A convincing AI-generated CAD model can pass a product demo and still fail a regulated-customer audit. Approval depends on a traceable design history: who made each decision, what evidence supported it, what changed, and who accepted the risk.
Consider this invented composite. At 4:40 p.m. in Berlin, Kojo, a founder from Accra building inspection equipment, was sharing his screen with a prospective customer’s engineering and quality teams. His AI-assisted workflow had turned a written brief into a clean enclosure model, complete with mounting points and manufacturing drawings. The engineering lead rotated it on screen and said the geometry looked ready.
Then the quality manager asked Kojo to open the record behind one mounting point.
The question the model could not answer
Kojo had expected questions about dimensions, materials and tolerances. He had the finished files, the original prompt and several exported versions.
What he could not show was why the mounting point had moved between versions, which requirement caused the change, or who had checked that the new position would still tolerate vibration. The AI system had produced a plausible result. His team had reviewed the shape. Nobody had preserved the chain of decisions between the brief and the final file.
The quality manager tried another route. Which engineer had approved the material choice? What test evidence had they used? Had the model inherited an assumption from an earlier product?
Kojo searched through chat messages while six people waited. One engineer remembered discussing the material during a call. Another thought the vibration requirement had arrived in an email. The only signed record covered the final export, not the reasoning that created it.
The review ended without approval.
That mattered because the prospective customer had reserved a production window. If Kojo could not reconstruct the design history before its internal decision, the pilot could move to another supplier. The model on his screen still looked convincing. For the next several days, that barely mattered.
A polished output can hide an unfinished decision
AI makes the visible part of design move faster. A team can explore alternatives, generate geometry and prepare a strong demonstration before its old documentation habits have caught up.
That speed creates a dangerous mismatch. The output looks final, so people treat the underlying decisions as settled.
A regulated customer sees a different object. The CAD file is one result inside a controlled process. Reviewers need to follow a requirement into a design choice, see the evidence behind that choice, identify later changes and confirm that someone with the right responsibility approved them.
A prompt log rarely provides that history. It may show what someone asked the system to generate, but it does not necessarily explain why the team accepted one result and rejected another. Version names such as `final_3_revised` provide even less.
This is closely related to the failure in finished CAD that hides the wrong manufacturing assumption. A technically impressive artifact can carry an untested choice all the way into a customer review. AI shortens that journey.
Build the evidence while the design is moving
Kojo’s team first tried to reconstruct the missing history from memory. That failed quickly. The engineer who had changed the mounting point could explain the immediate reason, but could not identify the customer requirement that triggered it. A message thread contained part of the discussion. An old screenshot contained another part.
With the customer’s decision still open, they changed the task. They stopped trying to write a perfect retrospective and built a decision record around the areas the reviewer had challenged.
For every material change, they captured the requirement, the options considered, the chosen option, the evidence reviewed and the person accepting the decision. They linked those records to model versions. Where evidence did not exist, they marked the decision unresolved and ran the missing check instead of inventing certainty after the fact.
This was slower than preparing another polished demo. It was also the first work that directly addressed the reason approval had stopped.
The useful discipline starts before the next model is generated. Decide which changes require human approval. Preserve rejected options when the rejection affects safety, performance or manufacturing. Connect test results to the exact version they support. Record the boundary between what the AI proposed and what an accountable person accepted.
That boundary matters as AI becomes more capable inside CAD tools. Generating a custom feature from an instruction can save engineering time. The audit question remains: why did this feature belong in this design, under these constraints, and who verified the consequence?
Make the next review inspectable
Kojo returned to the review with a less theatrical presentation. He opened the requirement record first, followed it into the changed mounting point, showed the test evidence and named the engineer who approved the decision. Two assumptions remained unresolved, and he said so plainly.
The reviewers could now separate completed work from open risk. They did not have to infer confidence from the quality of a rendered model.
For a founder on limited runway, traceability can feel like paperwork that belongs after product-market fit. A regulated buyer will often treat it as part of the product. If the customer cannot inspect how a consequential design decision was made, the customer inherits uncertainty it cannot approve.
Before your next AI-assisted CAD demo, pick one feature with safety, compliance or manufacturing consequences. Ask someone outside the design session to trace it from requirement to evidence to approval. If they need the original engineer in the room to explain the chain, the design history is still living in someone’s head.
Kojo’s next screen share began with that chain already open.
Comments
No comments yet.