no vendor can replicate is a firm's own data: twenty years of precedent in the document management system, the matter history, the negotiated positions, the client preferences, and the way the firm's best lawyers reason through a problem.
If that data is the firm's lasting advantage, then the question that matters most is where it goes when the firm adopts AI. When the AI runs on a vendor's shared platform, the firm's most valuable asset flows into infrastructure the firm doesn't control. When the platform runs inside the firm's own environment, the firm keeps the asset and gains the capability.
Why AtlasAI runs in the firm's own tenant
That's why AtlasAI is built the way it is. It can run as a single tenant inside the firm's own Azure environment. Documents stay where they already live, including in iManage, and AtlasAI reads only what each person is already allowed to read. It builds a knowledge graph of how the firm's matters, documents and people relate, and puts search, chat, drafting and review on top of it. The firm holds the keys and the audit trail.
The firm also owns the codebase, and that matters more every month for the reason I opened with. When agents can build software this quickly, a firm with its own AtlasAI repository can point its engineers and their agentic tools at that code and extend the platform itself: a new integration, a new review workflow, a practice-specific agent. It doesn't wait in a vendor's roadmap queue, and it doesn't send its data anywhere new to get the result.
Three questions for any legal AI evaluation
Features are becoming a commodity, and the gap between vendors on any single feature will keep shrinking. The questions I'd put at the top of any evaluation are these:
- Where does our data live when we use this, and who can see it?
- Who owns what gets built on top of it?
- Can our own people extend it when we need something the vendor hasn't built?
The firms that answer those three questions well will keep their advantage as software keeps getting cheaper. The firms that don't will find they've handed the one thing competitors couldn't copy to a platform someone else controls.