Using AI to Review Device History Records Before Lot Release

What the AI Can Check Reliably

Start with checks that are rule-based and verifiable against source data, because those are easy to validate and easy to explain to an auditor.

• Completeness: required fields, initials, dates and signatures present for every step
• Limits: recorded measurements against drawing tolerances and control plan
• Cert matching: material lot, heat number, grade and spec against the received cert
• Calibration: gauge and equipment IDs against current calibration records
• Cross-record consistency: quantities started, scrapped and shipped balance
• Open items: any NCR, deviation or hold tied to the lot or its materials

Reading Scanned Records and Certs

Many shops still have paper travelers scanned into PDFs and mill certs arriving as images. Modern document models can extract table values, handwriting and stamps, but extraction error rates are not zero, especially with poor scans or unusual handwriting. Design the workflow so every extracted value links back to the source image region, flag low-confidence extractions for human confirmation, and sample-audit the results monthly.

For an electronic DHR where data is already structured, the AI job is simpler and more reliable, because it works from the system records instead of reading images.

Validation, Part 11 and Where the Human Stays

If AI output influences a release decision, treat it as software used in the quality system and validate it for intended use with a risk-based approach consistent with FDA's computer software assurance thinking. Define the intended use as pre-screening and reviewer assistance, not automated release, and keep a human signature on every release. Record the model version, prompts or rules, inputs and outputs for each review so you can reconstruct what the tool said.

Test with historical lots, including ones with known errors, and track false negatives, since a missed error is the metric that matters most.

Data Privacy and Customer Confidentiality

Device records and drawings are often confidential OEM intellectual property, and some include controlled information. Sending them to a public AI API may breach your confidentiality terms or quality agreement. Run open-weight models on your own servers or in a private cloud, restrict access by role as your ERP does, and keep retrieval limited to the records the user is authorized to see.

Log queries for audit. Many OEM supplier audits now ask directly how you use AI on their data, so have a documented policy and an answer ready.

When It Is Not Worth It

If you release a few lots a week and your reviewer already catches issues the same day, AI review is a distraction. It pays off where release is a bottleneck, record volume is high, review is done by one or two people, or recurring findings show completeness and transcription errors. Quantify first: count how many lots are rejected back for correction at review and how many hours of reviewer time per week go to checking basic completeness.

If the number is small, fix training and traveler design instead.