Law Is the Test Case: Why AI Will Eat Every Wrapper and What Survives

On 29 October 2026, a live database of AI hallucination cases listed 2,241 court decisions identified so far. Of those, 884 involved lawyers. The database is kept by Damien Charlotin and was last updated that same day, so 2,241 is the count for 29 October and no other date.

Imagine a clerk at a registry desk in the early afternoon, a stack of filings at her elbow, a kettle going cold, and a browser tab she refreshes the way other people check the weather. Each time she looks, the figure has crept up. She is not a villain and she is not a prophet. She is simply the first person in the building to understand that a case name that does not exist is not a typo. It is a small, confident lie with a court's stamp waiting to be put on it.

I am a founder in Nairobi who builds AI products, and I have been thinking about a clerk like her for a month. She is a composite I invented to hold the number. Her registry is the best laboratory I know for the most important claim in my own business. So I will use law as the test case. I will tell you where I think I am right, where I am only hoping, and what you should ask the next AI salesman who walks into your office.

A Conviction, Labelled as One

Here is my bet, stated plainly. Frontier models are converging. Within months of one lab shipping a capability, others ship something close to it. I believe prices keep falling, though this essay will not put a number on that. Some strong models arrive with open weights that anyone can run. If that is right, the model becomes an input everybody can buy, and a thin product, mainly a prompt, a chat window and a monthly subscription, is a wrapper. Wrappers get eaten.

What stays valuable is what a model cannot download: proprietary data and the customer relationship. By relationship I mean distribution, trust, the workflow a firm has built around you, and the liability a company will carry for what your software produces.

This is the founder's conviction. It is not a fact. The rest of this essay tries to separate the two.

The Law Is the Test Case

Why law? A legal citation is either real or it is not, which makes the output checkable in a way a poem or a marketing deck is not. The cost of a wrong answer lands on a client's liberty, a client's money, or a judge's patience. And the incumbents who own the content have the most to lose. In my reading, they have also been moving fastest, as the Thomson Reuters and LexisNexis releases below suggest.

Consider Thomson Reuters. It launched general availability of the next generation of CoCounsel Legal on 20 August 2026, US only at launch. The release says the product "Built on Anthropic's Claude Agent SDK." Read that sentence twice. The company that owns the Westlaw content built its agent layer on a model vendor's toolkit. The model is a component. The content and the workflow are the house.

LexisNexis went the same way. On 24 August 2026 it announced agentic capabilities in Lexis+ with Protégé, and its release says the product "applies the best-fit supported model by default while allowing professionals to select a preferred model." A product that lets you swap the engine is the strongest version of my argument, because it concedes that the engine is replaceable. The open question is where the value sits.

Startups are raising money on the same bet. In October, Legora announced a $550 million Series E at an $8.8 billion valuation, led by Atreides Management of Boston.

A valuation is a price, not a moat.

Where the Thesis Holds: Data, Trust and the Invoice

My side has less measured evidence than I would like. The case I can make rests on where errors are caught, not on how often they occur.

My reading, which is judgment and not a finding of any study, is that whatever separates a better legal tool from a worse one probably lives in the layer around the model: which content it retrieves, how answers are grounded, and what checks sit between the draft and the filing. That layer is where data and product discipline live. The model is only one part of it.

The database has a limit of its own. It counts decisions identified so far, and a decision is not the same as a lawyer. The count does not tell you how many lawyers were careless, how many were sanctioned, or how many were simply unlucky with a tool. Treat the figure as a live tally that will keep moving.

Now the invoice. I believe the hourly invoice has long been the legal industry's quiet business model. Suppose a model drafts in four minutes what used to take four hours. The invoice then becomes a confession.

Here is the uncomfortable part. My reading, which is judgment and not a legal finding, is that a relationship in law is less a marketing asset than a compliance system. The part a client or a court can actually inspect is the firm's verification habits, not the vendor's brochure.

The Steelman: The Wrapper With the Best Hands Can Win

The opposing view deserves its strongest form, and the strongest form comes from inside the incumbents.

