Paid first intake for PDFs, scans, screenshots, forms, invoices, and image-heavy knowledge work Scope: up to 10 representative files or images, maximum 50 pages total, and one defined extraction workflow; extra pages, fields, unsupported layouts, or rework are quoted separately Sample-set review with Docling/OCR and qwen3-vl:8b visual reasoning smoke test Private-data handling, access, retention, and source-citation boundaries documented Field-level error table, unsupported-layout notes, and clear go/no-go: Team RAG pilot, Business Secure workflow, BYO server plan, or stop No page-volume SLA or recurring production promise before the sample set passes review One-time intake; recurring production management is quoted separately Use this when you are not ready for a monthly managed service yet but need a practical answer on whether your documents and screenshots can become a local AI workflow on RTX 4000 Ada class hardware. We benchmark first and only widen scope after the fit is measured. On the 20 GB GPU, the standard visual model is Qwen3-VL 8B (Ollama package about 6.1 GB); Qwen3-VL 30B and 32B are outside this 20 GB pilot because 30B leaves no reliable runtime/KV-cache headroom and 32B exceeds available VRAM.