A customer request for quote (RFQ) arrives with a spec pack attached: forty pages of drawings, a parts list split across three sheets, and a revision table that changed two part numbers last month. Before anyone can price it, someone has to turn those tables into line items in your quoting template. Usually that is a person at a keyboard, reading one cell and typing it into another.
To extract a bill of materials (BOM) from a spec pack, you identify the kind of document, confirm where it may go, choose the method that suits it, and check every line against its source cell. That is four jobs, not one:
- Know the document: text, scan or drawing.
- Decide where the documents may go.
- Pick the method that suits them.
- Check every line against its source.
Under all four sits one question that matters more than any tool: when a line is wrong, how fast can you see it?
Key findings
- In a 2026 survey of 23 wire harness respondents, 70% said poor or incomplete customer information limits their quoting (Wire Harness Manufacturers Association (WHMA), 2026). Extraction can only move what the drawing holds.
- The same survey’s most common estimate of hands-on quoting work is three to five hours per assembly (WHMA, 2026). That covers more than entering BOM lines: retyping is one part of a multi-hour job, so count your own hours first.
- On a 2025 table benchmark, multimodal AI models were far better than an optical character recognition (OCR) pipeline on cell text, while the pipeline was slightly ahead on table structure. The tables came from scientific articles, not drawings (Association for Computational Linguistics (ACL) workshop, 2025). So test any method on drawings your compliance team has cleared before you trust it.
- Under the Defense Federal Acquisition Regulation Supplement (DFARS), a defense contractor that uses an outside cloud provider for covered defense information must make sure the provider meets requirements equivalent to the Federal Risk and Authorization Management Program (FedRAMP) Moderate baseline (DFARS 252.204-7012, 2024). The question is which provider, and on what terms.
What does extracting a BOM from a spec pack involve?
It means turning the parts tables in a customer’s documents into line items in your own quoting template: part number, description, quantity and maker. Two costs sit in that job. The hours go into the typing, cell by cell. The risk gets in earlier, at finding the right table and mapping its columns.
A spec pack is rarely one clean table. It holds title blocks, notes, revision histories and sometimes several BOMs. The parts list can run across sheets, with the header repeated on each one. Vendors label the same column differently: “Make” in one pack, “MFR” or “MFG” in the next. Before any method can help, someone has to decide which table is the BOM and which columns map to which fields in your template.
Get that decision wrong and every line typed after it is wrong too.
Is your bill of materials in text, a scan or a drawing?
Ask two quick questions: can you select the text in the parts table, and is the BOM the only table on the sheet?
If you can select the text, it is a text document. You can copy the table, but merged cells, wrapped descriptions and repeated headers come across broken, and you fix them by hand.
If you can’t, it has to be read as an image, whether it was scanned or exported with its text turned into outlines. When you pull a BOM from scanned drawings, small print, stamps, faint lines and handwritten changes are where reading goes wrong.
If the BOM is one table among several, treat it as an engineering drawing, whether its text can be selected or not. A revision table beside it can change the meaning of some lines, so the risk is reading the wrong table, or the right one at the wrong revision.
Many spec packs mix all three. Sort the pages before you choose a method.
Where are you allowed to send a customer’s spec pack?
Ask your compliance team before any customer document leaves your systems, and before anyone tries a tool on a real one, ours included.
As the key findings note, the DFARS clause does not ban the cloud for covered defense information (DFARS 252.204-7012, 2024). So for each customer, ask which providers and terms apply, rather than assuming everything must stay in-house.
In our view, the risk that is easy to miss is the quick test: one engineer pasting one controlled drawing into a public AI tool to see what happens. A 2020 law-firm summary of the International Traffic in Arms Regulations (ITAR) encryption rule says that releasing encrypted technical data in readable form “requires an authorization to the same extent as the export of the technical data unsecured by encryption” (Baker McKenzie, 2020). So giving someone the means to read encrypted controlled data needs the same authorization as exporting the data to them. The summary does not mention AI tools: whether a tool provider’s access counts is our inference, and your compliance team decides. This post is not legal advice.
Which methods work, and where does each break?
