AI in companies looks like one workflow, not a transformation programme

The usual lists of AI-in-companies examples sort by category: marketing, customer service, production, finance. That doesn't help anyone make a decision, because each category lumps together a dozen unrelated implementations. The more useful question is what a single implementation actually looked like: what scope, what input, and what control point.

Case 1: document processing instead of data entry

A manufacturing company received purchase orders as PDFs, in different formats from different customers. The fix wasn't generic "AI for documents." It was an intelligent document processing pipeline with a narrowly defined job: read the PDF, extract line items, validate against SAP master data, create a checked SAP order. Anything that fails validation goes to a person, not into the system. The scope was narrow enough that the error rate was measurable from day one, and approval authority stayed exactly where it was before.

Case 2: computer vision on the production line

A second case had nothing to do with language. Industrial computer vision models watched production lines for anomalies that signalled a fault developing before it turned into downtime. The detail that made this different from a generic pilot: the model was trained and validated against a specific, known fault class, not "anything that looks unusual." That makes the result checkable, because you know what the system is looking for and what it isn't.

Case 3: enterprise search with a permissions layer

A media company with over 500 users had the opposite problem: too much content, spread across too many systems, for a conventional search to find. The AI-powered search platform that resulted was technically demanding but simple to describe: a search that understands what's being asked and only returns content the person asking actually has access to. The permissions logic mattered as much as the language model.

What these three cases have in common

None of them was an "AI transformation" programme. Each had a single workflow as its boundary, a clearly defined input, and a point where a person kept the final decision or the system could be checked against a known reference. That constraint is the difference between a use case you can actually evaluate and an idea that looks good on a slide. An AI Discovery Sprint scores which of your own ideas already has that shape, and which one needs it before it earns a budget. For an overview of the capabilities these cases come from, see our AI page.