AI for manufacturing, starting with the data the machines already produce

Usually. Older equipment often exposes serial output, a file drop or a vendor interface even without a modern API. The survey stage establishes this before anything is promised, because the answer genuinely varies by asset.

Not on day one, and anyone who says otherwise is selling. Prediction needs months of machine history including actual failures. If you are not instrumented yet, the honest first project is capturing that data — which delivers visibility value immediately and makes prediction possible later.

Read-only integration does not affect operation and can normally be introduced without downtime. Anything that writes to or controls equipment is planned with your engineering and safety people, not around them.

Assume it will drop, because it does. Data is buffered locally and forwarded when the connection returns. A pipeline that assumes a perfect network loses data quietly, which is worse than not having it.

For consistent, visible defects it can be very reliable and tireless. We would still route borderline cases to a person — the failure mode of an over-trusted inspection system is that it passes a defect confidently.

It changes your security exposure, so segmentation, authentication and least-privilege access are part of the design rather than an afterthought. This is a genuine risk and worth treating as one.