A Prompt Does Not Allow You to Skip DIKW

Many AI-related posts create the impression that expertise can be injected into a model through a prompt. The magic happens after you write:

• “Act as a procurement expert.”

• “Act as a CFO.”

• “Act as a strategy consultant.”

If expertise could be created through prompting, organisations using the same models and similar prompts should produce broadly similar recommendations. 

It's not the prompt.

In practice, some organizations obtain useful outputs, while others receive generic observations, robo-consultant-style recommendations, and conventional wisdom. 

The difference is usually not the prompt; it's what's feeding it.

The flowchart illustrates the DIKW (Data, Information, Knowledge, Wisdom) hierarchy, emphasizing that prompts initiate reasoning but do not replace the need for expertise. Starting with data intake, the process involves deterministic logic to transform data into information, followed by a procurement knowledge system to develop expertise, and a commercial decision framework to develop wisdom. The AI reasoning engine then generates insights that lead to executive reporting and action. The key takeaway is that prompts can access but not create data, information, knowledge, or wisdom.

The diagram I asked ChatGPT to create based on a real AI-enabled solution illustrates the point: 

1. The reasoning engine appears near the end of the process rather than at the beginning. 

2. Before any reasoning occurs, data must be collected and structured. 

3. Information must be generated from that data. 

4. Knowledge must be accumulated through experience, frameworks, methodologies, and organizational learning. 

5. Only then can judgement be applied.

This is not a new idea. It is simply the DIKW hierarchy:

Data → Information → Knowledge → Wisdom

A prompt does not allow bypassing these stages.

A procurement team without category strategies, supplier intelligence, governance models, and sourcing experience does not become an expert simply by asking AI to behave like a category manager. A finance function does not acquire commercial judgement because it instructs AI to act as a CFO. 

The model may generate language associated with those roles, but generating language and exercising judgement are not the same thing.

Prompt engineering and diminishing returns.

This distinction explains why prompt engineering often produces diminishing returns.

Organizations continue to refine instructions while the real constraint lies elsewhere. Missing knowledge or weak decisions cannot be solved through better prompting. 

The most valuable asset is therefore not a prompt; it is a knowledge system behind the prompt.

That is why the debate should be less about prompt engineering and more about knowledge engineering. A prompt can initiate reasoning but It cannot leapfrog the journey from data to wisdom.