Not Every Problem Needs AI
Abdulrhman Alodhayb explains why successful AI adoption depends on human judgment, governance, and organizational change—not simply having an AI strategy.
Abdulrhman AlOdhayb
Senior Technology Advisor to the Minister of Defense Office in Saudi Arabia
Human Judgment: AI adoption succeeds when leaders use models to sharpen decisions while retaining accountability for consequential outcomes.
Proven Results: AI-assisted governance cut review preparation time by 60 percent and surfaced major program risks weeks earlier.
Hidden Barriers: Complex institutions struggle when AI lacks undocumented relationships, shared context, and mechanisms for spreading lessons across teams.
Better Workflow: A disciplined AI workflow curates inputs, tests contradictions, applies expert review, and captures outcomes for continuous improvement.
Leadership Shift: Tomorrow’s technology leaders need judgment, humility, change-management skill, and strategic clarity more than technical authority alone.
Abdulrhman AlOdhayb works in technology and transformation within a complex institutional environment. He is Senior Technology Advisor to the Minister of Defense Office in Saudi Arabia.
We sat down with Abdulrhman to learn about the opportunities and pitfalls of AI adoption in complex, high-stakes environments. Here's what he said.
Technology Is Never the Hard Part
I am Abdulrhman Alodhayb, a technology and transformation executive with 19 years of leadership inside one of the most complex institutional environments imaginable: national defense. I have spent nearly two decades at the Ministry of Defense of Saudi Arabia, leading digital health transformation, enterprise architecture, and, most recently, serving as Senior Technology Advisor to the Minister of Defense Office.
Continue Reading for Free
Create a free account to finish this article, plus get ongoing access to timely insights and practical resources.
Have an account? Log In
My career has lived at an intersection that most technology leaders rarely encounter — where engineering decisions carry national security implications, where stakeholder alignment means bridging military command structures with clinical systems and enterprise platforms, and where the cost of a failed transformation is measured not in lost revenue but in institutional trust and operational readiness.
The thread running through all of it has been a single conviction: technology is never the hard part. The hard part is building the organizational conditions for technology to deliver value. I have designed ICT strategies for hospitals, contributed to a Ministry-wide Target Operating Model redesign with BCG, and guided enterprise-wide digital initiatives where failure was simply not an option.
Today, as artificial intelligence moves from experimentation into the core of system architecture, software development, and organizational enablement, I find myself at the most interesting moment of my career. I am most preoccupied with questions not about the technology itself but about how leaders build the judgment, governance, and resilient architectures that allow AI to accelerate delivery without eroding quality or security.
That is what I am here to talk about.
Working in the Strategic Layer
The technology estate I work across is substantial in every dimension. We operate enterprise-wide systems spanning military hospitals, command and control platforms, unified communications infrastructure, and national defense data frameworks. The architecture is inherently hybrid: on-premises sovereign deployments where data residency and security classification are non-negotiable, integrated with modernized platforms where we actively drive digital transformation.
I work at the strategic layer, advising at the ministerial level, leading PMO and IT strategic management functions, and aligning technology roadmaps across multiple directorates and organizational units. The human complexity rivals the technical complexity. Coordinating transformation across a multi-layered military and civilian workforce, each with different operating cultures and risk tolerances, is itself a discipline.
Product complexity here differs from commercial technology. We are not shipping features. We are redesigning how a national institution operates. This includes electronic health record systems for a network of military hospitals, enterprise resource platforms, and, most recently, contributing to a Ministry-wide Target Operating Model that reimagines how technology structures and enables the entire defense enterprise.
The deployment model is sovereign-first by necessity. Security, classification, and national data governance are not constraints we work around. They are the architecture. That reality shapes how I think about resilience, redundancy, and what responsible AI adoption looks like when the stakes are national rather than commercial.
