A finished CAD file can still fail at supplier handoff when the AI completes the geometry using an unstated manufacturing assumption. Before approving the file, the founder must verify that the proposed process, tolerances and machine constraints match what the supplier can actually produce.
In September 1999, NASA’s Mars Climate Orbiter approached Mars after a journey of more than nine months. The spacecraft was expected to enter orbit. Instead, communication stopped.
The geometry of the flight had been calculated. The commands had been sent. Yet two teams had worked with different assumptions: Lockheed Martin supplied thruster data in pound-force seconds, while NASA’s Jet Propulsion Laboratory in Pasadena expected newton seconds. The mismatch sent the spacecraft much closer to Mars than planned.
NASA lost the orbiter.
Arthur Stephenson chaired the mishap investigation board. Its report traced the failure to the units mismatch, then went further. The problem survived because the project’s checks, communication and operating practices did not expose it before the spacecraft reached Mars.
That is the useful parallel for an AI-generated CAD file. Correct geometry inside one system can still encode the wrong reality outside it.
The assumption hidden inside the finished model
The founder’s Friday problem begins with a file that looks complete.
The holes align. The enclosure closes. The dimensions appear coherent. The AI has resolved the visible geometry and removed the unfinished edges that made the earlier version look uncertain.
The supplier then reviews it and asks which manufacturing process the part was designed for.
That question changes the handoff. The AI may have produced features suited to CNC machining when the supplier intends to use sheet-metal fabrication. It may have treated a wall, corner or internal cavity as geometrically possible without accounting for the supplier’s tooling. The file can pass a visual inspection while asking the factory to make a part using a process it does not offer.
The founder now faces an awkward choice. Send the file because the schedule says Friday, or reopen a model that already looks done.
Limited runway makes the first option tempting. Supplier review feels like a late-stage formality, particularly after an AI tool has completed work that previously required several rounds with an engineer. But a confident render carries no evidence that the manufacturing assumptions were checked.
The Mars Climate Orbiter failed through a different technical mechanism and at a vastly different scale. The shared pattern is narrower: one side produced valid output under one convention, while the receiving side acted under another.
A supplier handoff needs an interface contract
A CAD file describes shape. A supplier handoff also needs to describe how that shape is expected to become a physical object.
Before release, I would put the assumptions beside the model in plain language:
- Name the intended manufacturing process.
- Identify the material and the few dimensions that affect fit or function.
- Mark which tolerances are essential and which can move.
- Ask the supplier to flag any feature that requires different tooling, setup or construction.
- Request confirmation before treating the file as approved for production.
This short note does something the polished model cannot. It gives the supplier a chance to challenge the reasoning behind the geometry.
The same principle applies when AI generates code, contracts or automation logic. The output often appears finished because the missing assumption is invisible on the screen. The receiving system exposes it later, usually after money or time has been committed.
That is why “the file opens” is a weak acceptance test. “The supplier can produce this using the agreed process” is stronger.
A related failure appears when geometry is mathematically valid but physically difficult to build. I explored that boundary in What Happens When Valid CAD Geometry Cannot Be Manufactured?. The recurring issue is authority: who can declare the work complete, and what evidence must they see first?
Friday is the moment to slow the commitment
The practical move is small. Do not ask the AI whether the CAD is finished. Ask it to list the manufacturing assumptions it used, then treat that list as unverified until the supplier responds.
An AI system can help identify likely concerns, prepare drawings and revise geometry. It cannot confirm a supplier’s actual process unless that information has been supplied and checked. When the model lacks that context, fluent completion can hide uncertainty instead of removing it.
This distinction matters most for a founder managing a short runway. A delayed handoff is visible immediately. A wrong handoff stays quiet until a quote changes, a sample fails or a production slot disappears.
NASA’s investigation did not treat the 1999 units mismatch as an isolated typo. The report examined why the project failed to catch incompatible conventions before the final operation. A founder should use the same frame at a smaller scale: the defect lies partly in the geometry, but also in the handoff that allowed an assumption to travel without a challenge.
On Friday afternoon, add one final approval field beside the file: “Supplier confirmed process and constraints.” Until someone on the manufacturing side can fill it in, the CAD is ready for review, not ready for production.
Comments
No comments yet.