Insights · Contracts · November 2026
What “you own everything” should mean in an AI contract
The sentence appears in almost every AI proposal written this year, including the one on our own systems page, and on its own it is worth nothing at all. An AI system has five layers and a different party tends to own each one by default, so a promise of ownership only means something once the contract names which layers it covers and what happens to them the day the engagement ends.
Five layers, five different owners
Buyers tend to picture one thing being bought. What is actually assembled is a stack, and the ownership question has a different answer at every level of it.
The integration code is the easy part and almost always ends up yours. Everything above and below it is where the argument lives: your own data and the outputs it produces, the prompts and evaluation sets that took months of tuning, any model weights fine-tuned on your material, and underneath all of it a foundation model that belongs to a company neither of us has a contract with.
Name these six, or the sentence is decoration
Ask for each one by name. A supplier who will put all six in writing is telling you something; a supplier who redrafts around them is telling you more.
The code and configuration. The easy yes. Get it anyway, in writing, including the right to hand it to another supplier.
The prompts. Practitioner checklists flag these as the thing suppliers most often reserve as their own know-how, which is understandable, because on a mature system the prompt library is most of the work. It should still be yours.
The evaluation sets. Almost nobody asks. Your evaluation set encodes what good output looks like in your business, it is expensive to rebuild, and without it your next supplier is guessing.
Weights fine-tuned on your data. Reviews of 2026 agreements report that the default is supplier ownership. If your material trained it, say plainly who holds it afterwards.
Your data. Including the right to have it deleted at termination and returned in a format a machine can read, which rules out a PDF export of your own records.
The outputs. Some agreements grant you a licence to outputs rather than ownership of them. Those are different words and the difference matters when you want to resell, publish or build on what came out.
Four clauses that decide it in practice
Training rights, and the word "improvement". This is the clause to read twice. Language permitting use of your data for "service improvement" can be read to cover training, fine-tuning and derivative models, and several guides report the default is opt-out rather than opt-in, with at least one noting suppliers may price the opt-out as an upgrade. The position worth holding is simple: no training on our material unless we say yes in writing, per use.
Model deprecation notice. One 2026 procurement review describes a pilot built on a model version that was retired without the notice period the buyer thought they had. Put a number on it. Ninety days is reasonable and costs the supplier nothing to agree if they mean it.
Exit, written while everyone is friendly. What is returned, in what format, within how many days, and what is destroyed. Negotiating this after a relationship sours is negotiating from nothing.
Indemnity that reaches the training data. Guides recommend cover for claims that the tool, its outputs or the data it was trained on infringe somebody else's rights. That last one is the part suppliers leave out and hope you do not ask about.
What ownership cannot buy you
Here is the part that belongs in this article because we have to live with it too.
You will never own the foundation model. When we build you a system, it calls a model made by a company we buy from on the same terms as everyone else, and if that company changes its pricing, its behaviour or its availability, your contract with us does not insulate you from any of it. Anyone promising otherwise is selling something they do not have. What a good contract buys you is portability: the prompts, the evaluation sets and the integration code are yours, so swapping the model underneath is an afternoon of work rather than a rebuild.
The second limit is technical and no clause repairs it. Once your material has influenced a set of model weights, removing it is genuinely difficult, which means a deletion right exercised after training has happened is weaker than it reads. Decide about training rights before anything is built. That order is the whole protection.
And the honest caveat on everything above: these are practitioner positions from a market still arguing with itself, not settled law, and none of it is legal advice. Take the list to a solicitor who reads technology contracts for a living. An hour of their time is cheaper than the clause you did not notice.
Read our terms before you need them
Our systems work puts all six in the engagement letter before anything is built: the code, the prompts, the evaluation sets, the configuration, the data and the outputs, with no platform fee, no licence that expires and nothing to log into. Ask us for the ownership clause on its own. You should be able to read it in a minute, and you should ask every other supplier on your list for theirs.