How AI Makes the Invisible Visible
I introduced AI-informed governance checkpoints into our technology program review cycle last year. Before this change, program reviews focused on traditional delivery metrics: hitting milestones, tracking budgets, and logging risks. The process was disciplined but inherently backward-looking. We measured what had already happened rather than anticipating what might go wrong.
The shift came when I recognized that several enterprise programs generated enormous volumes of structured and unstructured data — across project documentation, vendor communications, change requests, and operational reports — but no one synthesized this signal to inform real-time leadership decisions. We treated this intelligence as an archive.
We changed the governance layer itself. We introduced AI-assisted analysis into the program review process, using language model capabilities to synthesize cross-program documentation, surface anomalies in delivery patterns, and generate early warning indicators before they became escalation items. This did not replace human judgment. It gave decision-makers a richer, faster picture of organizational reality, allowing them to apply judgment at the right moment and level.
The results were significant. Review cycles became more focused because leadership arrived with a pre-synthesized view of where attention was genuinely needed, rather than spending the first half of every session reconstructing status from raw reports. More importantly, we caught two significant program risks earlier than our traditional process would have surfaced, allowing intervention before impact rather than response after the fact.
This experience reinforced a principle I now hold firmly: AI does not transform organizations by replacing processes. It transforms them by making the invisible visible and giving leaders the clarity to act with precision rather than react with urgency.
Where Technical Teams Should Use AI
The boundary between AI-informed and human-owned decisions is not static. It shifts as trust is earned, capabilities mature, and the stakes of each decision domain become clearer.
In my environment, I actively rely on AI to inform or power technical work in several areas. One area is documentation synthesis and cross-program intelligence. Beyond that, I rely on AI to support architecture pattern analysis, specifically evaluating whether proposed solutions align with established enterprise standards or introduce technical debt that will compound over time. AI can now surface in hours what used to take a senior architect days to review — not as a final verdict but as a structured input that sharpens the human review that follows.
I also use AI heavily in the strategic planning layer: synthesizing technology landscape assessments, benchmarking against global best practices, and drafting the analytical scaffolding for roadmaps and target operating models. Where input volume is high and synthesis time is limited, AI delivers disproportionate value.
On the security side, AI-informed monitoring and anomaly detection has become a baseline expectation, not an innovation. Pattern recognition at the speed and scale that modern threat environments demand is simply beyond unaided human capacity. Here, I trust AI to surface and flag. I do not trust it to conclude or respond.
Where Tasks Should Remain Human
And governance itself remains human. I am deeply cautious about any organization allowing AI to govern AI. People who can be held responsible for the outcomes must own the oversight layer, ethical checkpoints, and decisions about where AI should and should not operate.
Everything that carries institutional consequence remains explicitly human. Architecture decisions that will shape the organization for five to ten years remain human-owned. AI analysis informs them, but people who understand the organizational context, political dynamics, and long-term implications ultimately judge them in ways no model currently can. Incident response decisions that affect operational continuity, especially in a defense environment, require human command authority. There is no automation of accountability.
Vendor and partnership decisions remain human. Assessing trust, value alignment, and long-term organizational fit requires human judgment that goes well beyond what AI can reliably evaluate today.
And governance itself remains human. I am deeply cautious about any organization allowing AI to govern AI. People who can be held responsible for the outcomes must own the oversight layer, ethical checkpoints, and decisions about where AI should and should not operate.
My operating principle is this: I trust AI to expand the surface area of human judgment, never to replace judgment itself. The moment an organization confuses the two, it has not embraced AI. It has simply outsourced its accountability.
The Positive Results of AI Integration
On the positive side, AI integration in our workflows most measurably improved program review cycle efficiency. After introducing AI-assisted synthesis into our governance process, leadership spent roughly 60% less time reconstructing project status from raw documentation per review session. That is not a trivial number in an environment where senior decision makers operate across multiple concurrent programs simultaneously. We effectively gave back hours of executive attention per cycle, redirecting it toward judgment and intervention rather than status reconstruction.
