AI can do most of this extraction work, but messy handwriting will force a human back in.

The core fields are consistent week to week, but variable table layouts across three different factory partners and hand-written sections mean the agent must adapt its parsing logic per document. This is repeatable in intent but not in structure, which raises failure risk.

The output schema is well-defined (six specific fields, one CSV), which is favorable. However, 'severity flags' and defect categorization from handwritten notes introduce interpretation ambiguity that the agent cannot resolve without a defined rubric.

PDFs are available as inputs, and OCR + document-parsing tools (e.g., Azure Form Recognizer, AWS Textract, or GPT-4V) exist. However, handwritten content and non-standard table formats push these tools to their limits, and no ground-truth validation set is mentioned.

Incorrect defect counts or missed severity flags could propagate into quality dashboards, supplier scorecards, or compliance records — causing real downstream decisions to be made on bad data. Errors are reversible in principle but costly to catch after the fact across 2,600 records.

Most extraction is mechanical, but ambiguous handwriting, crossed-out values, and contextual severity assessments will require human judgment on a meaningful minority of records. A full no-human pipeline is risky; a human-in-the-loop review step is strongly advisable.