Glossary
Data residency
Data residency is a commitment about the country or region your data is stored in, offered by a provider so that a buyer with location requirements can meet them without running the software themselves.
In plain terms
Somebody in your organisation needs to be able to say where the material sits. Perhaps a customer asked, perhaps a regulator does, perhaps it is simply policy. Data residency is the vendor's answer: they will keep it in a stated place. It sounds like one fact and is really several, because storing something somewhere is not the same as processing it there, and neither settles whose laws can reach it.
Why it matters
Because it is frequently the first hard constraint in an AI purchase, and unlike most requirements it can rule a product out entirely rather than making it awkward. A tool that cannot meet a location requirement is not a tool you negotiate about. It also arrives early: the question is usually asked before anybody has evaluated whether the product is any good, which means residency shapes the shortlist rather than the final decision, and a team that discovers it late has evaluated the wrong set of products.
How it works
Storage and processing are separate questions and vendors sometimes answer only the first. Material can be stored in one region and processed somewhere else, briefly, as part of answering a request. Both matter, and a commitment that covers storage alone leaves the more interesting half of the question open, so it is worth asking about each explicitly rather than accepting a single reassurance.
Residency is about location; sovereignty is about legal reach, and conflating them is the commonest error. Where the material sits is a fact about geography. Which laws can compel its production is a fact about the provider's corporate structure and about the arrangements between countries, and a provider storing data in your region may still be subject to demands from elsewhere.
AI tools complicate it because more parties are usually involved. A product may run its own service, call a model from another company, and use a further supplier for something else, each with its own locations. The residency question therefore extends down the chain rather than stopping at the vendor you contracted with, which is why the sub-processor list is the practical document to ask for.
Some parts of a product carry the commitment and others do not. Core storage and processing are usually covered; logging, telemetry, support tooling, backups and abuse monitoring often live somewhere else and are frequently outside the headline promise. That is not concealment, it is how services are built, and it is a specific thing to ask about rather than to assume.
It is usually delivered as a regional deployment, and regions are enumerated rather than infinite. A provider offers a set of places and you choose from it; if the place you need is not on the list, no amount of contract drafting produces it. That is why the question belongs at shortlist stage, when the answer can still change which products you are considering.
Running the software yourself is the strongest form and the most expensive one. Self-hosting or an air-gapped deployment removes the question by removing the third party, and substitutes an operating burden. For material that genuinely cannot leave, that trade is often the only one available, and it should be recognised as a trade rather than a victory.
It has to be re-checked rather than established once. Providers add regions, change suppliers, move services and acquire companies, and any of those can alter an arrangement that was accurate when it was signed. Where the requirement is real, the practical habit is to reconfirm at renewal and to ask to be told about changes to the onward provider list, because that is the part most likely to move without anybody mentioning it.
The commitment is only as good as where it is written down. A statement on a marketing page, an answer in a support conversation and a term in a contract are three different things, and only one of them is enforceable. Where residency matters enough to shape a shortlist, it matters enough to appear in the agreement.
What a residency answer needs to cover
Seen in the wild
Ask which regions an assistant can be deployed in on business terms, and whether the answer covers processing as well as storage.
ChatGPTCheck where an enterprise search product holds its index, given the index is a copy of material that already lives in your systems.
GleanRun a model on your own hardware, which answers the question by removing the third party and replacing it with operating work.
OllamaLook at what a routing layer does, since a request passed to another provider inherits that provider's locations rather than yours.
OpenRouter
Common misconceptions
People assume
They store it in our region, so we are covered.
In fact
Storage location is one of several questions. Processing may happen elsewhere, telemetry and support tooling often do, and legal reach follows the provider's structure rather than the data centre's address. A single reassuring answer usually addresses one of these and is offered in good faith about that one.
People assume
The cloud region we chose settles it for everything.
In fact
It settles it for the parts your provider runs in that region. Where a product calls another company's model or uses further suppliers, those have their own arrangements, and the commitment you hold with your vendor does not automatically extend down the chain unless it says so.
People assume
If we never upload anything, residency does not apply to us.
In fact
Anything the system is given is material that went somewhere, and that includes what people type. A question containing a customer name, a pasted paragraph of a contract or a description of an incident is content, and it travels the same path as an attached file. Residency applies to the conversation, not only to uploads.
People assume
This is a technical question for the platform team.
In fact
It is a contractual question with a technical component. What is deliverable is technical; what is promised, to whom, and with what consequence if it is not met is contractual. Teams that treat it as purely technical often end up with an arrangement that works and cannot be evidenced to anybody who asks.
Telling them apart
Data residency vs Data sovereignty
Data residency
Where the material is physically stored, and often where it is processed.
Whose laws can reach it, which follows the provider's structure as well as the location.
One is answered with a place; the other is answered with a jurisdiction and a corporate structure.
Questions
- What is the most useful thing to ask a vendor?
- Ask about storage and processing separately, then ask which parts of the service are not covered by the commitment. Logging, backups, support tooling and abuse monitoring frequently sit outside it. That question surfaces the gaps a general reassurance conceals, and it does so without needing you to know how the service is built.
- Does residency mean our data is safe?
- It means it is in a stated place, which is a different claim from being secure or being beyond anybody's reach. Security is about controls, sovereignty is about legal jurisdiction, and residency is about geography. All three may matter to you and each needs asking about separately, because a strong answer on one is often offered in response to another.
- When should the question be asked?
- At shortlist stage, before anybody evaluates quality. Regions are enumerated rather than negotiable, so a product that cannot deliver the place you need is out regardless of how well it performs. Teams that ask late have often spent their evaluation effort on a set of products that was never viable.
- Does self-hosting solve it?
- It removes the third-party question and replaces it with an operating one: hardware, updates, security and somebody to run it. For material that genuinely cannot leave your control, that is frequently the only available answer. It is a trade rather than a simplification, and pricing the operating side honestly is part of making it.
- Who inside the organisation should own the question?
- Whoever owns data protection, with somebody technical to say what is deliverable and somebody commercial to get it written down. It falls between roles more often than most requirements, which is how an arrangement that genuinely works ends up impossible to evidence when a customer or an auditor asks for it in writing.
- How does this differ for AI tools specifically?
- More parties are usually involved and more material moves. A product may call another company's model, so the chain extends past the vendor you contracted with, and AI tools tend to receive whole documents rather than structured fields. Both raise the stakes on a question that ordinary software often answered with one region name.
Key takeaways
- Storage and processing are separate commitments; ask about each.
- Residency is geography, sovereignty is legal reach, and they are routinely conflated.
- Logging, backups and support tooling often sit outside the headline promise.
- The chain extends to any provider your vendor calls, so ask for the sub-processor list.
- Ask at shortlist stage: regions are enumerated, so it decides which products are viable.
Last checked July 2026