The second measurable outcome was risk detection. As I mentioned earlier, we identified two significant program risks earlier than our traditional review process would have. In both cases, the window between early detection and a material impact on delivery was measured in weeks. In a commercial environment, this translates to cost avoidance. In a defense environment, it translates to operational readiness — a harder number to quantify but a far more consequential one.
In strategic planning, AI integration reduced the time to produce first-draft technology assessments and roadmap scaffolding by approximately 70%. That does not mean the final output required 70% less effort. It means human effort shifted from generation to judgment, from writing to editing, from synthesis to validation. The output quality improved because senior minds spent their time on what required their expertise rather than what was mechanical.
The Negative Results of AI Integration
Now for an honest account of what has worked less well than expected — because the technology leadership community is poorly served by stories that only celebrate wins.
The first challenge was adoption. In a large institutional environment with deeply embedded operating cultures, introducing AI into workflows is not a technology problem. It is a change management problem. We had programs where we deployed the tools, granted access, and the capability sat largely unused because we had not redesigned the human systems around it — the incentives, habits, and trust — to accommodate it. Technology adoption without behavioral adoption is just expensive shelfware.
The second challenge was output quality consistency. In early deployments, we saw significant variance in AI-generated analysis quality depending on the input quality and structure. Garbage in, garbage out is not a new principle, but AI makes garbage look more authoritative. A poorly structured prompt producing a well-formatted but analytically weak output is, in some ways, more dangerous than a blank page because it creates false confidence. We learned to invest heavily in prompt discipline and input governance before trusting the outputs in high-stakes contexts.
The third and most important lesson was about expectations. Organizations that frame AI integration as a cost-cutting exercise tend to see underwhelming results because they optimize for the wrong outcome. Organizations — including our own programs — that have seen genuine impact framed AI as a capability amplifier. The question was not, “How do we do the same work with fewer people?” It was, “How do we do work that was previously impossible, or do it at a quality level previously unachievable with our resources?"
That reframing is the difference between AI as an efficiency tool and AI as a transformation lever. And, in my experience, only the second version produces outcomes worth talking about.
Why AI Integration Falls Short in Complex Environments
The gap between what AI promises in a boardroom presentation and what it delivers inside a complex institutional environment is still significant, and pretending otherwise does the profession a disservice.
AI has most clearly underdelivered for us in cross-system integration intelligence. We reasonably expected, based on what the technology demonstrated in controlled environments, that AI could serve as an intelligent orchestration layer across our enterprise platforms, understanding the relationships between systems, identifying integration failures before they cascaded, and recommending corrective actions in real time.
We discovered that AI performs extraordinarily well inside well-defined, well-documented domains, and extraordinarily poorly across the messy, underdocumented, politically complex boundaries between enterprise systems in large institutions. No document contains the knowledge required to navigate those boundaries. It lives in people, in relationships, in institutional memory accumulated over years. AI had no access to that knowledge and no reliable way to acquire it at the pace we needed.
AI has also underdelivered in natural language interfaces for non-technical users. The theory was compelling: give operational staff and senior leaders a conversational interface to enterprise data, eliminating the dependency on technical intermediaries for every query and report. In practice, the gap between how non-technical users naturally express their intent and how AI systems reliably interpret that intent proved far wider than anticipated. Users became frustrated when outputs did not match expectations, and frustration in a large institution does not produce feedback loops. It produces abandonment. We ended up with a capability that worked well for technically literate users and poorly for the broader population it was designed to serve, which is precisely the opposite of what transformation requires.
The third area is strategic document reasoning. We expected AI to work meaningfully with the long-form, highly contextual, often ambiguous policy and strategy documents that form the backbone of how large government institutions make decisions. We found that AI can summarize these documents competently, but it struggles to reason across them, to identify genuine contradictions between a policy written in one directorate and a standard adopted in another, to understand the political weight behind a particular clause, or to recognize when a technically correct recommendation is institutionally impossible. That layer of judgment, which makes strategic advice useful rather than merely accurate, remained stubbornly human.
