Private AI in Saudi Arabia, explained

“Private AI” is used loosely. For a buyer it comes down to four questions, and a platform either gives you a clear answer to each or it does not.

What private AI means in practice

• Where does the model run? On your servers, in a cloud region you choose, or in a vendor’s shared service.
• Where is the data stored? Questions, answers, documents and the evidence behind them.
• Who can sign in? Your own identity provider, or accounts the vendor manages.
• What can the agent do on your systems? Read only, or change things, and who decides.

A public AI chat service answers the first two questions for you: prompts go to the provider’s infrastructure, often outside the Kingdom. Private AI keeps those answers in your hands.

Customer-controlled inference and execution boundaries.

Why Saudi organisations ask for it

Three sets of rules come up in almost every conversation. None of them bans AI; each one asks you to know where data goes and who controls it.

• The Personal Data Protection Law (PDPL), enforced by SDAIA, governs how personal data is processed and when it may be transferred outside the Kingdom. See the PDPL and AI guide.
• NCA’s Essential Cybersecurity Controls (ECC-2:2024) apply to national entities and cover identity, logging, data protection, cryptography and third parties. See AI agents and NCA ECC.
• SAMA’s Cyber Security Framework applies to banks and other financial institutions SAMA regulates. See self-hosted LLMs for banks.

How AcTree runs privately

The platform is installed where you choose. The model is one you connect, and sign-in goes through your own identity provider.

Customer-hosted deployments use packaged deployment configuration. An offline deployment requires an accepted bundle, local model and disconnected installation tests.

Bring your own model — OpenAI-compatible endpoints, Ollama, Anthropic, or Cloudflare AI Gateway / Workers AI — with keys envelope-encrypted at rest.

Customer-hosted installations support an identity provider through OIDC with PKCE. Hosted workspace SSO must be confirmed for the agreed deployment.

The architecture, in one paragraph

Inside your environment sit the control plane and its database, the console your teams use, and a gateway. Servers you want agents to work on are enrolled one by one. The model sits behind an endpoint you choose: an open-weight model on your own GPUs, or a hosted model in a region you approve. The architecture page shows the sovereign, hybrid and hosted splits side by side.

A host joins the fleet with one install command and an enrollment token; revoking the host unpairs it.

What the agents do on day one

The first use cases are the ones where evidence matters most: IT operations, security investigations and control evidence for GRC. Each answer shows how it was reached, and hosts are read-only unless an administrator decides otherwise.

Read-only execution policy is the default on every enrolled host; the full policy is an explicit, per-host choice.

Host investigations show the commands that ran, their exit codes and output, per host.

The use-case library has a short film for each team, from IT to HR, all with sample data.

A checklist for evaluating any private AI platform

1. Can it be installed entirely on infrastructure you control, including without internet access?
2. Can you choose and change the model, including one served on your own hardware?
3. Does sign-in go through your identity provider, with no shared default administrator?
4. Are administrative actions written to an audit log you can export and review?
5. Does each answer show its sources and the exact steps behind it?
6. Is the default on your systems read-only?