2.4 The AI stack: five layers of engineering intelligence

The previous section named the six mechanisms — the classical trio that computes (R, P, O) and the AI trio that learns (K, ML, L). Knowing what kind of mechanism a capability is only gets us halfway. We also need to say where in the engineering work it sits — because "AI for moulded parts" can mean anything from drafting a quotation to predicting warpage, and those are entirely different problems.

For that we use a simple five-layer model of engineering intelligence. It breaks the vague phrase "an intelligent system" into five separate functions, each of which can be present or absent, and each of which can be served by a classical or a learned mechanism.

2.4.1 The five layers

1. Intent — turning a requirement into an engineering task. Moulding example: reading an incoming request for quotation and drafting the engineering response. This is language work, so the natural mechanism is an L (language model).
2. Knowledge — making what the organisation already knows findable. Example: matching a new part's geometry against past mould projects and surfacing their gate designs and outcomes. This is the home of K (retrieval) — and, as §2.3 noted, the place where decades of CAD archives usually sit unused.
3. Reasoning — predicting the consequences of a design decision before committing to it. Example: predicting whether a geometry will warp. Today this layer is owned by P (physics solvers); the emerging challengers are ML surrogates and neural solvers that try to do the same prediction faster.
4. Execution — actually doing the work. Example: laying out a mould base and ejector pattern, or generating routine CAD from a template. This is the most mature layer, and it is owned honestly by R (rules and templates) and by people — not by AI.
5. Feedback — learning from what actually happened in production, so the next design starts wiser. Example: feeding measured warpage and scrap back to the engineer who chose the geometry. This is the hardest layer, for exactly the reason §2.3 gave: the measured outcome and the geometry live on opposite sides of the workflow's hand-offs, and rarely meet.

2.4.2 Why the layers are worth the trouble

The value of the stack is that it turns a vague question — "is this company using AI for moulded parts?" — into five precise ones: which layer, served by which mechanism, learning from what? A copilot that drafts quotations (Intent, via a language model) and a surrogate that predicts warpage (Reasoning, via machine learning) are both "AI", but they are not remotely the same capability, and lumping them together is how marketing claims survive.

The stack also makes the shape of the field visible at a glance. Two layers are comparatively easy: Execution has been automated with rules for decades, and Intent rides on general-purpose language models that any vendor can adopt. Feedback, at the far end of the loop, is hardest of all, for the reason §2.4.1 gave. And the hard, valuable middle — Knowledge and Reasoning, exactly where moulding-specific learning would have to live — is the part that depends on the scarce, label-hungry data of §2.3. Whether that middle is full or empty in practice, company by company, is the question the rest of this study is built to answer.

2.4.3 What this chapter gives the rest of the study

We now have the complete lens:

• the honest-AI test — learn from data, or compute from equations?
• six mechanisms — R, P, O that compute; K, ML, L that learn.
• five layers — Intent, Knowledge, Reasoning, Execution, Feedback.

Every tool, product and claim in the chapters ahead is examined with these three instruments. Chapter 1 gave us the part and the process; this chapter gave us the vocabulary for the intelligence applied to them. With both in hand, we can turn to what companies have actually built.