Finally, I want to name something that rarely appears in vendor case studies: AI has not delivered the speed of organizational learning we anticipated. We assumed that as the organization adopted AI tools, it would develop collective intelligence faster, and that insights surfaced by AI in one program would naturally propagate and improve decision making across other programs. What happened is that AI insights remained siloed within the teams that generated them, because organizational structures, knowledge-sharing mechanisms, and cultural norms around documentation and transparency had not evolved to capture and distribute what AI was producing. The technology generated more signal than the organization knew how to absorb.
I share all of this not as a critique of AI's potential, which I believe remains profound, but as a reminder that every gap I just described is a human systems problem wearing a technology costume.
An AI Workflow for Optimizing Strategic Program Assessments
AI makes no single decision; it processes, synthesizes, and surfaces. Every judgment, every validation, every recommendation remains human.
I'll walk you through a workflow that is one of the most valuable in my current practice: AI-augmented strategic program assessment. I will describe it end-to-end because its value lies not in any single step, but in how the steps connect.
This workflow typically triggers when a program reaches a critical decision gate. In our environment, that might be a major enterprise platform migration, a digital health system rollout across a network of military hospitals, or an enterprise architecture review tied to a Target Operating Model redesign. These are programs where the decision carries multi-year consequences, and the inputs are voluminous, heterogeneous, and politically loaded.
The workflow begins with structured input aggregation. Before AI touches the analysis, my team deliberately curates inputs: gathering program documentation, vendor proposals, technical architecture submissions, previous review findings, relevant policy frameworks, and stakeholder position papers into a structured corpus. This step is entirely human and entirely intentional. The quality of AI's downstream output is directly proportional to the quality and structure of its upstream input. We learned this lesson the hard way in early deployments and now treat input governance as a non-negotiable discipline.
Next, we perform AI-assisted synthesis. We pass the curated corpus through a language model with a carefully designed prompt architecture that asks specific analytical questions, not for open-ended summarization. We do not ask the model to tell us what the documents say. Instead, we ask it to identify contradictions between stated objectives and proposed architectures, surface assumptions asserted but not evidenced, flag dependencies appearing in one document but absent from another, and highlight risk patterns matching profiles we have seen fail in previous programs. The output is a structured analytical brief, not a summary.
The third and most human step is critical review and recontextualization. A senior advisor — in most cases, me or a trusted technical lead — reads the AI-generated brief not as a conclusion but as a starting point for interrogation. We ask: What did the model surface that we had not consciously articulated? What did it miss that we know from institutional experience? Where is it technically correct but institutionally naïve? Here, nineteen years of operating inside this environment adds value no model can replicate. The AI brief sharpens the questions. Human judgment provides the answers.
The fourth step involves decision preparation. We use the validated insights from the human review to construct the executive decision package: framing the decision, outlining options with their genuine trade-offs, detailing risks with realistic likelihood and impact, and presenting the recommendation with its explicit assumptions. Because AI synthesis handles the mechanical work of processing the corpus, this step is faster and richer than a purely manual process would be. Leaders arrive at the decision gate with a clearer picture and more focused questions.
The fifth step is outcome capture, where most organizations drop the ball. After the decision and program move forward, we document not just what was decided, but also what the AI analysis surfaced that proved accurate, what it missed, and what assumptions it made that turned out to be wrong. This feedback does not go into a folder; instead, it goes back into our prompt architecture and input governance standards for the next program. The workflow improves with each iteration because we treat every run as a learning event.
This workflow has had a meaningful, measurable impact. It has reduced assessment cycle time for major program reviews by approximately 65%. It has improved the depth of risk identification, surfacing issues earlier in the decision process rather than discovering them during execution. And perhaps most importantly, it has elevated the quality of executive conversation at decision gates because leaders spend their time on genuine judgment calls rather than reconstructing status from raw information.
