Residents submit pothole photos that must become structured work orders. How should the city integrate generative and multimodal models in the agent workflow?

Pothole photos from residents become useful municipal work orders when vision understanding extracts defect cues and text generation drafts structured fields (location hints, severity language, category codes) for crew systems. Integrating generative and multimodal models in that workflow means the image is inspected and the work-order draft is produced under agent tool policies—not that photos are ignored, posted only to social media, or replaced by random IDs. Multimodal drafting still requires human or system validation before a work order is authoritative, but the development pattern correctly combines vision with structured text generation.

Ignoring images and requiring free-text-only defect descriptions forever fails by concept: it discards evidence residents already provided and reduces accessibility for people who struggle to describe damage. Using vision alone to post to social media without creating a work order misses the operational goal of entering public-works queues. Generating work-order IDs randomly without inspecting the uploaded image fabricates process artifacts and ignores the multimodal input that should ground the ticket.

Exam caveat: vision models can misread lighting, weather, or non-pothole defects; confidence thresholds and staff review remain important for safety-critical street conditions. PII in images (faces, license plates) may need redaction policies before storage.

Operational check: submit known pothole and non-pothole images in staging, verify drafted work-order fields reflect visible evidence, and confirm no work-order ID is finalized without the governed create-ticket tool succeeding.