Harvey is training their own model. Here's the question your firm should be asking instead.

Harvey announced they're training the first in a series of legal foundation models. They're hiring across post-training, agents, and the whole data stack.

If you read that news the way most of the legal world will read it, you'll feel reassured. The smartest, best-funded company in legal AI is now building a model specifically for legal work, possibly trained on documents from firms like yours.

Before you celebrate, let me say the part nobody wants to say out loud.

You don't own that intelligence

It does not matter whether your firm picks Harvey, Legora, or whatever well-funded name lands six months from now. And it does not matter that the model may be fine-tuned on your firm's own documents. If the model lives on the vendor's cloud, runs on the vendor's stack, and belongs to the vendor's corporate entity, then the firm doesn't own the intelligence.

The firm is renting it. The firm is a tenant.

Tenants lose everything two ways

The landlord changes the locks.

The Anthropic moment earlier this year was the preview. Fable 5 and Mythos 5 were switched off by government directive. The company disagreed with the directive. The models still went away. The lucky part was that the other models on the platform kept running. The unlucky version — where the model your firm is built around is the one named in the directive — is a few words away.

The landlord disappears.

Spirit Airlines ceased operations last month and is in liquidation right now. "Too big to fail" retired a decade ago. The legal-AI category has burned through billions in venture funding in the last eighteen months. Some of these companies will still be standing in five years. Most won't.

If your firm's most valuable knowledge layer lives entirely on a vendor's stack, your continuity is a function of their next funding round.

The real question

The legal industry spends most of its AI conversation arguing about which tool is smartest, which agent is most useful, which RAG pipeline is most accurate. Those are real questions. They are also the wrong ones to be asking first.

The first question is this:

When the tool, the cloud, or the company is gone — by directive, by acquisition, by bankruptcy, by anything — what walks out the door, and what stays inside the firm?

What ownership actually looks like

A version of legal AI that is genuinely owned by the firm has a specific shape. It runs in the firm's environment. It learns from the firm's matters and no one else's. It sits in a graph the firm owns — a curated ontology of clients, matters, parties, clauses, obligations, and defined terms — and that graph is exposed through APIs and SDKs the firm controls.

Any agent — Claude, GPT, open-source, in-house — can connect to that graph. Any application — drafting tools, conflicts assistants, intake bots, partner dashboards — can be built on top of it.

The vendor's job in that picture is not to own the intelligence. The vendor's job is to deliver the system, hand over the keys, and let the firm own what it builds. When the firm decides to swap the model underneath — or the vendor — the curated ontology and the data underneath it are not portable. They are already at home.

That is what owning your intelligence actually looks like.

A smarter tenant is still a tenant

Most of the market is not built that way. Most of the market is running a race to be the smartest tenant. The pitch is always the same — "we'll train a model just for you, on our stack, on our terms" — and the more sophisticated the pitch gets, the more the firm pays for the privilege of being a more important tenant.

But a smarter tenant is still a tenant.

Training a model on someone else's platform is not the frontier. The frontier is intelligence the firm actually owns — the kind no directive can disable and no bankruptcy can take with it.

Rent the model. Own the memory.

The model layer is going to keep moving. It should. The frontier of capability is shifting every quarter and that is healthy. The firm's posture toward the model layer should be one of swappable utility — rent it, use it, change it when something better comes along.

The memory layer — the curated graph of how this firm thinks, what it has done, who its clients are, which clauses survived the negotiation, what each defined term means in the firm's universe — that layer compounds. It is the asset. It is the moat. It is the only part that is genuinely yours.

Rent the model. Own the memory.

There is a version of this where ownership goes back where it belongs. I have thoughts on how. If that's the conversation you want to have, my DMs are open.