I want to emphasize this workflow's design philosophy. AI makes no single decision; it processes, synthesizes, and surfaces. Every judgment, every validation, every recommendation remains human.
How to Avoid Pitfalls in AI Adoption
The organizational immune system is real, powerful, and activates faster than any AI tool deploys. Every large institution has developed, over years and decades, informal mechanisms that resist change that threatens existing power structures, established workflows, and professional identities. I understood this intellectually before we began, but I underestimated how specifically and quickly those mechanisms would activate in response to AI.
When people sensed AI tools were changing not just how work was done but who held value in the organization, resistance was not expressed as opposition, but as compliance without adoption. Teams accessed tools but did not use them. They received outputs but did not act upon them. Processes were nominally updated but behaviorally unchanged. Passive resistance in a large institution is almost invisible until you measure outcomes. By the time measurement reveals the problem, months of momentum are lost.
On a related note, I wish I had known how much the rollout's framing would permanently shape its reception. In our early deployments, we made a communication error that I consider one of the initiative's most consequential mistakes. To build executive sponsorship and justify investment, we framed AI tools in the language of efficiency: faster, leaner, more productive.
That framing was accurate in a narrow technical sense but deeply counterproductive in every human sense. When people hear efficiency in the context of a new tool rolling out to their team, they hear headcount. They hear their role evaluated for redundancy. They engage with the tool not as something that will make their work better, but as something that might make their position unnecessary. The defensive posture that framing created took months to dismantle and never fully dissolved in some parts of the organization.
And personally, I now know that I should have been more transparent with my teams about my own uncertainty. I entered this journey with more confidence in my directional conviction than in my operational knowledge of how AI would behave within our specific institutional context. I knew the destination was right. I did not fully know the path. In my effort to project the leadership confidence that large programs require, I sometimes communicated certainty I did not possess, which created expectations that reality then complicated. I have learned that in genuinely novel territory, intellectual honesty about uncertainty is not a weakness of leadership. It is a prerequisite for building the trust that sustains teams through the inevitable moments when the path proves harder than the plan anticipated.
Why AI Requires a New Type of Technology Leader
The AI era systematically devalues a particular profile of technology leader: the expert who accumulated deep technical knowledge over years of disciplined study and practice, built authority through mastery, and led as the most knowledgeable person in the room. That profile dominated technology leadership for decades and produced extraordinary results when knowledge was the bottleneck and expertise the competitive advantage. AI does not eliminate that profile overnight. But it erodes its foundations faster than most people in that profile acknowledge, and institutions that depend on that profile without developing its successor accumulate a leadership risk they have not yet registered.
The profile the AI era requires is fundamentally different. The AI era requires leaders who derive authority not from what they know but from the quality of their questions. Leaders whose primary intellectual contribution is not synthesis (because AI can synthesize), but judgment. Leaders fluent in organizational dynamics, human psychology, and institutional power, as well as technology. Leaders who maintain genuine intellectual humility in the presence of a tool that constantly projects confidence, and who have developed the internal stability to say "I do not know" and "I need to think about this" in environments that reward decisiveness above almost everything else.
And perhaps most importantly, the AI era requires leaders who clearly and thoroughly answer a question previous generations of technology leaders rarely needed to ask: What human value am I adding that AI cannot replicate, and am I investing in that value with the same discipline I bring to my technical development?
I ask myself that question regularly. It is not a comfortable question. It requires honest self-assessment about which parts of my professional identity are genuinely irreplaceable and which parts I protect out of habit rather than necessity. I believe that process of honest self-examination is the foundational leadership practice of the AI era. Not prompt engineering, not model evaluation, not AI governance frameworks — important as they are. It is the willingness to look clearly at what you bring that a model cannot, and to invest deliberately and relentlessly in exactly that.
The AI era produces complexity at a rate and scale with no historical precedent in the technology profession. Leaders who will navigate it well do not adopt AI fastest, deploy it most ambitiously, or talk about it most fluently on stages like this one. They do the harder, quieter work of developing the human qualities complexity demands: judgment, humility, courage, institutional wisdom, and the clarity of purpose that allows them to direct powerful tools toward worthy ends.
