Ting Guides · For the CTO and CIO
What should an AI vendor due diligence checklist cover?
The short answer
Evidence, not answers. Ask the vendor to run their system on a sample of your own data while you watch, then check where your data is processed, which third-party models it depends on, who owns the code and prompts, how quality is measured after launch, and what happens to your data and system if you leave.
Reviewed by Badr Al-RashedUpdated
What it depends on
Buying a product or commissioning a build
A packaged product is checked by trial: run it on your data and read the contract. A custom build is checked by method: the vendor's specification, test approach and the people who will actually do the work.
What the system can touch
A tool that drafts text for staff to review carries less risk than one that messages customers or writes to your ERP. The more it can do on its own, the more you need to see its permissions and its logs.
What data it will see
Customer records, health or financial data raise the bar on hosting, access control and sub-processors. Public product information barely does.
How replaceable it is
If the vendor disappears, can you keep running it? Check what you get on exit: a data export, the configuration, the prompts, the test sets and, for custom work, the code.
Checks specific to Saudi buyers
General information, not legal advice. Each regulatory point summarises the official source cited beside it, as checked on the date shown under Sources; confirm how it applies to your company with your legal adviser.
Where each component processes data
Ask for the hosting location of every component, including the model provider. Article 29 of the Personal Data Protection Law lists the purposes for which personal data may leave the Kingdom and the conditions, including an adequate level of protection and the minimum data needed; the Transfer Regulation sets the exemptions from those two conditions and the safeguards they require, such as standard contractual clauses (Article 4). "Hosted in the region" is not an answer; a list of components and locations is.[1][2]
Arabic on your own material
A demo in English, or in Arabic written for the demo, tells you little. Bring your own: scanned forms, dialect messages, voice-note transcripts. Score them apart from the English results.
The risk category of the use
SDAIA's AI Ethics Principles sort AI systems from little or no risk to unacceptable, and require high-risk systems to undergo "pre- and post-conformity assessments". Agree with the vendor which category your use falls into, and what evidence that requires.[3]
Support in your working week
Confirm who handles incidents, in which language and on which days. A vendor working Monday to Friday from another time zone leaves gaps in a Sunday-to-Thursday operation.
Seven checks, the evidence that settles each, and the red flag
ExhibitWorked example
The procurement team of the fictional Rasikh Holding vets a vendor that will build a contract-review system. Each check is closed by a document or a run you can see, never by a verbal assurance.
| Check | Evidence to ask for | Red flag |
|---|---|---|
| Does it work on our data? | A live run on cases you choose, hard ones included, with the outputs saved | A scripted demo only, or "we'll test after signing" |
| Where is our data processed? | Every component, sub-processor and location, the model provider included | One word: "cloud" or "secure" |
| Which models does it rely on? | Named models and versions, and the plan when a provider changes or retires one | "Proprietary AI" with no detail |
| Can it take actions? | Each action, the permission behind it, and the log a reviewer can read | The model decides what it may do |
| How is quality measured after launch? | The metrics, the review rhythm and an anonymized report from a live deployment | One accuracy figure with no test set behind it |
| What do we own? | Contract terms for data, configuration, prompts, test sets and custom code | Prompts and configuration stay with the vendor, with no export |
| What happens if we leave? | An exit plan: export format, notice period, deletion confirmation | Exit isn't in the contract |
The decisionNo contract while any row still shows its red flag. Close each check with the evidence in its row, then sign.
What Ting does and doesn't do here
What we do
- We review the vendor's proposal and architecture against your requirements, and turn each claim into a test or a document you can check.
- We run the live test on your data with you, and score the Arabic separately.
- We stay on your side through delivery: milestone demos tested against agreed cases, evidence reviews and acceptance advice.
Where Ting helps: AI Vendor Oversight
What we don't do
- We don't certify vendors. We report what was evidenced against the criteria you agreed.
- We don't negotiate commercial or legal terms. Those stay with your procurement and legal teams; we give them the technical evidence.
- We don't treat an ISO/IEC 42001 certificate or a framework logo as proof that one specific system works. It shows a management system exists, which is useful but not the same thing.
Sources
Primary sources only, each read in full on the date shown.
- Personal Data Protection Law (Royal Decree M/19 of 9/2/1443H, amended by Royal Decree M/148 of 5/9/1444H)SDAIA · checked
- Regulation on Personal Data Transfer outside the Kingdom, version 2.0 (August 2024)SDAIA · checked
- AI Ethics Principles, version 1 (May 2025)SDAIA · checked
- NIST AI 600-1: Generative AI Profile (July 2024)NIST · checked
- LLM06:2025 Excessive AgencyOWASP GenAI Security Project · checked
- ISO/IEC 42001:2023, AI management systemsISO · checked
