Staff are already pasting company data into public AI tools, and we need a safe way to use AI instead of a ban

The situation

Your people have discovered how useful generative AI is. They summarize documents, draft emails, write code and analyze spreadsheets with public chatbots. Some of that content is contracts, customer records, source code and internal plans. Nobody approved it, and nobody can see it.

A ban will not work, and you know it. Staff will use the tools on their phones. What you need is a way to give the organization AI that is safe for your data, with rules people can actually follow.

Why this is hard

• Public AI tools are easy and good. Any approved alternative has to be nearly as convenient or it will be bypassed.
• Data leaves the organization in prompts and uploads, and the terms of consumer services are not written for your confidentiality, regulatory or data-residency obligations.
• The PDPL applies to personal data pasted into a prompt just as it applies to a database. The NCA Essential Cybersecurity Controls (ECC) and sector rules such as the SAMA Cyber Security Framework do not stop applying because the tool is new.
• The choices are many: enterprise versions of public services, cloud AI platforms with Saudi regions, private models hosted in the Kingdom, or a mix. Each has a different security, cost and capability profile.
• Policy without controls is ignored. Controls without policy are resented.

What goes wrong when it is ignored

• Confidential or personal data ends up in a third party’s systems with no record of what was shared.
• A customer or regulator asks how AI use is governed and the organization has no answer.
• Shadow AI spreads: unmanaged subscriptions, browser plugins and scripts calling external models from internal systems.
• Leadership eventually imposes a blanket ban, productivity drops, and use continues in secret.

How I approach it

1. Find out what is really happening. An AI inventory: which tools staff use, for what, with which data, and which internal systems already call external models. This is usually more than anyone expected.
2. Classify the use cases. Low-sensitivity tasks that any approved tool can serve, tasks that need an enterprise service with contractual data protections, and tasks where data must stay in the Kingdom or inside your own environment.
3. Choose the model and hosting arrangement for each class. Options include enterprise tiers of public services, cloud AI platforms in Saudi regions on AWS, Google Cloud or OCI, and private or sovereign models on Saudi-hosted or on-premises infrastructure. The choice is driven by data sensitivity, regulation, latency, cost and the skills you have to run it.
4. Put a control point in place. An LLM gateway, a single enterprise entry point through which applications and users reach approved models, gives you model allowlists, identity, logging and data-protection checks in one place rather than per team. I design the architecture; the product choice follows the requirements.
5. Write a short acceptable-use policy. What may be shared with which tool, what may never be shared, and what to do when unsure. One page, tied to the tools you actually provide.
6. Make it visible and governed. Logging of prompts and model use at the gateway, periodic review of the inventory, and a route for staff to request new tools so the next wave of adoption is managed rather than hidden.

What you receive

• An AI usage inventory and data-exposure summary.
• A use-case classification with the permitted tool and hosting option for each class.
• A target architecture for enterprise AI access, including an LLM gateway design where appropriate.
• A model and hosting options assessment: enterprise services, cloud AI in Saudi regions, private or sovereign AI.
• A one-page acceptable-use policy and the controls that enforce it.
• A governance routine that connects to your ISO/IEC 27001 or NCA program if you have one.

Where to start

A free 30-minute Architecture Decision Review at /free-architecture-review/ is enough to look at what your staff are using, what data is at stake and which hosting options are realistic for you. Or describe the situation through the contact form and choose the private and sovereign AI topic.