“We’ll host it locally, so we’re PDPL-compliant” is a sentence heard often in AI projects in the Kingdom, and it is misleading. Local hosting matters. For a system that handles personal data, it is often the most important architectural decision you will make. But the Personal Data Protection Law is about how personal data is collected, used, kept and protected across its whole life, and where the servers sit is only part of that.
What the PDPL expects, in brief
The PDPL came into force in September 2023, with a one-year transition before enforcement began in September 2024. The Saudi Data and Artificial Intelligence Authority (SDAIA) supervises it, and implementing regulations — including specific rules on transfers outside the Kingdom — sit alongside the law itself. In broad terms, organisations that process personal data are expected to:
- Have a lawful basis for each processing activity — consent, or one of the other bases the law recognises.
- Collect data for a specific, stated purpose, and not use it in ways incompatible with that purpose.
- Collect only what the purpose needs, and keep it accurate.
- Keep it no longer than necessary, then destroy it, subject to limited exceptions.
- Honour data-subject rights: to be informed, to access and obtain a copy, to correct, and to request destruction.
- Protect data with appropriate technical and organisational measures, with extra care for sensitive categories such as health, credit, genetic and biometric data.
- Notify SDAIA of breaches that could cause harm — the implementing regulations set a 72-hour window — and inform affected individuals where required.
- Transfer data outside the Kingdom only where the conditions in the law and its transfer regulations are met.
What local hosting does for you
Hosting models, indexes and logs in the Kingdom — on-premises or in a private cloud region — brings real advantages:
- It largely takes cross-border transfer off the table for that system, removing one of the more demanding sets of conditions in the law — provided nothing else in the chain, such as remote support, monitoring tools or backups, reaches outside it.
- It removes an outside AI provider as a processor of your prompts and outputs, along with its retention, abuse-monitoring and sub-processor arrangements.
- It puts logs, backups and deletion under your own control, so you can show where data is and act on a destruction request.
- It makes the security controls — network isolation, encryption, access control, monitoring — yours to design and audit.
In many sectors it also lines up with other requirements: the NCA’s cybersecurity controls and sector regulators such as SAMA set their own expectations for where and how data is hosted.
What it doesn’t do
None of the following changes because the server is in Riyadh:
- Whether you have a lawful basis to use customer data in an AI system at all — especially for a purpose different from the one it was collected for.
- Whether the system collects more than it needs. An assistant that stores every conversation by default fails data minimisation wherever it runs.
- Whether retention rules are applied. Local logs kept forever are still kept forever.
- Whether you can answer an access or destruction request that touches AI data.
- Whether the right people see the right data. A local assistant that ignores document permissions can disclose personal data to colleagues who should never see it.
Local hosting narrows the risk. It does not discharge the obligations.
Where AI creates new personal-data surfaces
Traditional systems keep personal data in databases with known fields. AI systems spread it into places that data maps often miss:
- Prompts. Users paste in e-mails, CVs, medical notes and customer complaints. Whatever the system was designed for, it will receive personal data.
- Conversation logs. Useful for quality review and audit — and a growing store of personal data with no natural expiry.
- Vector indexes. Retrieval systems split documents into chunks and store them with embeddings. Deleting the source file does not delete its chunks unless the pipeline is built to, so deleted content can keep appearing in answers.
- Fine-tuning data. A model can memorise records used to adapt it, and they cannot simply be removed from its weights afterwards.
- Model outputs. Summaries, classifications and drafts about a person are personal data in their own right, and inferences can be sensitive even when the inputs were not.
- Third-party telemetry. Monitoring, analytics, error-tracking and evaluation tools may copy prompts and outputs to services outside the Kingdom without anyone deciding they should.
Practical controls
- 1Map the data flow for each AI use case end to end, including logs, caches, indexes, backups and every tool that receives a copy.
- 2Record the purpose and lawful basis for each use case, and check any new use of existing data against the purpose it was collected for.
- 3Minimise at the edge: mask or pseudonymise identifiers before they reach a model when the task does not need them.
- 4Set retention for each surface — prompts, logs, indexes, outputs — and enforce it automatically.
- 5Build deletion that propagates: a destruction request should reach the source record, its chunks and embeddings, cached answers and logs.
- 6Keep personal data out of fine-tuning unless there is a clear basis and a documented need; prefer retrieval, which can be updated and deleted.
- 7Carry existing permissions into retrieval, so an assistant only returns what the user could already open.
- 8Review every third-party component for where it sends data, and switch off telemetry that leaves your boundary.
- 9Include AI systems in your breach-response plan, with logs good enough to tell what was exposed and to whom.
Working with your legal team
The projects that go well put engineers and the legal team in the same room early. Engineers can show where data actually flows; lawyers can say which flows need a lawful basis, a notice, an impact assessment or a contract. Neither can do the other’s part, and a data-flow map both sides have signed off is worth more than any single control.
That is how we work too. We design AI systems that support these obligations — residency, purpose limitation, retention and data-subject rights — and document them so your legal and compliance teams can assess them. We do not give legal advice and we do not certify compliance; that judgement stays with your advisers.