AI technologies

Not on the business terms we use. The paid APIs of the major providers state that customer data isn't used to train their models, and we never put your data into free consumer tools whose terms can allow it.

Every project starts with a data-flow statement that names each service and the country where your data is processed and stored. For sensitive data, we design the solution to run on your own infrastructure or inside the Kingdom.

We remove or mask personal data first wherever we can. If it still has to be transferred, we help you meet the law's transfer conditions with SDAIA's standard contractual clauses and a written transfer risk assessment.

Yes. Search results are filtered by each user's existing permissions, data stays in its source systems where possible, and every answer is logged with its sources.

It depends on the data and the use case. We classify your data first: public and low-sensitivity data can go through provider APIs on business terms, while restricted and sensitive data goes to a private model. Many solutions combine both behind one gateway, and we decide by testing on your own data, not on public leaderboards.

Yes. Open-weight models can run on your servers or in a fully isolated environment with no internet connection. We agree in writing who owns the hardware and who handles updates and model upgrades.

It depends on the model size, the number of users and how much text each request carries. We size the servers and GPUs after a short assessment, and only then give you a price.

Most projects start with retrieval-augmented generation, because facts change and answers can cite their sources. Fine-tuning helps with format, tone and terminology, or to make a smaller model cheaper to run.

It varies by model and task, so we test on your own Arabic text with native-speaking reviewers before choosing. Formal Arabic is usually strong; dialects and text that mixes Arabic and English need more testing, and we tell you plainly what we find.

Answers come only from the documents you specify, cite their sources and can say “I don't know”, and sensitive answers go through human review. Before launch, we measure quality on a test set agreed with you and rerun it on every change. No one can honestly promise zero errors.

Only within the limits you set. Agents can prepare and propose, but a person on your team approves anything consequential, every tool has a narrow scope, and every action leaves an audit trail.

Yes, we design for them. Each project is mapped to the principles and its risk level, alongside the PDPL and your data classification, and we hand you the file. We describe this as alignment with the principles, not as a certification.

Yes, through their APIs and Model Context Protocol servers, starting with read-only access. On WhatsApp, the assistant serves your own customers, with their opt-in and a clear way to reach a person.

The gateway can switch to a backup model automatically, and we pin model versions so nothing changes without testing. When a model is retired, its replacement has to pass the same test set first.

There are three parts: building, with a fixed quote per phase; running, based on model usage or servers; and support. We also show you what lowers running costs, such as caching, batch processing and routing simple tasks to smaller models.

You do. Fine-tuned weights, prompts, code and test sets are yours, within the licence of the base model we started from, and we hand them over in full.