Why does AI governance need a framework, not a policy document?
Most Saudi enterprises that begin adopting AI agents already have a policy document. It usually says the right things: human oversight, data protection, no sensitive data in prompts. The problem is that a policy describes intent, while an agent operates in production. Between the two sits a gap that only a framework closes: named owners, a risk assessment that produces decisions, permissions that are technically enforced, security tests that are run and recorded, and a release gate with a signature.
A framework is different from a policy in three ways. First, it assigns accountability to roles, not departments. Second, it makes each control testable — you can show evidence that it works, not just that it was written. Third, it defines the change process, because an AI agent is not a static system. Its prompts, knowledge base, and tool access change over time, and each change is a new risk surface.
This matters more in the Saudi context because PDPL expectations, data residency, and Arabic-language handling all add requirements that generic global templates do not address. A framework built for a US SaaS company will not tell you how to handle candidate CVs in Arabic, or how to keep customer conversations inside the Kingdom. The framework has to be yours, and it has to be operational.
Who owns what? Assigning accountability across the AI lifecycle
Governance fails when ownership is shared vaguely. A workable model names four accountable roles for every AI agent, and writes them down before any code is built.
The business owner is the executive who benefits from the workflow — the head of procurement for a quotation agent, the CX director for a WhatsApp agent. This person owns the outcome and the acceptance criteria. The product owner translates the workflow into agent behaviour: what it reads, what it drafts, what it escalates. The technical owner is accountable for access control, logging, and the integration surface. The risk and compliance owner — often from legal, security, or data governance — holds a veto at the release gate and signs the risk assessment.
Crucially, the human approver inside the workflow is also a named role. If an agent drafts a quotation and a human approves it, that human is a control, and the framework must say who they are, what they see, and what they are accountable for. An approval step with no named approver is theatre, not governance.
How do you assess risk before an agent touches production data?
Risk assessment should produce decisions, not a score. Start by classifying the data the agent will touch: public, internal, confidential, or personal. Personal data — customer records, candidate CVs, employee information — triggers PDPL obligations and should default to the strictest controls. Then classify the action: read-only, draft-only, or write. Reading a document is a different risk from sending a customer message or updating a record.
Next, map failure modes explicitly. What happens if the agent retrieves the wrong document and answers confidently? What if it misreads an Arabic CV because of dialect or formatting? What if a prompt injection in an inbound email tries to make it exfiltrate data? Each failure mode gets a likelihood, an impact, and a named mitigation. The output is a short risk register that the compliance owner signs — not a 40-page document nobody reads.
Finally, decide the autonomy level. LeenAI starts systems with bounded autonomy rather than open permission to act, and this is the right default. A new agent should begin read-only, then move to draft-with-approval, and only then to bounded write actions with audit logging. Each step up is a governance decision with evidence behind it, not a schedule milestone.
What permissions and controls must be enforced in the system itself?
Documenting a control is not the same as implementing it. A framework must specify which controls live in the system and can be demonstrated, not just described. Role-based access control is the foundation: the agent should have its own identity, with the minimum permissions needed for its workflow, and no standing access to systems it does not need. Data minimisation means the agent retrieves only the fields required to complete the task, not whole records.
Audit logging is the second enforced control. Every retrieval, every draft, every tool call, and every human approval should be logged with who, what, when, and which source was used. This is what makes an agent defensible in an audit and useful in a weekly review. Retention rules should be documented before code is written, so logs are kept long enough to investigate and not longer than necessary.
Data residency and prompt hygiene round out the technical controls. Where personal data is involved, the processing should stay within approved environments, and business users should be trained to avoid entering sensitive personal or confidential data into AI prompts. That last point is a control too — a human one — and it needs training, not just a policy line. For teams building knowledge agents, LeenAI's OpsRAG is designed around cited answers from approved documents, which makes the retrieval boundary visible and testable.
How should you test security, and what does change review look like?
Testing is where most governance programmes are thinnest. A control that has never been tested is a hypothesis. Security testing for an AI agent should include adversarial prompt tests — attempts to make the agent ignore its instructions, reveal its system prompt, or act outside its permissions — plus data-boundary tests that confirm the agent cannot retrieve documents outside its approved scope. Arabic and English inputs should both be tested, because retrieval quality and instruction-following can differ by language.
Change review is the discipline that keeps the framework alive after launch. Every change to prompts, knowledge sources, tool permissions, or model version should pass a lightweight review: what changed, why, what was tested, and who approved. This does not need to be heavy, but it does need to be recorded. LeenAI identifies the ongoing discipline of maintaining rules, reviewing agent activity weekly, and updating rules as the business evolves as a major cost of AI adoption — and change review is exactly where that cost shows up.
A weekly review cadence makes this practical. The business owner looks at a sample of agent activity, the technical owner checks logs and escalations, and together they decide whether any rule needs updating. Over time this becomes the operating rhythm of the agent, and the evidence trail for the next audit.
What does a credible go/no-go release decision require?
The release gate is the moment the framework either works or is bypassed. A credible go/no-go requires four artefacts: a signed risk assessment, evidence that the enforced controls were tested, a defined autonomy level with its boundaries, and named humans for every approval step. Without these, the decision is opinion.
This is also where acceptance criteria matter. A serious buyer should see an acceptance pack with KPIs, UAT notes, and operating guidance — proof that the agent was measured against a baseline, tested with real users, and handed over with instructions for running it. LeenAI builds pilots around this idea, and you can see how we scope pilots with fixed deliverables and defined acceptance criteria. The customer owns the system and all artifacts, which matters for governance: you cannot govern what you do not own.
If you are evaluating where to start, talk to us with one workflow. Within 24 hours we reply with scope, timeline, and SAR cost — and a clear view of the governance controls that workflow will need. Proof before claims means the framework comes first, and the agent earns its autonomy step by step.




