Comment to add before publication: Revise before publication. Replace absolute claims about complete logs, tamper-evidence, and vendor edit access with testable requirements for event coverage, integrity protection, and administrative accountability. Define approval scope, expiry, and reapproval after material changes. Require authorised testing in an isolated environment, including access isolation, tool permissions, duplicate actions, and shutdown. Expand data-flow evidence to include extraction, embeddings, model processing, telemetry, backups, and support access. Distinguish contractual incident-notification commitments from applicable legal obligations. Clarify controller and processor responsibilities rather than placing all compliance responsibility on the customer. Verify LeenAI-specific security commitments and remove the unsupported 24-hour promise. Add evidence dates, tested versions, gaps, remediation owners, and explicit purchasing outcomes.
How to Evaluate the Security and Governance of Enterprise AI Solutions: A Saudi Buyer’s Guide What should you establish before evaluating an AI vendor? Start with the workflow the system will support. Define its users, data sources, expected outputs and permitted actions. A document-search assistant and an agent that sends customer messages need different controls.
Ask what the system can access, what it can change, which actions require approval and how those boundaries are enforced. Then identify the evidence needed to verify the answers.
Keep security, operational quality and data protection connected, but assess them separately. An accurate answer can expose restricted information. A well-protected system can still produce an incorrect recommendation. A documented policy can exist without effective implementation.
The evaluation should produce a purchasing decision supported by evidence, known limitations and assigned responsibilities. How should you assess access controls and credentials? Request a permission matrix connecting each capability to the access it requires. If the system prepares quotation drafts, it may need to read approved product and pricing information without modifying those sources.
Check both the agent’s service identity and the requesting user’s permissions. Broad backend access must not allow the assistant to reveal documents that the user cannot otherwise access.
Ask the vendor to demonstrate:
- Access using accounts with different permissions.
- Rejection of unauthorised retrieval and actions.
- Removal of access after a role change or revocation.
- Separation between customers, departments or workspaces where required. Credentials should be managed through an appropriate secret store, managed identity or other controlled mechanism. Ask who can access them, how they are rotated or revoked, and how exposure in prompts, logs and deployment files is prevented. Configuration may reference a secret without containing the secret itself.
Read-only access can reduce modification risk, but it does not prevent disclosure through answers, previews, exports or external processing. What evidence demonstrates data minimisation? Ask for a field-level explanation of what the workflow needs. An order-status assistant may need an order identifier and delivery status without requiring payment details or the customer’s full history.
Review the data actually transferred, not just the intended design. Use a permitted test case to inspect what reaches each service and what is retained.
Request a data-flow diagram covering source systems, document extraction, embeddings, retrieval, model processing, monitoring, logs, backups and support access. Identify subprocessors and processing locations.
Also ask about secondary use. Can prompts, documents or outputs be used for training, analytics or service improvement? What settings and contractual terms govern those uses?
A statement that data is not used for training does not answer questions about retention, access or deletion. Those require separate evidence. What makes a human approval gate meaningful? Define approvals by action and consequence. An internal summary, a customer commitment and a financial transaction should not automatically share the same approval rule.
The reviewer should see the proposed content, relevant evidence, intended action and recipient. Approval should apply to that specific version. Material changes should trigger a new review.
Test what happens when:
- The approver rejects or edits the output.
- Approval expires.
- The recipient changes.
- The approver is unavailable.
- Execution fails or returns an uncertain result. Silence should not become approval. Where automated execution is permitted, define limits, monitoring and a way to suspend it.
An approval record proves that approval occurred. It does not prove that the reviewer detected every error, so assess review quality as well as completion. What should audit records allow you to reconstruct? Define the events needed to investigate the workflow. These may include request identifiers, source references, configuration versions, tool calls, approvals, overrides and confirmed outcomes.
Distinguish an attempted action from a successful one. A request to send a message is different from confirmation that the receiving system accepted it.
Ask how records are protected against unauthorised access and alteration. Administrative access does not automatically invalidate an audit trail; the relevant questions concern restrictions, monitoring, integrity protection and accountability for changes.
Avoid unnecessary sensitive content in logs. Set retention, export and deletion rules, and define what happens if logging fails.
Test reconstruction using one sample case. Can an authorised reviewer establish what was requested, which sources were used, what was approved and what happened next? The evidence should support investigation without claiming to expose the model’s complete internal reasoning. Which security tests should you request? Agree the scope and authorisation before testing. Use an isolated environment with synthetic or otherwise permitted data.
Relevant scenarios include: Test What to verify Restricted-document request No disclosure through answers, citations or previews Instructions embedded in a document No unauthorised change to task or permissions Cross-customer request Effective isolation between accounts Invalid tool parameters Enforcement of action, recipient and value limits Repeated request after a timeout Duplicate prevention or controlled reconciliation Credential revocation Access stops as designed Logging failure The agreed response is applied Shutdown during queued work New work stops and pending actions are handled predictably
Request the tested version, date, method, results and unresolved findings. A demonstration on selected examples is useful evidence, but it is not equivalent to a comprehensive security assessment.
Prompt disclosure and sensitive-data disclosure should also be distinguished. Evaluate the consequences of what was exposed rather than treating every disclosure as identical. How should you assess privacy and incident responsibilities? Have the relevant legal and privacy owners assess the proposed processing against applicable requirements. Establish the parties’ roles and obligations rather than relying on a general “PDPL-ready” label.
Review permitted processing, data-subject request handling, retention, deletion, subprocessors, processing locations and relevant transfer arrangements.
A vendor’s contractual promises do not establish the customer’s overall compliance. Equally, the customer’s responsibilities do not remove obligations that apply to the vendor’s own role.
For incidents, request named contacts, escalation procedures, evidence-preservation arrangements and cooperation duties. Distinguish contractual notification commitments from applicable legal reporting requirements. How do you turn the checklist into a purchasing decision? Record evidence and gaps in a shared evaluation table: Area Evidence to request Decision question Access Permission matrix and isolation tests Are boundaries enforced? Credentials Storage, access and revocation controls Can access be controlled and withdrawn? Data handling Field inventory and complete data flow Is processing understood and bounded? Approvals Workflow rules and tested records Is approval specific and enforceable? Audit Sample records and integrity controls Can a case be reconstructed? Security Test results and open findings Are material risks addressed? Operations Monitoring, shutdown and incident procedures Can the team operate and respond? Exit Export, deletion, licence and handover terms Can the organisation manage transition?
Mark items as evidenced, partially evidenced, untested or not applicable, with a reason. Assign owners and deadlines to unresolved gaps.
Possible outcomes include acceptance for a limited scope, remediation and retesting, insufficient evidence, or rejection. Keep the decision tied to the assessed workflow and version.
To discuss a defined workflow and its evaluation requirements, explore how LeenAI scopes pilots or contact LeenAI. Bring the systems involved, data requirements, intended actions and the evidence your team needs before approval.




