I think we’re seeing a pattern where AI is sometimes treated as a universal solution, when in reality it’s just another tool in the stack.

My rule of thumb is:

If the answer can be derived deterministically, let a deterministic system do it.

For tasks like CAD validation, financial calculations, schema validation, or compliance checks, traditional tools already provide precise and verifiable outputs. Replacing them with an LLM often adds uncertainty without adding much value.

Where RAG becomes powerful is when the challenge shifts from computing the answer to understanding the answer.

For example, a CAD validation engine might identify 27 geometry issues. A RAG system can then:

• Explain why those issues matter.

• Retrieve relevant standards or documentation.

• Surface similar historical cases.

• Suggest potential remediation paths.

In that setup, the traditional tool remains the source of truth, while the LLM acts as an interpreter and advisor.

The most effective AI systems I’ve seen don’t replace existing engineering tools, they sit on top of them, making their outputs more accessible and actionable for humans.

So rather than asking whether RAG can solve a problem, I think the better question is:

Which parts of this workflow require deterministic accuracy, and which parts require human-friendly reasoning?

That’s usually where the boundary becomes clear.

I think the real question isn’t whether RAG can be used for these problems, it’s whether it should be the primary decision-maker.

A good rule of thumb is:

If there’s already a deterministic way to verify something, let the deterministic system do the verification.

Take CAD validation, financial calculations, code compilation, or compliance checks. These domains have well-defined rules and established tools that can produce consistent, repeatable results. Replacing them with an LLM often introduces uncertainty where none previously existed.

Where RAG shines is in everything around the decision:

• Explaining why an issue occurred

• Retrieving relevant standards or documentation

• Surfacing similar historical cases

• Helping users navigate complex systems

• Translating technical outputs into actionable recommendations

In other words, I see RAG as a reasoning and communication layer, not a validation layer.

The most successful architectures I’ve encountered don’t choose between AI and traditional tools. They combine them:

1. Traditional systems generate the facts.

2. Retrieval provides context.

3. The LLM helps humans understand and act on the results.

When AI is used to augment expertise rather than replace deterministic processes, the outcomes tend to be both more accurate and more trustworthy.

I think the mistake isn’t using RAG for these problems, it’s expecting RAG to replace the traditional tools that already solve them well.

To me, RAG is strongest when the challenge is finding, synthesizing, and explaining information. Traditional engineering tools are strongest when the challenge is computing, validating, or enforcing rules.

Using the CAD example:

• A geometry engine should determine whether a model has non-manifold edges, invalid surfaces, or tolerance violations.

• A rule-based system should check manufacturing constraints.

• A RAG system can help explain why the issue matters, surface relevant standards, retrieve similar past cases, or guide an engineer toward a fix.

In that sense, RAG becomes a layer on top of deterministic systems rather than a replacement for them.

I’ve seen a similar pattern in other domains. The most successful AI applications often combine:

1. A deterministic system that generates trustworthy facts.

2. A retrieval layer that provides context.

3. An LLM that translates those results into something humans can quickly understand and act upon.

When we ask an LLM to perform tasks that already have precise algorithmic solutions, we sometimes trade accuracy for flexibility without gaining much value.

The more interesting question might be:

Where is the boundary between “reasoning about the output of a system” and “being responsible for generating the output itself”?

In my experience, AI tends to create the most value on the reasoning and communication side of that boundary, while traditional tools remain the source of truth.