Can engineering documentation be automated?

A released part is not a model. It is a package. The model is one file in it, and on most jobs it is the file that took the least time.

And behind all of it, the engineering report: the verdict against each requirement, the analysis that produced it, every assumption made and why, and provenance for every figure. That is what a reviewer signs, and it is the part most tools do not produce at all.

Designing a bracket is an afternoon. Releasing one is a week, and the week is not spent on the shape.

It goes on placing views so nothing overlaps, chasing a dimension that points at the wrong edge, deciding which characteristics are critical, rolling a parts list up through three levels of assembly, finding the standard-part numbers, and writing the note that explains why the wall is 3mm and not 2.5mm. None of it is difficult. All of it is exacting, and every piece of it has to agree with every other piece.

If a machinist has to ring you to ask what something means, the drawing failed. That is a low bar and a great deal of work goes into clearing it.

This is also why so much of it gets skipped or done at the last minute, and why the parts that never get analysed are usually the same parts whose paperwork is thinnest.

The specific failure that automation has to solve is not slowness. It is drift: the paperwork and the part describing two different objects.

It happens the ordinary way. The model is revised after the drawing was made. A tolerance is tightened on the sheet but not in the analysis that justified it. The parts list is exported before the last fastener change. Nobody is careless; the documents are simply produced by different people, from different sources, at different times.

A generated set cannot drift, but only if it is generated from the same run that produced and proved the model. A tool that reads a finished model and draws it has removed the drafting hours and left the drift exactly where it was, because the model it read may already disagree with the analysis nobody re-ran.

That is the difference worth paying for, and it is a question about architecture rather than about output quality.

A generated drawing is judged the same way a drawn one is. It should be possible to hand it to a shop with no covering note.

• Every field is filled. Material, finish, general tolerance, projection, scale, revision, who released it and when.
• Every dimension points at something real. Taken from the model's own topology, not typed, so a revision moves them.
• No two views sit on top of each other, and section arrows point the way the section was cut.
• Fits are resolved to limits. H7 on the sheet is convenient for the designer and useless to the inspector.
• Critical characteristics are marked, and marked because the analysis depended on them rather than because they looked important.
• The revision block says what changed, not that something did.

Documentation is the part of the job with the longest life. A part released this year gets remade in eleven years by a shop that has never heard of the tool that drew it.

So the output has to be in formats that every system reads and nothing is required to open: STEP for the geometry, PDF and DXF for the sheets, CSV for the lists. Those open in SolidWorks, Onshape, Fusion, Creo, and in whatever replaces them.

The practical test is what happens when you stop paying. If the software stops and the files stop with it, you were renting your own drawings.

• Ask where the drawing comes from. The same run that proved the model, or a separate pass over a finished file? Only the first cannot drift.
• Ask to see a real sheet, not a render. Views, sections, a filled title block, a revision entry.
• Ask what happens on a revision. Does the whole set regenerate, or does someone reconcile it?
• Ask how critical characteristics are chosen. If the answer is a rule of thumb, the inspection sheet is decoration.
• Ask whether every figure in the report traces back. Pick a number and ask which run produced it. If nobody can say in under a minute, the report is prose.
• Ask what you keep. Neutral formats, or a proprietary bundle.

Drawing automation exists and some of it is good, but almost all of it starts from a finished, already-approved model. The unsolved part is producing the documents out of the same run that designed the part and proved it, so that the pack is internally consistent by construction rather than by somebody checking.

That is the layer LMCAD is building. It is in beta and goes on sale on 1 January 2027. The other two halves are covered separately: what agentic CAD means as a category, and what automated engineering analysis actually requires.