Who Owns Your Intelligence?

AI is a technology. Intelligence is what it helps to create. Once you separate the two, a vital question emerges. Who ends up owning that intelligence? In public, that question is being argued at the level of nations and the largest technology companies.

More importantly, there is a corporate version of the same question, and it is barely being asked. Your organisation is accumulating something inside someone else’s infrastructure. No one has yet established who owns it. It does not appear on the asset register, in the AI use-case inventory, or in the third-party risk assessment. Yet it compounds every day.

The asset is not the model

The first thing to be clear about is that the AI model your organisation has chosen (ChatGPT, Microsoft Copilot, Claude, or another) is not the thing.

Model access is becoming a commodity. Capability converges, prices fall, providers are swapped. Any organisation that believes its AI advantage rests on which model it licensed is describing a position unlikely to survive the next procurement cycle.

What does not commoditise is everything the organisation builds around the model to make it useful: the standing instructions, the evaluation sets that define what good looks like in the business, the correction history that shows where the system was wrong, the retrieval structure over internal documents, and the workflow knowledge embedded in how the system is actually used, by whom, and at what point in a process.

None of that is the model. All of it is specific to one organisation. It cannot be bought, and a competitor cannot replicate it simply by signing the same contract.

That is the asset. Call it your accumulated intelligence. It is the only part of an enterprise AI deployment that compounds.

Boards are being taught to think about AI as procurement: identify the use case, build the business case, buy, deploy, measure. In that process, the question of ownership never arises, because you own what you bought. But the thing that matters was not bought. It accrued.

Four claims, no resolution

Ask who owns the accumulated intelligence and four answers present themselves, all plausible, none settled.

The organisation. It paid for the work. The intelligence encodes its processes and its judgement. On any ordinary reading of an employment relationship and a software licence, it is the organisation’s.

The individual who built it. Most accumulated intelligence is not built by a team. It is built by one person, over months, through use. Where that person has worked in a personal or departmental account rather than an enterprise instance, the corpus sits under their login. When they leave, the practical question is not who has the stronger legal argument. It is who still has access.

The provider. The intelligence lives in the provider’s infrastructure. What can be exported, in what format, on what notice, and whether anything survives export in usable form is governed by the provider’s terms, not by the customer’s intentions. Enterprise AI exit and portability terms increasingly need to be negotiated expressly, rather than assumed.

Nobody, yet. The category did not exist when most of these contracts were written. Contracts speak to data, intellectual property in documents and code, and confidentiality. An accumulated conversational corpus is none of those cleanly. Silence in a contract is not the same as an answer.

The uncomfortable position is the fourth one, because it is the most common.

Why this is not a data question

Boards are practised at data ownership. It appears in every material contract they have approved for twenty years. The temptation is to file this under the same heading and move on.

That would be a mistake, because the asset here is not the data.

The data is usually the organisation’s, and usually protected. What is at issue is the arrangement of judgement around the data:

• Which questions get asked.
• Which answers were rejected, and why.
• What the system was told to assume.
• What it learned not to do.

That layer is not covered neatly by a data processing agreement, because it is not personal data and is often not the organisation’s data in any conventional sense. It is closer to an accumulated working method than to a dataset.

Trade secret protection is the nearest existing frame, and it depends on the organisation having taken reasonable steps to keep the thing confidential and identified. Most organisations have not identified it. Several have never named it.

The three questions

The argument becomes concrete under three questions. They are the ones I would put to an executive team.

If you changed your AI model provider next quarter, what could you take with you? Not in principle. In practice, in a format that works. If the answer is a chat export nobody can act on, the organisation does not own a portable asset. It owns a dependency.

If the person who built it resigned on Friday, what would leave with them? In most organisations the honest answer is a great deal, and there is no process that would even detect it.

If a regulator asked how a particular decision was reached, could you reconstruct it? This is where the ownership question stops being commercial and becomes a governance question. Under the direction of travel in UK and EU oversight, organisations are increasingly expected to evidence how AI-influenced decisions were reached and who was accountable.

If those three answers are unclear, the organisation has been building an asset it cannot move, cannot protect, and cannot produce on demand.

The gap widens quietly

This does not go wrong on day one. It goes wrong the way governance usually goes wrong.

A capable person starts using a general model for real work. It is useful, so they use it more. The intelligence accumulates because that is what makes it better. No one registers the corpus, because no one thinks of it as a thing. Then it becomes load-bearing. A process depends on it. A client relationship runs through it. A regulated workflow touches it.

At no point does any single step look large enough to require a decision. By the time the question is asked, the honest answer is that the organisation does not know what it has, where it sits, or what it could take with it.

That is the same pattern as agent drift, arriving through a different door.

What this means

The public debate asks whether humanity gets a stake in the wealth intelligence creates. It is a serious question, and not one for a board to answer today (if ever).

However, in a business using AI models for workflow, decision assistance, and decision making, and a plethora of other tasks, the intelligence being produced compounds in value. Boards need to understand what ownership of intelligence means for them and their business.

Most have not, because no one has put it on the agenda as an ownership question. It has been treated as an IT matter, which is how several categories of enterprise risk are often treated, shortly before they stop being IT matters.

What to do

Four things, none of them expensive.

• Name the asset. Until the accumulated intelligence has a name and an owner, no process will touch it.
• Establish where it lives. Enterprise instances or personal accounts. Which provider. Which jurisdiction. Who has access.
• Read the contract for the specific question of what leaves with you, rather than the general question of who owns your data.
• Decide the employment position deliberately, in advance, rather than discovering it during a resignation.

The board does not need to resolve the general ownership of intelligence. It needs to resolve the ownership of its own.