Buying computer vision for construction sites: a procurement field note

Computer-vision pilots on construction sites are now cheap to start and hard to scale. Operators we interviewed describe the same pattern: a successful 90-day pilot, then a stall when the tool has to integrate with the site's actual safety, scheduling and incident workflows.

The failure is rarely the model. It is the operating model.

First: what does the vendor's incident workflow look like end-to-end, from camera event to closed-out corrective action? Second: who owns the data, including derived event metadata, after contract end? Third: what is the false-positive rate at the safety officer's desk after one month of tuning, not in the demo? Fourth: how does the system behave when the site network drops? Fifth: what evidence does the vendor have from sites resembling yours, climate, trade mix, sub-contractor structure?

These five questions cut the candidate field for most operators by half within a single round of evaluation.

Built-environment AI procurement is a craft. The biggest gains come from clearer questions, not better tools. We are publishing a longer field guide for procurement leaders later this year.