Each of these five methods works on some documents, and none works on all of them without a person checking. By governed extraction we mean a method where each extracted line stays tied to its source cell, and a person approves it before export.
| Method | Works when | Breaks when |
|---|---|---|
| Copy and retype by hand | The document is short and the text can be selected | The pack is long, scanned or holds several BOMs |
| Import from the design tool | The customer sends the native design file | You get only the drawing or a scan |
| OCR scripts | Scans are clean and the layout repeats | Layouts vary by vendor, or columns are unruled |
| General AI tools | You need a quick first read and the document may go to the tool | The document may not leave your systems, or the output does not show each line beside its source cell |
| Governed extraction with review | Volume is high and every line must be checked against its source | The drawing never held the data, or no reviewer has time to check each line |
Now apply the question from the top: when a line is wrong, how fast can you see it? With retyping, only if someone reads the line again. With a design-tool import, at the import, and only when a field fails to map. With OCR scripts and general AI tools, at whatever spot check you run afterwards. With governed extraction, beside its source cell, when the reviewer checks each line. Whatever the method, a wrong line nobody checks is found when the quote is already out.
Our explainer on retrieval-augmented generation covers how AI tools find and use information in documents.
How do you check an extracted BOM line by line?
Put every extracted line beside the cell it came from, and check the columns before the values. A swapped column ruins every line at once. A single wrong quantity ruins one.
A practical order:
- The right table. Is this the BOM, at the current revision, for the assembly being quoted?
- The right columns. Does each column map to the right field in your template? Watch for maker and part number trading places.
- The line count. Does the extracted list have the same number of lines as the source, across every sheet?
- Quantities and units. Each, meters, feet: a unit read wrong prices the line wrong.
- Maker and part number. Check each against the source cell, not against memory.
When the document leaves something out, mark it and ask the customer. Never fill a gap from what the part “usually” is. That is the kind of gap the WHMA respondents describe (WHMA, 2026). No extraction method fixes a part number that was never on the drawing.
Where AI Intime fits
AI Intime is the layer that separates AI volatility from the business: a controlled path from AI capability to measurable, owned business outcomes. The large language model (LLM) behind it is a replaceable component, never inside the business. On top of the layer sit hyper-contextualized knowledge engines, each scoped to a domain, an account or a lab. Our use cases are built this way: BOM Extraction is the one built for reading spec packs.
Three use cases run in production today: Analytical Lab Reports, Knowledge Twin and BOM Extraction. The rest sit on the same foundation: we build custom use cases on the AI Intime platform for your processes, and the AI ROI Map collects real asks that could become one. Your use cases set the agenda.
BOM Extraction reads scanned documents, spreadsheets, images and Word files. It finds the candidate tables in a spec pack, including several BOMs in one document, and handles vendor terms such as “Make”, “MFR” and “MFG”. The user picks the BOMs and reviews each extracted item beside its source cell. Human review catches problems such as swapped columns. The checked lines then export to your own quoting template.
Our published figures: 5 min per BOM, down from hours; 95%+ accuracy on hard scans; zero data leakage, air-gapped; any LLM of your choice, swappable. The AI ROI Map entry for BOM extraction sets out where the cost of this work sits, including keyboard hours and transcription errors.
Recommendations by role
- IT (information technology): start here. List where customer spec packs are allowed to go today, and who approved it, and count the tools people already use with customer documents. Any tool on that list without an approval is the first thing to settle, before the first tool trial.
- Estimators and quotations engineers: sort your next ten enquiries by the same two questions (can you select the text; is the BOM the only table) before you choose a tool. That mix tells you which method to try first.
- Applications engineers: own the line-by-line review. Write down the column checks and the vendor terms your customers use (“Make”, “MFR”, “MFG”), so every reviewer checks the same way, and keep a count of lines corrected per BOM. If that count does not fall as reviewers learn the checks, the method is the problem, not the reviewers.
- Sales operations: count the keyboard hours per enquiry, from the spec pack arriving to the BOM sitting in the template. That number tells you whether changing the method is worth it at all.
Bring one enquiry your compliance team has cleared for a demo, and watch the line items land in your template. Book a demo on your own spec pack