Responsible AI in Saudi Arabia is a compliance requirement with a regulator attached, not a values statement you publish and forget. Between the Personal Data Protection Law (PDPL), SDAIA’s AI ethics principles and your sector regulator, the governance has to exist before the model goes live. This post covers the obligations that bite, the one question that sets how much governance you need, the five checks that make a deployment defensible, and where owners get it wrong.
IBM put the average data breach in the Middle East at $8.75M in 2024, the second-highest of any region, against a global average of $4.88M (IBM via The National, 2024). An AI system that ingests customer data is one more way to have that breach.
The obligations that actually bite
The PDPL took effect on 14 September 2023, full compliance was required by 14 September 2024, and SDAIA is the regulator enforcing it (Morgan Lewis, 2024). These are the obligations that matter in practice.
- Personal data is regulated the moment you touch it. Train or run AI on personal data and you are under PDPL: lawful basis, data minimization, purpose limitation, individuals’ rights, and limits on cross-border transfer, governed by Article 29 and SDAIA’s Transfer Regulations, now under enforcement (trade.gov, 2024). “We fed it customer data” is a regulated act, and outside the Kingdom, a regulated transfer too.
- Purpose limitation follows the data into the model. Data collected to deliver a service was not collected to train a model that scores people. A new purpose needs its own basis.
- You may have to explain the decision. A model that cannot justify its output is a liability wherever a regulated decision is made: credit, hiring, insurance, healthcare.
- Bias is a legal exposure, not just an ethical one. Skewed or discriminatory outcomes hurt people and expose you. You test for it, you don’t assume it away.
- Someone has to own it. “The AI decided” is not a defence. A named person is accountable, and the regulator will want that name.
For the engineering view of the law itself, see PDPL compliance for tech leaders.
The one question that sets your governance level
Owners hear “AI governance” and imagine a committee and a long delay. Most AI needs neither, just one honest question answered before any code is written: does this system make, or materially shape, a decision about a person?
flowchart TD
A(["A proposed AI use"]) --> B{"Touches personal data?"}
B -->|No| C(["Light: acceptable-use policy and a named owner"])
B -->|Yes| D{"Makes or shapes a decision about a person?"}
D -->|No| E(["Standard: data inventory, lawful basis, access controls"])
D -->|Yes| F(["Full: add impact assessment, bias testing, human review, records"])
- Light. A drafting assistant on non-personal documents, a coding assistant. Write down what it may be used for, name an owner, move on.
- Standard. A support bot that reads customer records, invoice extraction with names on the invoices. Personal data is in play, but nobody is being judged.
- Full. Anything that scores, ranks, approves, flags or prices a person: loan pre-screening, CV shortlisting, fraud flags. Every check below applies.
Most AI ideas land in Light or Standard, where governance is cheap. The effort goes to the handful in Full.
The five checks that make a deployment defensible
You don’t need a 200-page framework. You need these five things to be true, each with a document you could hand to an auditor tomorrow.
flowchart TD
A(["1. Know your data and lawful basis"]) --> B(["2. Human oversight where it matters"])
B --> C(["3. Test for bias and accuracy"])
C --> D{"Affects people materially?"}
D -->|Yes| E(["4. Document decisions and keep records"])
D -->|No| F(["5. Set the responsible-use boundary"])
E --> F
- Know your data. The artifact is a data inventory: every field, where it came from, why you hold it, where it is processed. An API hosted abroad needs a transfer mechanism on that row. For the practical side, see using your company’s data with AI safely.
- Build in human oversight where decisions materially affect people: in the loop (approves each decision) or on the loop (samples and can halt), with real authority to overrule. The artifact is a review procedure and a log proving reviews happened.
- Test for bias and accuracy before launch, then watch for drift. Split the evaluation set by nationality, gender, age band and region and compare error rates against thresholds you wrote down first. The artifact is a dated test report, re-run on a schedule.
- Document what the system does and why, and how you validated it. A one-page model card: purpose, inputs, training data, known limits, owner, review date. Add an impact assessment for the Full tier. This is what survives an audit.
- Set the responsible-use boundary. A short acceptable-use policy: what the AI may and may not be used for, and what happens when someone breaks the rule. Staff pasting customer data into a public chatbot is the failure it prevents.
None of this is a reason to avoid AI. It is what lets you deploy it in a regulated market and sleep at night, and in our experience far cheaper built in from day one than retrofitted.
A worked example: pre-screening financing applications
Say you run a consumer financing company supervised by SAMA and want a model to pre-screen applications so the credit team reads only the strong ones. The numbers are made up to show the shape.
You receive 3,000 applications a month, and four analysts spend most of their time on the first pass. The model auto-declines the weakest 40% and passes the rest through with a score.
- Gate. It shapes a decision about a person, so it is Full tier from day one.
- Data. Applications, bureau data, transaction history: all personal. The lawful basis is the applicant’s contract, documented per field. Hosted in-Kingdom, so no Article 29 transfer, which alone saved a month of legal work.
- Oversight. Auto-decline is never final. The 1,200 monthly declines go to a queue where one analyst reviews a 10% sample and can overturn, and every letter offers a human review.
- Bias test. You run the model on last year’s 36,000 decided applications and compare decline rates by nationality, gender and region. Say one region is declined at 1.4 times the rate of the others with no difference in repayment. That is a finding, not a blocker: remove the proxy feature, re-test, record both runs.
- Records. One model card, one impact assessment, one bias report, one review log. Fifteen pages, perhaps.
Three analysts move to higher-value work; the fourth runs the queue. When SAMA or SDAIA asks why an applicant was declined, you answer in a day with a paper trail instead of a week with a scramble. The governance cost roughly three weeks, once. Without it the project ships faster and is a liability on day one.
Where this goes wrong
The failure modes are consistent, and most are false economies.
- Governance as a document, not a system. A policy has to show up as access controls, review queues and test jobs, or it is decoration.
- Bolting it on after launch. Retrofitting a lawful basis to data already used for training is between painful and impossible. Retrofitting human review means rebuilding the workflow.
- Trusting the vendor’s word. “Our model is unbiased and PDPL-compliant” is a sales claim, and the exposure is yours. Ask for the test methodology, hosting region and sub-processor list, in writing.
- Skipping the risk controls to move faster. Gartner predicts at least 30% of generative AI projects will be abandoned after proof of concept by the end of 2025, and inadequate risk controls are among the reasons it names (Gartner, 2024). The proof of concept that ignored governance never gets sign-off for production.
- Forgetting the security layer underneath. AI governance sits on top of ordinary cybersecurity. If your controls do not meet the baseline yet, start with NCA ECC compliance and add the AI layer on top.
How to start without a 200-page framework
The pattern that works is boring and reliable. List every AI use in the company, including the unofficial ones. Run each through the gate and sort them into Light, Standard and Full. Pick one Full-tier system, ideally the one that touches the most people, and take it through the five checks properly. Measure what it cost in time. Then turn what you produced into a template: inventory, model card, test report, review log. The second system takes a fraction of the time, because you are filling in a form instead of inventing a process.
After two or three, you have a governance practice without ever writing a framework, and the next regulator question is a lookup, not a crisis.
Deploying AI on personal or regulated data in the Kingdom? SDCG builds responsible AI governance aligned to PDPL and SDAIA principles into your systems from the start, so you can adopt AI confidently and defensibly. We’re independent, so we’re not steering you toward anyone’s product. Book a free 30-minute review.