Posted by Ziyaad Ahmed, co-founder of Legal Futures Associate Qanooni AI
Ahmed: You cannot rent institutional knowledge
Something has changed for me in the last 12 months. Not the thesis I have been building on for four years. What has changed is how I answer when senior partners ask me which legal AI they should buy.
Four years ago, I would have walked them through a vendor comparison. Now I tell them the question itself is wrong.
The reason is straightforward. There are now more than 400 legal AI vendors in the market, with new entrants arriving every quarter. Almost every senior partner I speak with is currently trying to make the same decision, and almost none of them can.
They have four vendors in trial. Two more on next quarter’s list. Every pitch is competent. Every demo is polished. Every deck has a slide on hallucination controls and a slide on integration with the tools the firm already uses.
And at the end of every meeting, the same question keeps surfacing. Which of these will still be the right choice in 12 months?
Sarah Blair, in her recent Nexian sessions, has been documenting what any of us in the market can already feel: the senior legal buyer, the person who signs off on strategic technology decisions inside a firm, has stopped being able to tell the vendors apart at the point of decision.
This is not, in my view, another software cycle. In every previous procurement decision a partner has ever made, the underlying technology was stable enough that the vendor comparison was the whole exercise.
Word processors, practice management systems, document assembly – you picked the best implementation and you were done.
AI does not work that way. The frontier model that powers a vendor today is not the frontier model that will power it 18 months from now. The interface that impresses in a demo this quarter will look dated by the next release cycle. The layer of the stack that most vendors are competing on is the layer that is going to churn hardest.
Which is why the question is not which vendor. It is which layer.
The three layers
Here is how I have come to see it. Legal AI in 2026 sits in three layers.
The model layer is the foundation models themselves. GPT, Claude, Gemini, the next one nobody has released yet. This layer is rented.
Whichever model is best this quarter will not be best next quarter, and the firms buying into it will pay a subscription to swap between them for as long as they use the technology. There is nothing wrong with that.
But rented capital is rented capital. No firm ever built an enduring advantage by paying rent.
The product layer sits directly on top. A specialist interface tuned for legal work, sold as a capability gain.
This is where the majority of the current legal AI market lives. It is a real business and some of the vendors in it are building good products.
But this layer is replaced every two years or so. When the underlying model changes, or when a competitor builds a better wrapper, the product becomes a candidate for a rip-and-replace decision. The firms buying into it are not accumulating anything durable. They are paying for time saved on the next matter.
Above both sits the firm layer. This is where very few vendors are actually building. It is where the firm’s playbooks, its matter context and its partners’ judgment become a persistent, addressable representation the AI system reasons over. It is where the firm position stops being tacit and becomes IP.
Owned, not rented. Persistent, not swapped.
The counterintuitive part is that the firms that stop chasing the best model tend to end up with the best AI. Because the value the AI actually delivers is compounded not by the model behind it, but by the firm’s institutional knowledge the AI has learned to reason over.
You can rent a model. You can replace a product. You cannot rent institutional knowledge.
Under the hood, this means something specific. Every decision your partners make on a matter is captured to the firm’s knowledge graph and redeployed on the next matter, inside the tools they already use.
The system knows how your firm approaches a shareholder agreement. It knows which precedent you use for a particular clause. It knows which redlines your senior partner accepts and which ones they push back on.
When the underlying model changes, none of that changes. Your firm’s reasoning stays yours.
The test worth running before your next vendor pitch
The way I have learned to tell whether the AI you are being sold sits at the firm layer or the product layer is a single question.
Before your next vendor pitch, ask this: If we swap the model behind your system in 18 months, what of our institutional knowledge persists?
If the honest answer is a folder of prompts and outputs, you are looking at a product layer tool. If the honest answer is a persistent representation of the firm’s matters, playbooks and reasoning, you are looking at a firm layer tool.
Most current legal AI, when pushed on that question, gives an answer somewhere near the first. Most current legal AI is not a firm layer tool.
This is the same lesson I have watched the profession learn once already. In the 2000s and 2010s, legal process outsourcing (LPO) looked like a straightforward margin uplift for firms. Route the process work to an LPO provider, take the efficiency gain, keep the client relationship.
Two decades later, the firms that lost the process to LPO providers had also lost the margin, because the providers had become the ones who understood the client’s operations better than the firms did.
If your firm’s reasoning lives in someone else’s system, someone else eventually owns the value it produces.
Four years in
Natalie Foster runs Inspire Legal Group. She was one of the earliest to build on a matter-first system.
When I sat with her earlier this year, she said something that has stayed with me. She said they had stopped thinking of the platform as a tool. It is where their firm runs.
That is what a firm layer decision looks like 18 months in. The AI is not a capability add-on that competes with other capability add-ons. It is the infrastructure the firm reasons on.
The returns Inspire Legal Group is reporting back on the platform are approximately 50 times its cost. That number is not the result of a single feature or a better model. It is the result of a system that holds the matter as persistent state across the tools the team already runs on, and that gets sharper every time they close a file.
Four years ago, when the market was chasing the best model and the flashiest interface, we made a different bet. We bet that the model would commoditise. We bet that the interface layer would be replaced every couple of years by whoever built the next best wrapper.
And we bet that the layer where value would actually compound was the one closest to the firm’s own thinking.
So we built QMatters around the matter, not the document, and not the prompt. Every partner decision on every matter is captured into the firm’s graph, indexed, and made available to the next. The model underneath can change. The firm’s reasoning stays with the firm.
Four years in, the bet has aged well.