Large organizations in the Kingdom now come to any conversation about AI with a story ready. A virtual assistant in customer service, a document-reading tool in finance, a forecasting model in operations, each with a successful demonstration and an enthusiastic sponsor. Ask what any of them contributes to this year’s results, and the room tends to go quiet.
The pilots worked. They usually do. A capable team with a clean sample of data and a few weeks can make almost anything look possible. What separates the organizations pulling ahead is the unglamorous work that follows: choosing which ideas deserve to become products, governing them so they can be trusted, feeding them data that holds up at scale, and running them with the discipline applied to any system the business depends on.
The binding constraint on AI value has moved from technology to management. That makes AI a leadership agenda, and it should be run like one.
Run AI as a portfolio
Pilots multiply when nobody is required to choose between them. Each department funds its own experiment, each vendor proposes its own proof of concept, and the organization ends up with a long list of promising work and no single initiative large enough to move a cost line or a service outcome.
The remedy is the portfolio discipline a board would expect of any capital program. Every use case enters with a named business owner, a value hypothesis expressed in the business’s own measures, and an honest view of the data and process change it will require. Every use case also carries a kill rule: the evidence that would stop it, agreed before work begins. Funding is released in stages, and it flows toward the few use cases that clear each gate.
It feels slower at first. In practice it is the fastest route to production, because scarce engineers, data owners and executive attention stop being spread across experiments that were never designed to scale. A short list of use cases running in production is worth more than a long list in permanent pilot.
Publish the route to production
Responsible AI is often cast as a brake. In our experience, its absence is what stalls programs. When nobody knows who may approve a model for use with personal data, or what testing a customer-facing system needs before launch, every use case becomes a fresh negotiation with legal, risk and cybersecurity teams. Under that uncertainty, the safest answer is always to wait.
Good governance turns those negotiations into a standing route to production. Use cases are classified by risk at intake, so an internal summarization tool does not face the scrutiny reserved for a model that shapes credit, eligibility or clinical decisions. Each risk tier has published requirements for testing, documentation, explainability and human review. Obligations under the Personal Data Protection Law and the national data management standards are designed into the route once, so project teams do not rediscover them one use case at a time.
Leaders can test their AI governance with a single question. Does a well-prepared team know exactly what it takes to reach production, and how long that will take? If the answer depends on who is asking, the governance is not yet doing its job.
A pilot proves the technology works. Production proves the organization does.
Engineer for the second year
A pilot is judged in its first week. A production system is judged in its second year, after the data has drifted, the original team has moved on and the vendor has shipped several new model versions. That is where data readiness and machine learning operations earn their place.
Data readiness is narrower than many programs assume. It means making the specific data each priority use case depends on reliable, owned and governed, with quality monitored at the source. Waiting to clean the entire data estate first is a program that never ends. Over time, the AI portfolio becomes the most practical roadmap for data investment, because it shows exactly which data creates value and when.
Operations deserves the same deliberate design. Before any use case is cleared for production, its sponsors should be able to answer a short set of questions:
- Who owns the model’s performance once it is live, and what happens when it degrades?
- How are drift, bias and failure detected, and who is alerted when they occur?
- Can every version be traced, rolled back and explained to an auditor or a regulator?
- What will the system cost to run at full volume, and which budget carries that cost?
- Where do the data and the model reside, and does that meet the organization’s sovereignty requirements?
The machinery behind those answers, from deployment and monitoring to evaluation and cost control, belongs on a shared platform that every team uses. Without it, each use case pays for its own plumbing, and the portfolio never gets cheaper to grow.
Give agents a chain of command
The next wave of AI acts as well as answers. Agentic systems can read a request, gather information from several applications, draft a decision and execute the routine steps that follow. The productivity potential in process-heavy functions such as procurement, finance operations and beneficiary services is real. So is the risk of handing operational authority to software that nobody has been asked to supervise.
We advise leaders to design agentic workflows the way they would design a team. Each agent has a defined scope, explicit limits on what it may commit or change, and a named person accountable for its output. Decision rights are tiered: the agent handles the routine, recommends on the material and escalates the exceptional. Every action is logged in a form a reviewer can follow after the fact.
Human oversight works only when it is designed into the workflow. A reviewer asked to approve a constant stream of agent decisions without context becomes a rubber stamp. Oversight belongs where judgment matters, and it should arrive with the information needed to exercise that judgment well.
Keep the know-how in the building
The final constraint is the most durable. Organizations that rely entirely on vendors and advisers for AI find that every new use case restarts the learning curve, and that hard-won knowledge walks out when each contract ends. We say this as advisers. An engagement has done its job when the client can run what was built without us.
That capability has to exist at every level. Executives need enough fluency to set direction, challenge a value case and weigh risk. Product owners and process leaders must be able to redesign work around AI, which is where most of the value sits. A core technical team should own the platform, the standards and the vendor relationships. A center of excellence can anchor all of this, provided it is judged on use cases running in production and on capability transferred to the business.
None of this waits on a new model. It starts with a request any chief executive can make this quarter: every AI initiative on a single page, each with its owner, its value case and its route to production. Fund the few that have all three. Close or merge the rest, and judge the program from then on by what runs in production.