AI designed a multi-part mechanical assembly, unsupervised, in under an hour
The Jarvis Onshape MCP demo showed Claude building a 4-part monitor arm — joints, clearances, mates — starting only from a rough sketch and a description. It checked its own work using the tools available to it and iterated. This is CAD, not text generation. The geometry is real, exportable, and manufacturable.
Schematic capture and simple sheet-metal parts have been automatable for years. Complex multi-body parametric assemblies just crossed the line. The question isn't whether this affects the mechanical engineer role — it's how fast the function adapts.
Source: @Reshef_ on X, April 20 2026 — "Claude can actually do CAD now in @Onshape … Here it worked for an hour and built a 4-part monitor arm, starting only from a sketch and description."
01Where your week actually goes (pre-augmentation)
Typical distribution for a mid-level mechanical design engineer across product, tooling, or capital-equipment contexts. Actual numbers vary by company and project phase.
The largest block — CAD modeling and design iteration — is exactly what the Onshape demo automated. The second-largest — FEA and simulation review — is partially there. The pattern: anything that produces geometry or structured documentation is under active AI attack right now.
02Old role vs augmented role
- Spends 16+ hours per week drawing geometry that requirements already define
- Runs simulation, waits, adjusts, re-runs — iterating manually through design space
- Writes ECO descriptions from memory and tribal knowledge
- Assembles BOM line by line; chases suppliers for lead times manually
- Produces DFM notes that manufacturing ignores until tooling is cut
- Reviews tolerances against mental model of process capability
- Drafts FAI/PPAP packets by pulling data from six different systems
- Specifies intent: loads, interfaces, constraints, failure modes — AI generates geometry
- Reviews and adjudicates AI-proposed design variants; owns the selection rationale
- Validates simulation results, escalates edge cases, signs FEA reports
- Approves ECOs drafted by AI from change-request context; adds engineering judgment
- Reviews AI-assembled BOM for obsolescence, substitution risk, lead-time flags
- Owns DFM sign-off; AI flags manufacturability issues before the conversation happens
- Signs FAI/PPAP compiled automatically from production and dimensional data
03Day in the life — augmented mechanical engineer
04New job description
Core accountabilities
- Own engineering intent: translate requirements into precise load cases, interface constraints, tolerance priorities, and failure-mode consequence that AI can act on
- Review, select from, and adjudicate AI-generated design variants — document selection rationale with engineering basis
- Validate simulation outputs; own failure mode identification and escalation
- Approve ECOs, FAIs, and PPAPs; signature represents engineering judgment, not data assembly
- Own DFM sign-off; partner with manufacturing before tooling commitment
- Review AI-assembled BOMs for obsolescence, substitution risk, and lead-time exposure
- Develop the engineering intent specification capability on the team
What no longer defines the role
- Originating geometry for defined requirements
- Running parametric sweeps or basic FEA iterations manually
- Assembling BOMs from scratch or drafting ECO descriptions from memory
- Compiling FAI/PPAP packets from multiple source systems
- Drafting DFM notes that manufacturing hasn't seen yet
05KPIs that move
| Metric | Baseline | Augmented | Driver |
|---|---|---|---|
| Design variants evaluated per sprint | 2–4 | 15–40 | AI geometry generation; human selection |
| Time from brief to first manufacturable geometry | 3–10 days | 2–8 hours | Parametric AI CAD from intent brief |
| ECO cycle time | 5–14 days | 1–3 days | AI drafts from change context; engineer reviews |
| FAI / PPAP compilation time | 5–14 days | 1–2 days | Automated data assembly; engineer signs |
| DFM issues caught before tooling | ~40% of issues | ~85% of issues | AI manufacturability check in design loop |
| BOM obsolescence exposure at release | ~15% of line items | <3% of line items | Real-time component lifecycle monitoring |
| Design re-work after tooling commit | 25–40% of projects | 8–15% of projects | More variants evaluated; better DFM review |
06Skills to develop
Engineering intent specification
Writing precise, AI-actionable briefs: load cases, interface constraints, tolerance priority rationale, failure consequence. The new core competency.
