An AI strategy is a short set of choices: where AI will create value for your organization, what has to be true for it to work, and in what order you will move. Vision 2030 and SDAIA’s national strategy for data and AI tell you to move. They do not make those choices for you, and the prize for making them well is large: PwC projects AI could add about $135bn to Saudi Arabia’s economy by 2030, the largest gain in the Middle East (PwC, via US ITA).
This post covers the three questions a real strategy answers, a worked example, the traps we see most often, and how to start next quarter. First, what a strategy isn’t: a list of AI projects, a vendor platform, or a slide that says you will “embrace AI.” That’s all activity. A strategy decides which activity to fund and which to refuse.
The three questions a real strategy answers
Every AI strategy we have seen work fits on one page, because it only answers three things. The platform, the team, and the pilots follow from the answers.
flowchart TD
A(["An AI strategy"]) --> B{"Where does AI create value for us?"}
A --> C{"What has to be true for it to work?"}
A --> D{"In what order do we move?"}
B --> B2(["A ranked short list of use cases"])
C --> C2(["Data, operating model, governance"])
D --> D2(["A first win, then a sequence"])
Question one: where does AI create value for us
Value first, technology second. Start from the work, not from the models: a half-day session with the people who run operations, finance, customer service, and sales, listing where AI could move a number you already report. Cost per transaction, time to serve a customer, error rate, or a capability you lack today, such as reading every contract instead of a sample.
Then rank. Two scores per use case, each from 1 to 5:
- Value. How much would it move a number leadership already watches? Prefer boring line items with big volumes over exciting ones with small volumes.
- Feasibility. Is the data there, is a business owner named, is the “right answer” clear most of the time, and can a person check the exceptions? A use case with no data and no owner scores 1, however good the idea.
This ranking is where most programmes leak. Gartner expects at least 30% of generative AI projects to be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs, or unclear business value (Gartner, 2024). Every one of those four reasons is a feasibility score nobody wrote down. A focused strategy says no far more than yes, and says it early, on paper, before money moves.
Question two: what has to be true for it to work
This is the part the mandate skips and the vendor deck hides. Three things have to be true, and the strategy should name the gap in each.
- The data is usable. Not “we have a lot of data.” Usable means the fields you need exist, are filled in consistently, can be reached without a three-week request to IT, and have a named owner. In our experience almost every organization overestimates its readiness here, so fixing the foundation for your top three use cases is part of the plan, not another team’s problem. Not sure where you stand? Our AI readiness check walks through the questions.
- Someone owns the operating model. Who builds, who runs, who governs? What stays in-house and what is partnered? How do models get deployed, monitored, and retrained when accuracy drifts? Write it down for the first use case, even if it’s “one engineer plus a vendor for twelve months.” A strategy with no operating model behind it is a wish.
- Governance is designed in, not bolted on. In the Kingdom this is concrete. PDPL, regulated by SDAIA, governs any personal data your models touch and any transfer of it outside the country. NCA’s Essential Cybersecurity Controls apply to the systems the AI sits in. In financial services, SAMA expects to see the controls before you scale, and its regulatory sandbox exists for testing under supervision. If you sell to government through Etimad, your buyer will ask where the data lives and who can see it. None of this is a reason to wait; it is a list of requirements to bake into use-case selection (the practical side is in responsible AI and the PDPL).
Question three: in what order do we move
Sequence is the strategy. Start with a contained, high-value use case that builds capability and confidence, prove it, then expand into the harder ones once the data and the team can carry them. Don’t bet the budget on a multi-year programme before you have shipped a single thing.
flowchart LR A(["Vision 2030 mandate"]) --> B(["Rank use cases by value and feasibility"]) B --> C(["Fix the data for the top three"]) C --> D(["Decide who builds, runs, and governs"]) D --> E(["Ship a contained first win in 90 days"]) E --> F(["Measure, then expand"]) F -.-> B
The loop back matters. After the first win you re-rank, because you now know how clean the data was, how long integration took, and what governance cost. A strategy that never re-ranks is a plan from a year ago.
A worked example: fourteen ideas, one first win
Say you run a 300-person distribution business with an ERP, a CRM, and three warehouses on two different stock systems. The numbers here are made up to show the shape.
Your half-day session produces fourteen candidate use cases. Three score 4 or above on both value and feasibility:
- Supplier invoice extraction. Say 1,500 invoices a month at roughly 5 minutes each, about 125 hours of keying and checking. The ERP owner is named, and supplier data isn’t personal data, so PDPL exposure is low. Value 4, feasibility 5.
- Customer service triage in Arabic and English. Two years of tickets sit in the CRM, and the right answer is clear for the repetitive half. Value 4, feasibility 4.
- Demand forecasting. High value, but the three warehouses disagree on product codes and the sales history is split across two systems. Value 5, feasibility 2.
The strategy writes itself from the scores. Quarter one: invoice extraction as the first win, a person reviewing the messy 20%, measured on hours removed net of software and build cost. Quarter two: scale it and start triage, with a governance check because tickets contain customer data. In parallel, the unglamorous work of unifying product codes across the warehouses, because that is what has to be true before forecasting is feasible. Quarter four: forecasting, now scoring 5 and 4. Eleven ideas were told no, in writing, with a reason. Those reasons are as much a part of the strategy as the three yeses.
Where AI strategies go wrong
The traps are the same everywhere, with a couple of local flavours.
- The strategy is a list of projects. Twenty initiatives, no ranking, no gaps named. It reads well and delivers nothing, because nobody decided what not to do.
- Platform first. Buying an AI platform before a single use case is scored. The platform then hunts for problems, and the licence runs while it looks.
- Data assumed, not checked. The strategy says “leverage our data” and nobody has opened the tables. It is the most common reason a first pilot slips by six months.
- Governance bolted on at the end. A pilot built on customer data reaches the audit before it reaches production, and stops there. The fix was a week’s work at selection.
- Alignment by wording. Pasting Vision 2030 language into the slides while the plan changes nothing. The national agenda rewards genuine capability, local talent, and measurable outcomes, not AI theatre.
How to start next quarter
The pattern that works is boring and reliable. Pick one use case from your ranked list, the one with the highest feasibility among the high-value ones, not the most impressive. Measure what it costs you today, honestly, in hours and errors. Run a contained pilot with a person reviewing the AI’s work, held to a number you already report. If it pays back, scale it and take the next one off the list. If it doesn’t, you spent a little to learn something cheap, which is the point of a pilot. Running that first one so it leads somewhere is covered in an AI proof of concept that works.
Good AI strategy is focused, honest about the data and capability gap, governed by design, and sequenced to deliver something early and grow from there. The mandate says move. The strategy says where, in what order, and what you will refuse.
Need an AI strategy that delivers, not just one that satisfies the mandate? SDCG’s AI consulting practice helps organizations build focused, deliverable AI strategies aligned to Vision 2030, grounded in real value and an honest view of what it takes. We’re independent, so the plan is built around your outcomes, not a product we’re selling. Book a free 30-minute review.