LexisNexis Legal CEO Sean Fitzpatrick, in the same announcement that introduced Protégé's agentic features, put the wrapper argument better than I can: "Legal professionals should not have to spend their time managing AI, switching among tools or repeatedly explaining how they want work performed." That is a claim about friction, not about models. A thin product that understands how a litigator works, and removes the clicks between a question and a usable draft, can win an afternoon even when the engine under it is the one the next product uses.

I believe this. Workflow is real. A product that feels like a colleague is worth more than one that feels like a search box. My answer is that good design is a wrapper's best defence, and a competitor with better design can attack that defence within a quarter.

The second steelman is factual. The Stanford AI Index 2026 reports that the top closed model led the top open model by 3.3% as of March 2026, up from 0.5% in August 2024. The report says "the open model performance gap reopened in 2025 after briefly closing in 2024," and that six of the top ten models on the Arena Leaderboard are now closed. That counts directly against my line that open weights will flatten everything. Parity at the top is real, but incomplete.

The third steelman is the one I find most uncomfortable. The incumbents are wrappers too. Thomson Reuters built its agent layer on a third-party SDK, and LexisNexis lets you choose the model. If the thin layer is what gets eaten, the sellers of the thick layer, the content and the trust, survive, and the incumbents already own the content. That argument favours the old companies, which is exactly why a founder should distrust it.

Models will converge. The record of what a client trusted you with will not.

That sentence is my conviction, and I would still defend it. But it does not defend itself. It needs a moat that is not a data set: a standard of verification, and a relationship that survives a model change.

Kenya, at the End of the Pipe

Kenya is where the test is clearest, partly because I live here and partly because the courts here have already had to rule on it.

In April 2026, the Milimani High Court in Nairobi was reported to have dealt with a self-represented litigant accused of filing AI-generated pleadings. The opposing party offered no technical or forensic evidence. The court's position, as reported by allAfrica, was that such allegations "need proof, such as evidence of invented citations or a forensic analysis." The same court held that "technology cannot replace compliance with established court rules," and that self-represented parties are held to the same pleading standards as anyone else.

Imagine the litigant in a corridor after the hearing, saying he read every line himself and checked each one before he signed. That scene is invented. I do not know the man. What the ruling shows is the shape of the problem: the accuser must prove the machine, and the filer must still answer for the page. Both are true at once.

The local legal-tech conversation is the same argument on a smaller scale. Court judgments are published, so the raw text is not scarce, and a competitor can in principle obtain the same judgments. If the text is not the moat, it is probably the curation, the indexing, and the way an advocate comes to trust a tool and the people who maintain it. That part is judgment, and I hold it loosely.

Three Questions Before You Buy

The test I now use has three questions. I would ask them of any AI product, whether it sells to lawyers, doctors, farmers or bankers. If the answers are vague, the product is probably a wrapper.

First, what do you own that the model cannot download? Name the data rights, the corpus, the feedback loop that improves only because your customers use you, and the record of corrections nobody else holds. If the answer is "the model," you are being sold a rental.

Second, who signs when the output is wrong? Find the person who answers for the error, the contract clause that says so, and the step where a human verifies before the filing leaves the building. A product that cannot say who is liable is asking you to carry its risk.

Third, if the model underneath changed tomorrow, what would still work on Monday morning? Look for the workflow your customer built around you, the integrations they cannot easily leave, and the relationships your team has earned. If the answer is only the interface, a rival with better design is one quarter away.

I have not solved the problem these questions raise. Imagine the clerk at the registry desk, the kettle cold beside her, pulling the reporter off the shelf to check one citation by hand, line by line. That scene is invented. No model can download that habit. Models will converge. The hand on the page will not.

Editorial note: This essay expresses the founder’s ideas, developed with AI-assisted research and writing (parallel Claude Haiku 5.5 agents) and checked against the linked sources on 30 October 2026. Convictions, proposals and forecasts are identified in the text; scenes framed as imagined are illustrative.

Cover photograph / illustration: Joe Gratz (Flickr, photo taken 23 March 2006). CC0 1.0 (public domain dedication, no attribution required). Original source. Illustrative image: a courtroom gavel. Photograph: Joe Gratz, CC0 1.0 (public domain dedication), via Wikimedia Commons. This image does not document the events described.