I believe the technology leadership community needs to have that conversation with far greater urgency than it currently does. And it is the reason I agreed to this interview: not to talk about what AI can do, but to think carefully, in public, about what we as leaders must become to be worthy of directing it.
Five Pieces of Advice for CTOs
I want to give advice that is useful rather than advice that sounds good on a conference stage: First, resist the pressure to have an AI strategy. Second, treat adoption as the primary engineering challenge, not integration. Third, protect your judgment layer fiercely. Fourth, pay serious attention to sovereign AI ecosystems…Fifth and finally, invest in your own judgment as a leader with the same discipline you invest in your organization’s technical capability.
I want to give advice that is useful rather than advice that sounds good on a conference stage, so let me share what I believe based on nineteen years of leading transformation inside one of the most complex institutional environments imaginable.
First, resist the pressure to have an AI strategy. Your organization needs a business strategy that AI serves. I have watched too many technology leaders spend enormous energy producing AI roadmaps, standing up AI centers of excellence, and announcing AI initiatives, yet they leave the fundamental question of what problem they are solving for the institution unanswered. AI without strategic clarity is not transformation; it is sophisticated activity masquerading as progress. Start with the outcome you are accountable for delivering, and work backward to where AI creates the most leverage. That sequence matters enormously.
Second, treat adoption as the primary engineering challenge, not integration. The hardest problem in AI deployment is not connecting the model to your systems; it is changing how your people think, decide, and work in the presence of AI assistance. In my experience, organizations underinvest in this by an order of magnitude. They allocate 80 percent of their budget to technology and 20 percent to people and process change, when the ratio that produces results is closer to the reverse. If your AI initiative does not have a serious change management and capability-building component, you are building on sand regardless of how sophisticated your architecture is.
Third, protect your judgment layer fiercely. As AI becomes more capable and more embedded in workflows, constant pressure will emerge, subtle and not so subtle, to extend its authority: to let it not just inform decisions but make them, not just flag risks but resolve them, not just draft recommendations but approve them. Resist this. Not because AI cannot be technically capable of these things in certain domains, but because the moment your organization loses the habit of human judgment, it loses something extraordinarily difficult to rebuild. Judgment is a muscle. Organizations that stop exercising it do not remain static; they atrophy. In a high-stakes environment, that atrophy costs you at the worst possible moment.
Fourth, pay serious attention to sovereign AI ecosystems, particularly if you operate in or adjacent to government, defense, or critical national infrastructure. Question the assumption that global hyperscale platforms will remain the default architecture for enterprise AI. In Saudi Arabia, we are witnessing the emergence of genuinely sovereign AI infrastructure through platforms like HUMAIN ONE. These platforms are built on Arabic-first language models, hosted on national compute infrastructure, and designed specifically for the governance complexity of large institutions. For CTOs operating in regulated or security-sensitive environments, the question of where your AI runs, on whose infrastructure, under whose jurisdiction, and trained on whose data, is not a compliance checkbox; it is a foundational architectural decision that will define your organization's AI posture for the next decade.
Fifth and finally, invest in your own judgment as a leader with the same discipline you invest in your organization's technical capability. This is the advice I consider most important and least discussed in technology circles. The CTOs who navigate this moment well do not necessarily understand AI most deeply in a technical sense. They combine technical literacy with institutional wisdom, strategic clarity, and the moral courage to make decisions that are right for their organizations even when they are unpopular in the short term. AI will keep advancing regardless of what any of us does. The variable is the quality of human leadership directing it.
We are not in a race to adopt AI fastest; we are in a race to adopt it wisest. In my experience, wisdom has never been a feature you can install.
Follow Along
You can follow along with Abdulrhman AlOdhayb's work on LinkedIn.
More expert interviews to come on The CTO Club.
