The incoming inspection process at a mid-size structural fabricator looks like this: a truck arrives, the driver drops off a package that includes physical MTCs or a stack of printed PDFs. The receiving clerk opens each cert, finds the heat number, types it into a cell in a shared spreadsheet, notes the PO number, and moves to the next one. On a busy receiving day, that's 40–60 heat number entries. The process takes about 90 minutes.
That 90 minutes produces a spreadsheet with heat numbers that may or may not be right. Transposition errors on alphanumeric heat codes (e.g., typing "A2B347" as "AB2347") are common and often go undetected until a heat traceability query fails months later. Some certs are photocopies of photocopies with contrast problems. A few come rotated 90 degrees. Some use "Melt No." where others use "Heat No." or "Cast No." — same data, different label.
The spreadsheet then gets manually keyed into the ERP system by someone else, introducing a second opportunity for error. The original PDFs get filed in a folder by date. If anyone needs to find a specific heat number later, they search the spreadsheet first and then dig through the folder if the spreadsheet entry is wrong.
What Makes Heat Number Extraction Hard (and What Doesn't)
The technical challenges in automated heat number extraction are well-understood:
Field label variation. Different mills use different labels for the same field. "Heat No.," "Heat Number," "Melt No.," "Cast No.," "Charge No.," and "HT#" all refer to the same thing. A simple OCR-plus-keyword approach fails on the variants it hasn't seen. AI-based extraction learns that these labels are semantically equivalent and extracts the associated value regardless of which label appears.
Document layout variation. Mill cert formats are not standardized. Some mills use tabular layouts with labeled cells. Others use free-text paragraphs ("Material from Heat 8A3291 was tested..."). Some organize by test type (chemistry section, mechanical section). An extraction model trained on one mill's format may fail entirely on another mill's format if it's relying on positional rules rather than semantic understanding.
Scan quality problems. Rotated documents, low-contrast photocopies, and handwritten annotations over printed text create OCR challenges. Modern document AI handles rotation automatically and applies image preprocessing to improve contrast before extraction. The accuracy gap between a clean digital PDF and a third-generation photocopy scan is real but manageable — typically 95–97% extraction accuracy on clean documents vs. 85–90% on degraded scans.
Multi-heat certs. Some certs cover multiple heat numbers — a coil-to-plate conversion where the cert references both the original coil heat and the plate production heat, or a combined cert covering multiple PO line items. Extraction needs to identify which heat number corresponds to which line item or product, not just extract a list of numbers from the document.
None of these are unsolved problems. The extraction models exist. The OCR engines handle scan quality. The question is whether the implementation is accurate enough for production use.
What Accuracy Rates Look Like in Practice
For high-quality digital PDFs from major mills, AI-based heat number extraction achieves 97–99% accuracy on the heat number field specifically. This is better than manual keying, which has a documented error rate of 2–5% on alphanumeric codes entered under time pressure.
For lower-quality scans (photocopied fax transmissions, third-generation copies), accuracy drops to 88–93%. At this level, a human review step for flagged low-confidence extractions is appropriate. The system extracts what it can confidently, flags what it cannot, and queues the flagged documents for manual review — which is a much smaller set than the full incoming volume.
The combined human-plus-AI workflow achieves better accuracy than all-manual at higher throughput: the AI handles 90–95% of documents without human intervention, and human review is concentrated on the 5–10% where the AI is uncertain.
Downstream Impact on Traceability and ERP Linkage
Heat number accuracy is not just a data quality issue. It is the foundation of material traceability in fabricated metal products.
When a quality event occurs — a field failure, a customer complaint, a recall — the first question is "what heat was this material from?" If the heat number in the ERP record is wrong, the traceability query fails. You cannot identify what other parts were made from the same heat. You cannot pull the original cert to verify the material properties. You cannot trace back to the supplier or mill for corrective action.
In pressure vessel, structural, and pipeline fabrication, heat traceability is not optional. ASME Section VIII, AWS D1.1, and many customer quality plans require that heat numbers be documented and traceable through the fabrication record to the finished product. An MTC filing system based on manual entry produces traceability records of variable accuracy. The errors are silent — they don't announce themselves until someone tries to use the record.
Automated extraction with validation (the extracted heat number is confirmed against the cert PDF after extraction) creates a record that is as accurate as the cert itself. The link between the ERP record and the original cert document is automatic rather than relying on someone to file the right PDF in the right folder.
The 90-minute daily data entry process also becomes a near-real-time intake: certs can be processed within minutes of receipt, heat numbers are in the ERP before the material reaches the shop floor, and the traceability record is complete before fabrication begins rather than assembled after the fact.