AI is changing the way software is built, including embedded software. Denying that would not be credible, so we do not. We use AI agents for code, unit tests, end-to-end tests and documentation.

But in embedded systems, errors have different consequences than they do in web applications. When a model misinterprets a requirement or lets objectives drift, the result is not just a “bug”; it is a product that may behave unreliably in the field.

That is why we treat AI like any powerful but fallible component in an embedded system: it needs to operate within a control loop. We understand its failure modes, from hallucinated objectives and misread requirements to drift during iteration, and we design our process accordingly.