Short answer: For a defined equipment family, AI can now read a gas-line diagram, pick released parts from your vault, place and connect them under your rules, and produce a native SOLIDWORKS assembly you can open and edit. It can't decide your design rules, approve a part that isn't in PDM, settle a diagram that is genuinely ambiguous, or sign the release. The useful line runs between applying rules and making judgments.
Start from the work, not the demo
Look at a normal week on a gas-equipment design team. A new order comes in with a gas-line diagram: several sticks, each with a manual valve, regulator, pressure transducer, filter, flow controller and pneumatic isolation valves, plus purge and vent branches. Someone opens the previous similar job, uses Pack and Go to copy it, and then spends days swapping components, re-mating them, re-routing tube, fixing broken references and reconciling the BOM against the diagram. Almost none of that is new engineering. It is careful transcription of decisions the team has already made many times.
That is the work to measure AI against. Chat assistants that write a macro and generative tools that sculpt a bracket are interesting, but they don't touch this problem. The question is whether a system can do the transcription part reliably, inside rules you control, with output you can check line by line.
What it does well today
Reading the diagram as connections, not text
Running OCR over a diagram gives you a list of tags. It doesn't tell you that a check valve points the wrong way for a branch, or that a purge line only connects when a given valve is open. What helps downstream is a graph: components as nodes, lines as edges, with flow direction and conditions attached. MST's product page describes recovering tags, connections, flow direction and conditions from the P&ID into a directed engineering graph. Once you have that graph, every later step can be checked against it. Each stick either has its regulator upstream of the transducer or it doesn't.
Placing released parts, not drawing new ones
A good system places the part your vault already holds: the regulator with the correct part number at the approved revision, with the mate references your library engineer built. MST describes binding approved PDM models to their part identities, revisions and actual connection interfaces. The practical effect is that the BOM coming out of the assembly is made of things purchasing can actually order.
Solving placement and routing together
Anyone who has laid out a gas box knows that placement and routing can't be separated. Moving a filter along the stick changes which tube runs are possible, and a route that needs an awkward compound bend may mean the component should have been rotated. MST's mechanical assembly AI treats this as a coupled problem. A layout solver chooses components, positions, orientations and interface alignment. A routing solver searches paths, fittings and bends between compatible ports. Collision, interface compatibility, bend rules, access and mounting are hard constraints, while things like envelope size, route length and fitting count are objectives you approve.
Producing a file you can keep working on
MST's core statement is specific: P&ID gas-line diagram + approved PDM parts + equipment rules -> native, editable SOLIDWORKS assembly, starting with semiconductor gas equipment. The output is a .SLDASM built from your own native parts. MST then reopens it and reads back components, transforms, mates, rebuild state, routes and traceability so they can be checked against what the task required. See the [MST AI Engineer product page](https://mst-us.ai/product/).
What it can't do
Define the rules
Whether a service break goes on both sides of the flow controller, whether a bypass is welded or uses face seals, how much wrench clearance you keep around a VCR nut, which side the pneumatic actuators face for tubing access: these are your company's decisions, and many of them exist because something went wrong once. A system can apply them consistently once they are written down. It can't come up with them. MST states that a new equipment class needs its own engineering rules and validation, developed with the customer.
Fill gaps with invented geometry
If the diagram calls for a component that has no released model, the correct output is a flagged gap. A placeholder dressed up as a real part is wrong. A generated part has no part number, no supplier and no revision history, and it will get past a hurried reviewer. [Q27](https://mst-us.ai/questions/ai-solidworks-assembly-from-gas-line-diagram/) covers how unresolved items are meant to surface.
Resolve ambiguity by guessing
Diagrams carry mistakes: a line that crosses another with no junction dot, a tag that appears twice, an instrument whose connection side isn't shown. A senior engineer phones the process engineer. A system should flag these. If it resolves them silently, it gives you a confident-looking model built on an assumption nobody approved.
Own the release
MST's launch note is explicit. It doesn't claim an unconditional global optimum, arbitrary whole-machine autonomy, universal CAD/PDM support or autonomous production release, and engineers keep approval authority ([launch announcement](https://mst-us.ai/mst-launches-mechanical-assembly-ai/)). In PDM terms, the generated assembly enters your workflow the way a junior engineer's work does. It doesn't move itself to Released.
Scenario: one gas box, two paths
A design engineer receives a diagram for a gas box that is close to a previous job but has an extra stick and a different flow controller model.
Manual path. Pack and Go the old assembly. Delete the stick that no longer matches, copy another, replace the flow controller with a model whose port spacing differs, and re-mate. Several concentric and coincident mates fail, and the fittings downstream need repositioning. Rebuild the tube runs, recheck clearance to the enclosure wall, then go back over the BOM against the diagram and catch a filter left over from the old configuration.
AI-assisted path. The diagram, the approved part subset and the written rules for this gas-box family go in. A native assembly comes out with each stick placed and routed, plus a read-back that lists components, mates and connections and flags one tag that matched no released part. The engineer reviews the routing near the new flow controller, adjusts one run for better wrench access, resolves the flagged part with the library owner, and sends the assembly to review.
The engineer's judgment is still needed in the second path. What changes is that the engineer spends it on access and on one genuine gap, not on re-mating.
Where the line sits
| Task | Suitable for AI today | Stays with the engineer |
|---|---|---|
| Turning diagram connections and flow into a graph | Yes, for supported diagrams | Ambiguous symbols and missing information |
| Choosing which released part to place | Yes, from the approved subset | Creating or approving new parts |
| Placing components and routing tube under rules | Yes, as a coupled search | Writing the rules and setting objectives |
| Building the native assembly | Yes, with read-back | Design review |
| Release in PDM | No | Always |
Tradeoffs to expect
- Up-front work moves earlier. Library cleanup and rule capture come before the first good result. That is real effort, and the team will spend it whether or not the tool earns its place.
- Narrow scope works better than broad scope. A system tuned for one equipment family is more dependable than a general tool, but it won't cover your one-off machines.
- The review changes rather than disappears. You check a complete proposal against a list of facts instead of building from scratch, and the habit of checking has to stay sharp.
A quick self-check
- Do we have a repeating assembly family with mostly released parts?
- Could a senior engineer write the layout and routing rules down?
- Are revisions and mate references in the vault clean enough to trust?
- Who would review and release a generated assembly?
If the answers are mostly yes and your team designs semiconductor gas equipment in SOLIDWORKS, [contact MST](https://mst-us.ai/contact/) with a short, non-confidential description of one representative task.
FAQ
Is this the same as generative design?
No. Generative design creates new part geometry against loads. Assembly AI as described here arranges and connects existing released parts.
Does it replace CAD engineers?
No. It takes on repetitive placement, mating and routing. Engineers write the rules, review the result, correct it and release it.
Will it open in normal SOLIDWORKS?
Yes. MST's output is a native, editable SOLIDWORKS assembly.
Can it handle any machine?
No. MST starts with semiconductor gas equipment, and a new class needs its own rules and validation.
Is the layout "optimal"?
Only against the declared objectives, constraints and search scope. MST doesn't claim a universal optimum.
What if a part is missing from the vault?
It is reported as unresolved, and the library owner and design lead decide what to do.