AI design review
Critical evaluation of AI-generated geometry: reading stress contours, identifying failure modes AI missed, understanding what the AI optimised for and why that might be wrong.
Design-space thinking
When you can evaluate 30 variants instead of 3, the engineering value shifts to knowing which axes of variation matter and why. Statistical thinking about design trade-offs.
Failure mode ownership
FMEA, fault trees, failure consequence analysis — the parts of engineering that require human accountability. These expand when the geometry work contracts.
Manufacturing partnership
Tighter DFM loops. Understanding process capability at the level that lets you write tolerances AI won't overspecify. This requires real manufacturing relationship, not just tooling handoff.
Supply chain fluency
Component obsolescence, lead-time exposure, qualified substitution logic. When AI flags these in the BOM, the engineer needs to evaluate them with real procurement context.
07Junior and senior reshape
- Traditional entry ramp (learn CAD by doing CAD) is largely gone
- New entry ramp: learn to write engineering intent briefs, validate AI geometry against requirements, own the test plan for a sub-assembly
- Failure mode identification becomes a primary competency from year one
- BOM review and component qualification as a real ownership area, not a task
- Faster path to meaningful work — reviewing 30 AI variants beats months of redlines on one
- Risk: engineers who arrive expecting to do CAD will be disappointed and must adapt quickly
- Domain knowledge for writing engineering intent is now the scarce resource
- Cover more product families, more design programs simultaneously
- Own the AI review framework for the team — define what "good enough" looks like in generated geometry
- Principal design authority on failure mode and consequence — this is not delegatable
- Manufacturing relationship owner: process capability knowledge AI cannot infer from data alone
- Build the intent-specification library the team reuses across programs
08What percentage of your week could be augmented?
Adjust the sliders to match your actual weekly hours. The estimate reflects current AI capability — not theoretical future state.
of your week could move to autopilot or augmented review
Get the full Mechanical Engineer transition playbook — new JD template, intent-brief framework, AI review checklist, and tool shortlist — when we publish it.
09Frequently asked questions
Is the Mechanical Engineer role going away?
No. Engineering judgment, failure mode ownership, tolerance sign-off, and the engineering change authority stay human. What moves to autopilot is the geometry generation, iteration, and documentation — the work that used to consume 60% of a design engineer's week. The role gets more senior, not smaller.
Do I need to learn AI or parametric prompting to keep this job?
You need to learn to specify design intent precisely — material constraints, load cases, interfaces, tolerances — in a form AI can act on, then critically review its output. The new skill is engineering judgment at speed, not model training.
How does this work with GD&T, PPAP, or AS9100 requirements?
AI-generated geometry can be output to any CAD format and annotated with full GD&T. The engineer reviews and stamps every tolerance and every PPAP submission. The AI does not sign — it drafts.
Will engineering headcount drop?
Individual engineers cover 3–5× more design variants per sprint. In most organisations that means fewer design iterations stall in queue — headcount holds but each person's output surface area grows. Companies that cut headcount fast typically lose failure-mode expertise they can't recover.
What CAD platforms does this work with?
Onshape, Fusion 360, SolidWorks, PTC Creo, Siemens NX, and CATIA all have MCP or API surface that agentic tooling can address. The Onshape MCP demo in April 2026 is the clearest public proof point — a 4-part monitor arm assembly built from a sketch and description in under an hour.
What about IP and proprietary design data?
All generation runs inside your VPC or on-prem. Design data never touches model training pipelines unless you explicitly opt in. IP stays yours.
What happens to junior mechanical engineers?
The traditional CAD apprenticeship path changes fundamentally. Juniors who adapt become design-review and simulation specialists faster. The new junior role is: specify intent, validate AI output, own the test plan.
What's the fastest way to start?
Pick one low-stakes design variant task — a bracket, a housing, a fixture — and run it through an agentic CAD tool with an experienced engineer reviewing output. Measure review time vs. original design time. That number tells you the business case better than any consultant slide.