In short
- A one-line hosting requirement covers the database and misses files, backups, AI calls and support access.
- Data residency is where data is stored; data sovereignty is whose law and control apply.
- Ask how each limit is enforced, and run the pilot where production will run.
The tender document runs to forty pages. On page twenty-eight, under non-functional requirements, there is one line: "All data must be hosted within the country." Every bidder ticks it. The evaluation moves on to the demos.
Six months after the award, the security team does its first review of the live platform. The database is in the country, as promised. The attachments are in a storage service in another region. The AI features send their requests to a model provider abroad. The vendor's support engineers log in from three countries. Nobody lied in the tender. Nobody was asked.
The requirement was real, and it was written too short to be met.
What goes wrong
Data sovereignty has moved from the security appendix to the main tender in the Gulf and across Europe, for government and regulated buyers above all. The way it is asked has not caught up. Four gaps recur.
One line, many paths. "Hosted in the country" describes where the main database sits. Programme data also travels in files and backups, in notification emails, in exports, in calls to other systems and in AI requests. A requirement written about hosting leaves every other path unstated, so each bidder answers for the part it knows best.
Residency is asked, sovereignty is assumed. Data residency is where the data is stored. Data sovereignty is whose law applies to it and who controls it: who runs the service, who can be compelled to disclose the data, who holds the keys. A tender that asks only about location can be met in full by a service that remains under another jurisdiction's reach.
The AI arrived after the template was written. Most sovereignty clauses were drafted before programme software had AI features. They say nothing about prompts, models or model providers, so an AI feature can send programme text abroad without breaching a single clause.
The answers are yes or no. A compliance matrix with a yes column produces yeses. It does not show how a limit is enforced, or what happens when someone changes a setting.
What good looks like
A buyer does not need to be a security specialist to write a better requirement. Five moves, and a short list of questions.
Ask for a map, not a yes
Ask each bidder to draw where a single status update goes, from the moment an owner saves it: database, files, backups, emails, integrations, AI requests, logs and support access. Each destination should carry its location and the law that applies there. A bidder who cannot draw the map has not checked it.
| Path | Question to ask |
|---|---|
| Database, files, backups | Where is each stored, including backup copies and disaster recovery? |
| AI requests | Which model receives the text, where does it run, under whose key, and is anything kept or used for training? |
| Email and integrations | Which outside services receive data, and can each one be limited or switched off? |
| Support access | Who can reach the live data, from where, under which law, and how is each access recorded? |
| Identity and keys | Do people sign in through our own identity service, and who holds the encryption keys? |
| Exit | In which formats can the whole record leave, and has that been tested? |
Put the AI inside the sovereignty clause
State that AI requests are a data transfer like any other. Ask whether the product can use a model you approve, whether it can use a model running inside your own network, and whether it can work with your own key. If the only answer is the vendor's chosen provider, the AI is outside your sovereignty requirement, whatever the hosting says. The reasoning is in sovereign AI for programme management.
Ask how each limit is enforced
For each path, ask what stops data going somewhere else. A policy document is one answer. A setting that refuses the connection is a better one. Ask what happens if an administrator changes the AI endpoint tomorrow: does the next request fail, or go through?
Run the pilot where production will run
A pilot in the vendor's own environment proves the features and nothing about sovereignty. Run it in the region, cloud or data centre you will use, with sign-in through your identity service and the AI on the model you will be allowed to use. The security review then happens once, during the pilot, rather than again after the award.
Price the exit before you sign
Ask for the export formats in writing, and test one during the pilot. A record you cannot take with you makes the next tender a formality.
Lodestar is hosted wherever a client needs it: on any cloud and in any region, in the client's own data centre, or air-gapped. Its AI runs on the model the client chooses, through the client's own API key or on a model hosted inside the client's own network. Lodestar is ISO/IEC 27001:2022 certified and holds a SOC 2 Type II report, both shared under NDA.
Questions for your next review
- Does our tender ask where data is stored, or also whose law and control apply to it?
- Which of our requirements mention AI requests, models or model providers?
- Have we asked each bidder to show how a limit is enforced, rather than whether it exists?
- Will our pilot run in the environment we will use in production?
A sovereignty requirement written in one line is met in one line. Ask about every path the data takes, and the answers start to differ.
