What jobs does computer vision do in a business?
Four families cover most of it. Inspection: detecting defects, wear or missing parts on a line, often before a person would. Reading: extracting text and structure from documents, where it overlaps optical character recognition. Monitoring: counting people or stock, verifying safety equipment, flagging events on camera feeds. And measurement: dimensions, planogram compliance, land or crop condition from aerial imagery. Each one replaces a visual check someone was doing by eye, or not doing at all, which is where the business case usually sits.
How did vision-language models change the field?
They collapsed many single-purpose models into one general one. Classic computer vision trained a separate model per task on thousands of labeled images; a modern vision-language model answers questions about an image directly, so document intake, inspection notes and photo triage can run on prompted general models, with trained classification specialists kept for high-volume, high-precision work. The same capability put vision inside agents: a vision agent uses what it sees to decide what to do next, from operating software by screenshot to checking imagery for brand fit. General models made the pilot cheap; the specialist model is now an optimization, not the entry ticket.
What does a computer vision deployment actually require?
Mostly things that are not models. The camera positions, lighting and image quality decide the ceiling; a definition of the event or defect precise enough that two people would label it the same way; a plan for the edge cases the model flags with low confidence, which usually means a human review queue; and a feedback loop that turns corrected mistakes into training or prompt updates. Buy the model, but budget for the physical and definitional work around it, because that is where vision projects stall.