In part 1, I showed a Value Chain for implementing AI. It covered the components needed, from the user's need down to cloud infrastructure. Now I’m going to add evolution to turn it into a Wardley Map, so you can see how to implement AI.
Evolution is the core innovation of Wardley Mapping. It shows how everything evolves — from the uncharted space on the left of the map to the industrialised on the right. The position of a component on this axis tells you how to manage it:
Far left — high uncertainty. An exciting, potential source of future value, but requires experimentation.
Far right — high certainty. Low margin, cost of doing business but generates revenues from volume operations.
Middle — transitional. Demand here is growing and profits are available for those who serve user needs best. Also the space of more intense competition.
Wardley Map for AI Deployment
The first five components ‘above the line’ are visible to end users, so they care about them more. Therefore, focus on effectiveness, or ‘doing the right thing’ here. It’s better to do the right thing badly than not at all — that’s how you get better at doing what users value.
‘Testing AI safely' sits furthest left as uncertainty is highest here. A recent MIT report found that 95% of AI pilots fail to deliver ROI, while there have been some large scale and costly failures. There are no best practices to follow. You have to figure it out yourself. But if you do, you can unlock a new source of value.
Testing AI requires choosing a 'use case’. But you can’t choose the same use case as your competitor. You have a different 'knowledge base' (component #7) so the technology will not work the same for you, as different inputs produce different outputs. Therefore, this is positioned towards the left of the map, showing the uncertainty here.
Choosing a 'use case’ requires ‘defining the problem to solve’ first. However, fear of failure means most people in organisations avoid the hard ones and focus on the problems they can fix, not the ones they should.
Defining the right problem to fix requires a ‘discussion’. This should be straightforward, but most organisations lack maps or any way of seeing their landscape clearly, so struggle to have such discussions.
You will also need an ‘interface’ to interact with the AI technology. These are becoming standard products, so positioned centrally on the map.
The next five components below the line are less visible to users, so they care about these less. Here, the focus is on efficiency, or doing things right. The first of these components is ‘orchestration’ — an important component that, through RAG, puts the intelligence in AI. This is less visible to users so they understand and care about it less. They simply expect it to be there. The focus here therefore is on efficiency — providing something reliable at a reasonable cost. This is a well-evolved component with high certainty about it. Therefore it’s positioned towards the right of the map.
The components below orchestration, such as the ‘knowledge base’, are almost invisible to users and not things they care about. They’re also highly-evolved, with a high certainty and good practices for managing them.
‘Embeddings’ is in the industrialised space as there’s complete and widespread certainty about this. This leaves no room for providers to create competitive differentiation.
‘LLMs’ are also industrialised. Some may push back on this, pointing to new iterations of the main models and debates about which are better. But DeepSeek and other Chinese LLMs are now open-source. This means users will not choose to pay a premium for a US model when cheaper alternatives work almost as well.
At the bottom of the map is ‘compute’ — the infrastructure on which the entire AI revolution runs and where most investment (and profits) are being made with GPUs and data centre build outs.
This map reveals that the technology for AI is well established. This is not a fad that will fade away. Therefore, the widely reported failure rates for AI implementation are not because the technology is under-developed. It’s because organisational practices are. The issue is not the tech — it’s the lack of fit to the organisation.
Organisations have chased best practices for a long time — imitating what works elsewhere, hoping it works as well for them. Organisations can learn from others, but AI amplifies the importance of unique inputs —different inputs produce different results. The inputs here are your knowledge base. To use AI successfully, you need to find your own path.
So what's holding organisations back from choosing the right things to focus on? That’s what I’ll explore in the next part.