Field note
The production AI systems audit
The useful question is not whether a company should use AI. The useful question is which workflow can survive data quality, approval paths, security review, and measurement.
Start with the business bottleneck
Most AI projects fail before model selection. The workflow is too broad, the owner is unclear, and the success metric is a sentence instead of a number.
A good audit narrows the work to one bottleneck, one accountable owner, one risk register, and one payback target. That constraint makes the build sharper.
Map the system before choosing tools
Model choice matters, but not before the team understands inputs, permissions, handoffs, review points, and failure modes.
The architecture should describe where human approval stays in the loop, where logs are stored, what data is never sent to a model, and how the workflow degrades when the model is wrong.
Ship the smallest measurable version
The first build should not try to make a sweeping AI claim. It should prove that one process can get faster, safer, or cheaper without breaking trust.
If the first system measures well, expand. If it does not, the team still leaves with better workflow knowledge and less sunk cost.