You Can Rent a Model. You Cannot Rent Case Law Discipline.

Ziyaad Ahmed of Qanooni AI argued in Legal Futures this week that legal AI has three layers: model, product, firm. Only the firm layer, he says, holds durable value. He is right about the model layer and the product layer. He is also right that firms should look past vendor pitches and ask what persists. Here is what his three-layer model does not quite capture, and why it matters for claimant firms picking specialist tools.

Ziyaad Ahmed's Legal Futures piece on 9 July is worth reading. His three-layer framing of legal AI (model, product, firm) is a clean way to make sense of a market that currently has more than 400 legal AI vendors in it, most of them pitching senior partners on features that will change again in eighteen months.

His central argument is simple. The model layer is rented. The frontier model that powers a vendor today will not be the frontier model that powers it 18 months from now, so buying at that layer is buying a subscription to swap between models for as long as you use the technology. The product layer is churned. The specialist interface tuned for legal work will be replaced when the underlying model changes or when a competitor builds a better wrapper. That leaves the firm layer as the one place where value compounds: the firm's playbooks, matter context and partner judgment captured as a persistent, addressable representation the AI reasons over.

Ahmed's summary line lands. You can rent a model. You can replace a product. You cannot rent institutional knowledge.

We agree with almost all of it. This is not a rebuttal.

What we want to add is a fourth observation his three layers do not quite capture. For claimant firm work, there is a durable asset that is neither a rented model nor a firm-specific knowledge graph. It is court-anchored discipline. Verified case law, current CPR practice, established evidence standards, and the non-hallucination rules that keep those anchors honest. That discipline sits as a layer of its own. It is worth understanding why, and what it means for how claimant firms should think about buying legal AI.

Where Ahmed's model works cleanly

For a large commercial firm advising on a bespoke shareholder agreement, Ahmed's argument holds without qualification. The value is in the firm's institutional read of the client, the sector, the counterparty, and the specific commercial context. That is not written down anywhere the AI can access without the firm building it. It sits in partner heads and grows every time a matter closes. If the firm captures that into a persistent representation the AI can reason over, the firm builds an asset. If it does not, the firm rents time-saved-on-next-matter until the vendor churns.

Ahmed's Qanooni is built to sit at that firm layer. Fair enough.

For the firms he is writing to, this is the right analysis. The move from renting a product to building a firm layer is where the durable value is, and any partner making a strategic AI investment decision should read the Legal Futures piece before signing the next enterprise contract.

Where the three-layer model runs into claimant work

Claimant firm work is not bespoke commercial advisory. It is protocol-driven, high-volume, and anchored to court practice.

A housing disrepair claim runs on a pattern the courts have already established. A credit hire challenge stands or falls against a case law library that has been evolving in a defined way for two decades. A clinical negligence chronology has to meet standards that are set by CPR, the courts and the profession, not by the firm's individual read of the matter. A witness statement is subject to Part 32 and PD 32 whether the firm has captured its institutional knowledge or not.

The durable asset for that work is not primarily the firm's individual reasoning about a specific claimant. It is the discipline that ensures every case follows the correct legal framework, cites the right authorities, meets the right evidence standards, and does not fabricate. That discipline is court-anchored. It exists in published law and in professional practice, not in a firm-specific knowledge graph.

Two things follow.

First, the discipline is genuinely durable. A model swap does not degrade it. A product interface change does not degrade it. It survives churn for the same reason a well-maintained case law library survives churn: it is anchored to something outside the vendor. The published law does not move when the vendor changes hands.

Second, the discipline is not something individual claimant firms should be trying to build themselves. Curating a UK case law library that is current across credit hire, RTA, clinical negligence, housing disrepair and public liability is a full-time legal-knowledge job. Keeping non-hallucination rules current as models evolve is a full-time engineering job. Building CPR compliance into the drafting layer is a full-time procedural job. Small and mid-sized claimant firms should not be trying to be a legal knowledge engineering house alongside their fee-earning work. They should buy a specialist tool that maintains the discipline for them.

That is where Ahmed's three-layer model has a gap. Between the product layer (churned) and the firm layer (owned), there is a specialist layer where the durable asset is court-anchored discipline rather than firm-specific institutional knowledge. For claimant firm work, that layer is where the compounding value lives.

What court-anchored discipline actually is

Three components, all of them concrete.

A curated, verified case law knowledge base. Every argument the tool drafts is grounded in cases it can actually cite. The knowledge base is maintained: new judgments are read, added, cross-referenced with existing authorities, and dated to the point in time they became authority. The system refuses to invent cases and refuses to overstate authority. When the underlying model changes, the case law knowledge base persists. It is not the model's output. It is the tool's input.

A hard non-hallucination rule. The tool does not invent defects, exhibits, dates, heads of loss, defendants or rates. If the case facts do not support a proposition, the tool refuses rather than fabricates. This is a discipline that operates at the drafting layer, not at the model layer. When the model changes, the rule stays.

Embedded CPR discipline. Witness statements meet PD 32. Rate arguments follow the GTA and the current case law. Correspondence with defendants follows the pre-action protocol and the CPR rules on evidence. The tool encodes these rules in its drafting logic, not in its model prompt. When the model changes, the rules stay.

Together, these three make up an asset that is genuinely durable. It is not the firm's institutional knowledge, and it should not be. It is the specialist knowledge base that all claimant firms need but no individual firm is best placed to build.

What this means when a partner is picking legal AI for a claimant practice

Ahmed's advice was aimed at senior partners of large firms considering enterprise AI buys. His test question, the one he suggests asking any vendor before the pitch, is whether any of the firm's institutional knowledge would persist if the model behind the system were swapped in 18 months.

For a claimant firm, the question needs a small addition. If we swap the model, what of the specialist knowledge base persists, and how do we know the tool is grounded in it?

The answer to the persistence question should be about the curated case law knowledge base, the non-hallucination rule and the encoded CPR discipline. Those persist. If the vendor's answer is "our model has been trained on a lot of law", that is not the same thing. Training is not curation. A curated knowledge base is something the tool references at run-time, verifiable and updatable. A trained model has absorbed law it may or may not still cite correctly six months later.

The answer to the grounding question should be that every drafted output can be traced back to a source the fee-earner can point to. If the vendor cannot show that trace, the tool is at the product layer, not the specialist layer, and the discipline is not durable.

We have been building LegalDocs Assist to that specialist layer. Every argument the tool drafts is grounded in verified UK case law. The system does not invent. The drafting logic embeds Part 32 discipline, current CPR practice and the practice-area-specific chronology rules for housing disrepair, RTA, clinical negligence, employers' liability and public liability. When the underlying model changes, the discipline stays. That is what we mean when we say the value at the specialist layer compounds.

The wider read

Ahmed's warning about legal process outsourcing is worth taking seriously. His point, that if the firm's reasoning lives in someone else's system the someone else eventually owns the value, is real for firms whose competitive edge is in bespoke reasoning.

But it is not the whole picture. There is a class of legal work where the reasoning is not bespoke, the durable knowledge is court-anchored, and the right specialist tool is one that has invested in the case law and CPR discipline the firm should not be trying to duplicate. Buying that specialist tool does not transfer value out of the firm. It transfers non-fee-earning legal knowledge work to somebody who does it full time, and it gives the firm's fee-earners a tool they can defend on the record.

You can rent a model. You can replace a product. You cannot rent institutional knowledge. And you cannot rent case law discipline either. But you can buy it from a specialist that maintains it, and for volume claimant work, that is the right build strategy.

© LegalDocs Assist — www.legaldocs-assist.co.uk