Most conversations about artificial intelligence start with the model: which one, how accurate, how expensive. In my experience, that is rarely where projects succeed or fail. Taking a computer-vision analytics platform from government-funded pilot projects to commercial deployments, and building workplace-safety AI for a multinational company across three countries, the hardest step was almost never the algorithm. It was getting a working system into the hands of people and processes that were ready to use it.
With colleagues from De La Salle University and Caraga State University, and partners in the Asia-Europe for Artificial Intelligence (AE4AI) Network from the United Kingdom, the Netherlands and Brunei, I set out to test that intuition more rigorously. Our paper in Technologies maps the AI value chain end to end and asks where value is created, and where it gets stuck.
Five layers, each built on the one before
We describe the AI value chain as five connected layers. The value created in each layer becomes the foundation of the next.
- Hardware. Chips, servers, data centers and cloud infrastructure. This layer depends heavily on a small number of global suppliers.
- Data management. Collecting, cleaning, labelling, governing and securing the data that models learn from.
- Foundational AI. General-purpose models and algorithms, often built by nation-states, large firms and research institutions at very high cost.
- Advanced AI capabilities. Turning foundational models into specialized, tested and ethically reviewed applications.
- AI delivery. Deploying AI into real workflows as products and services, including AI-as-a-service, and keeping them running.
Each layer has its own stakeholders, from component suppliers and data stewards to regulators, quality-assurance teams and the communities affected by AI decisions. The chain only works when these actors coordinate across technical, regulatory and social lines.
Where the chain breaks
We ran a SWOT-based bottleneck analysis for every layer. Some constraints are familiar. Hardware is resource-intensive and globally dependent. Data management still relies on manual annotation and carries privacy and security risk. Foundational models are costly to build and can pass inherited bias downstream.
The most important finding sits in the last layer. Infrastructure for deploying AI is increasingly scalable, yet the final integration of AI into business processes remains the main barrier to adoption. We call this the last-mile problem. It is not mainly a technical issue. It is a socio-technical gap: technology is advancing faster than organizations can absorb it. The decisive constraints are managerial competencies, digital skills, process redesign and resistance to change.
As an engineer who began in semiconductor design, I find this both humbling and useful. We tend to invest in the layers we can see and measure, such as compute, data and models. The return on that investment is lost if the delivery layer, and the people in it, are not ready.
What the Philippine case shows
We tested the framework on the Philippines as a representative developing economy. A bibliographic analysis of the country’s AI research shows two main sources of influence. Links to Europe and the Americas shape the foundational side of our work: core algorithms, machine learning models and established frameworks. Links to Asian and Middle Eastern peers shape the application and delivery side, in areas such as logistics, finance, agriculture and disaster management.
In other words, the Philippines is not building its AI value chain alone. It is absorbing foundational knowledge from global leaders and applying it to problems it shares with its regional neighbours. That is a sensible strategy, as long as we build the local capability to deliver.
What to do about it
The framework is meant to be used, not just cited. Here is how I apply it with different audiences.
- Companies. Locate yourself on the chain before buying technology. Most firms do not need to build foundational models. They need clean, governed data, a clear path from pilot to production, MLOps practices, and leaders and teams who can redesign work around AI. Measure time-to-deployment, not just model accuracy.
- Government agencies. A national or regional AI roadmap should address human-capital readiness alongside digital infrastructure and data governance. It should also manage dependence on foreign technology while growing local innovation where we have an advantage.
- Universities. Design interdisciplinary curricula and executive programs aimed at the skill gaps in the advanced-capability and delivery layers. This is the thinking behind the case-study-based training I deliver for industry and government.
The framework can also become practical tools: a risk-assessment checklist for each layer, a maturity scorecard, or a policy-impact dashboard. That is how I use it in AI readiness assessments and roadmapping work.
Five questions to ask before your next AI project
- Which layer of the value chain are we actually investing in, and which layers are we depending on others for?
- Is our data governed, labelled and secure enough to train and run the system we want?
- Are we building on a foundational model, and do we understand its limitations and inherited biases?
- Who will own the system in operations, and what processes must change for it to be used?
- Do our managers and staff have the digital skills to use, question and improve it?
If the honest answer to the last two questions is "not yet," that is where the next investment should go.