AI and engineering data
Why XKEMA Does Not Start With AI
Infrastructure before inference. The order is a design decision, not a limitation.
Every hardware tool bought in the last two years has an AI feature. Ours does not, and the reason is not caution for its own sake.
A language model answering questions about your hardware is only as good as what it can read. If what it can read is a folder of PDFs, a spreadsheet somebody renamed in 2023 and a schematic export nobody can tell is current, then the model is not reasoning about your product. It is guessing about your filing.
The position. Infrastructure before inference. Structured engineering objects, the relationships between them and a reliable notion of which version is authoritative come first, because they are what any future automated reasoning would have to stand on. This is a sequencing decision, not a claim that AI is useless in hardware.
The question a model cannot answer today
Take a question a hardware engineer asks weekly:
Which released products use this voltage regulator, which revision of each is current, and is there an approved alternative that has already been through qualification?
Three facts are needed, and none of them lives in a document:
- Usage. Which products contain the part. This is a relationship between
a component and several BOMs, each belonging to a board, each belonging to a product.
- Authority. Which revision of each product is the released one. This is
a property of a configuration, not of a file.
- Approval. Whether an alternative exists and who accepted it. This is a
decision with an owner and a date.
A model with access to every file in the company can produce a fluent answer to that question. What it cannot do is produce a correct one, because the relationships are not written down anywhere it can read them. They live in the heads of three people and in the layout of a shared drive.
Why the order matters
The software industry solved this by accident. Source code arrived in systems that recorded, as a side effect of normal work, what changed, when, by whom and against what baseline. The dataset that makes software assistants useful was not built for AI. It was built for teamwork, and AI inherited it.
Hardware did not get that by accident. The outputs of hardware design are documents and fabrication files, and the systems that hold them are document systems. They store the artefact and lose the relationship.
So the order is forced:
- Give the engineering objects an identity: component, product, board, BOM
row, version, document.
- Record the relationships between them, including which product uses which
part at which revision.
- Make released state explicit, so the difference between a draft and a
baseline is a fact and not a filename convention.
- Then, and only then, automated reasoning has something to reason over.
Skipping to step four produces a confident assistant on top of an unreliable dataset. In hardware the output of that is not a bad paragraph. It is a production run of the wrong revision.
What this is not
Three honest clarifications, because this argument is easy to overstate.
- This is not a claim that AI cannot help hardware teams. Plenty of
narrow tasks already work: reading a datasheet table, drafting a test procedure, converting one file format into another.
- This is not a claim that XKEMA has AI features. It does not. The
argument here is about what would have to exist first.
- This is not a criticism of the tools you already use. ECAD tools are
good at design capture. ERP systems are good at transactions, inventory and purchasing. Neither was built to answer the question above, and expecting them to is the actual mistake.
What a good hardware team should know
The useful test is not "do we have AI". It is:
- Can one person answer the usage question above without asking anyone?
- Does the answer come with a revision, or just a part number?
- Is there a difference, visible to a tool, between a draft and a released
configuration?
- When somebody approves an alternative, is that decision stored as a
decision, or as an email?
A team that can answer those four has a dataset. A team that cannot has documents, and documents are a poor substrate for any kind of automation, human or otherwise.
Where structured engineering data helps
XKEMA organises the engineering context between an ECAD tool and an ERP: components, products, BOMs, versions and the relationships between them, with released state recorded explicitly rather than inferred from a filename.
The immediate benefit of that is not AI. It is that the question at the top of this note becomes answerable by a person in one place. Whether that same structure later makes automated reasoning more reliable is a reasonable expectation and a design direction, not a feature that exists today.
Sources
- NIST AI Risk Management Framework 1.0 (NIST AI 100-1) — the validity and reliability characteristics an AI system is expected to demonstrate, and why they depend on the data and the conditions of use.
- NASA Systems Engineering Handbook Rev 2, section 6.5 — configuration management as the discipline that makes a product's state knowable at all.