What You Actually Own When You Commission An AI System

When an organisation commissions a bespoke AI system, the contract usually specifies who owns the code. That is one of at least six things that change hands, and it is rarely the one that determines whether the organisation is genuinely independent at the end.

The six, named

Each of these has an owner whether or not anyone wrote it down. Where it is not written down, the default tends to favour the supplier, not because of bad faith but because the supplier is the party holding the artefact.

1. The application

The software built to the specification. This is the part contracts handle well and the part that matters least in isolation, because an application without the layer beneath it is not independently operable.

2. The model

Whatever performs the inference. If it is an external API, nobody in the room owns it and the commercial terms can change under both parties. If it runs locally, the question becomes whether the organisation may continue running it if the relationship ends, and on what terms.

3. The tuned knowledge

The most commonly overlooked item. When a system is grounded on an organisation's records, indexes, embeddings and retrieval structures, that derived layer encodes the organisation's own knowledge in a form it did not previously possess. It is genuinely valuable and frequently unaddressed, and the organisation should hold it.

4. The training and grounding data

The source records were the organisation's to begin with, which makes ownership feel settled. The live questions are different: was any of it copied outside the perimeter during development, does the supplier retain a copy for support or improvement, and what obligations survive termination.

5. The audit trail

The record of what the system did. Under most regimes this is the organisation's evidence and its retention obligation, so it needs to be held by the organisation and verifiable without the supplier's participation. An audit trail only the vendor can interpret is not evidence, it is a dependency.

6. The right to leave

Not an asset, but it behaves like one. If continuing to operate requires a licence that can be withdrawn, a key that must be renewed by a third party, or an endpoint that can be switched off, then whatever the other five clauses say, the organisation is a tenant.

Where the line should sit

Suppliers who build on their own platform have a legitimate interest in keeping the platform. The workable split is not complicated, and stating it plainly at the outset removes most of the friction later: the organisation owns the system built for it and the knowledge layer built on its data, while the underlying platform and its intellectual property remain the supplier's and are licensed to the organisation.

What makes that split acceptable rather than merely tidy is what the licence permits. A licence that allows continued operation on the organisation's own hardware, offline, without a call home, is materially different from one that requires periodic contact with a vendor server. The first is ownership with a supplier attached. The second is a subscription wearing a bespoke coat.

“If the system stops working when the supplier does, the organisation did not commission a system. It rented one with extra steps.”

Questions worth asking before signing

These are unglamorous and they are answered quickly by a supplier who has thought about them. Hesitation on any of them is itself informative.

• If we end this relationship, what specifically stops working, and on what timescale?
• Who holds the indexes and embeddings built on our records, and can we export them in a usable form?
• Does any of our data leave our infrastructure at any point, including during development and support?
• Can we verify the audit record without your participation, and can you demonstrate that today?
• Does the system require network contact with you to keep running, and what happens if that contact is refused?
• What survives termination: which licences, which obligations, which copies?

Why this is worth doing early

None of these questions is difficult at the scoping stage and all of them are difficult eighteen months in, when the system is load bearing and the answers have consequences. The organisations that end up genuinely independent are not the ones that negotiated hardest. They are the ones that named all six items before the work started, while both parties could still choose a different arrangement without anything breaking.