AI BOM Scrubbing for Electronics Manufacturers

What Is Wrong With Customer BOMs

Customer BOMs are exported from CAD tools, edited by hand in Excel, and sent without a standard format. Typical defects include description-only lines with no manufacturer part number, manufacturer names spelled five different ways, quantities that disagree with the reference designator list, merged cells, multiple parts crammed into one line, and alternates mentioned in a notes column. Another common problem is stale revisions, where the BOM and the assembly drawing do not match.

Each defect is cheap to find individually and expensive when found late, because a wrong quantity discovered at kitting becomes a short kit, and a wrong part discovered at placement becomes rework.

What AI Is Good At Here

Language models handle the fuzzy work that rule-based importers fail on: reading a column layout it has never seen, recognizing that a manufacturer name is a variant of one in your database, interpreting a description such as a 10k 1 percent 0402 resistor and suggesting matching catalog parts, and parsing notes that say use equivalent or no substitutes. Deterministic code should do the checking: counting reference designators, comparing quantities, and validating that a proposed part's package matches the footprint. The pattern that works is AI proposes, rules verify, a human approves.

Models left to decide alone will occasionally invent a plausible-looking part number, so every proposed match needs a source link.

• Normalize columns and manufacturer names across inconsistent files
• Propose manufacturer part numbers for description-only lines
• Cross-check reference designator counts against quantities
• Flag lines where package, value, or tolerance disagree with the description

Lifecycle, Availability and Alternates

Once lines are matched, the next scrub is risk. Check lifecycle status (active, not recommended for new designs, last time buy, obsolete), stock across your approved sources, and lead times. For risky lines, AI can draft alternate candidates by comparing parametric data such as capacitance, voltage rating, dielectric, tolerance, and package, then list the differences an engineer should evaluate.

Alternates still require customer approval and often engineering review, especially for crystals, power parts, and anything with tight timing or thermal requirements. The value is in producing a shortlist with evidence, not in autonomously substituting parts, which would violate configuration control on most regulated builds.

Keeping Customer Data Private

BOMs and drawings are customer intellectual property, and many are subject to NDAs, ITAR, or CUI marking. Sending them to a public AI API can breach contract terms. A private deployment runs open-weight models on your own hardware, retrieves from your approved manufacturer list and past jobs, and logs what was processed.

Netray builds this as a read-only layer over your ERP and document stores, with the estimator or engineer approving any change that writes back. For shops handling only non-sensitive consumer work, the privacy case is weaker, and a cloud tool may be adequate if your customer agreements allow it.

When Scrubbing Is Not Worth Automating

If you build a small number of repeat products from your own controlled BOMs, a scrubbing tool adds little, because your BOMs are already clean. The payoff appears when you quote many new BOMs from many customers, where each file is a fresh mess. Measure first: sample twenty recent RFQs and record how long scrub and part matching took, how many BOMs contained errors discovered later, and how often those errors cost money.

If scrub time is under thirty minutes per quote and errors are rare, spend the effort on labor models and feedback from actuals instead.