A 640-Line BOM Scrub Ties Up Your Buyer for Three Days. Agent Teams Split the Same RFQ and Return It Before Lunch.

Twenty-six RFQs came in. You quoted fourteen.

Stare at that number. In a 60-person contract assembly shop running three SMT lines, the quoting desk is one materials buyer and one estimator, often the same person before 7 a.m. RFQs arrive as a zipped folder: Gerbers or ODB++, assembly drawing, fab drawing, centroid file, test spec, and an Excel bill of materials with 640 lines on it. The customer wants pricing at 25, 100 and 500 pieces with a five-day prototype turn, by Friday.

Twelve of the twenty-six never got answered. Some aged out — the OEM's sourcing engineer placed the order with whoever came back first, which in this business is often whoever came back at all. Each one is not a lost quote, it is a lost program: a board built somewhere else for four years, plus spares, plus the revision B eighteen months later.

A BOM scrub is the line-by-line job of turning a customer's parts list into a price you can stand behind: every part number confirmed against their approved manufacturer list, priced at the quantity break you would actually buy at, checked for lifecycle and lead time, and flagged where nobody has stock. None of it is clever. All of it is slow.

Where the three days actually go

Take the 640-line BOM. Maybe 380 lines are unique part numbers once you collapse duplicates. Your buyer opens Digi-Key, Mouser, Arrow, Avnet and TTI and prices three quantity breaks each. She checks minimum order and standard pack quantity, because a line priced at cut tape and bought as a 3,000-piece reel is a $340 surprise on a 25-piece build. She checks whether the part is non-cancellable and non-returnable, because on a prototype that terms line is the whole risk. She checks lifecycle status in SiliconExpert or Z2Data and finds four parts end-of-life with a change notice from October, and one connector whose only stock sits at an independent distributor at nine times book price.

Then she reconciles every substitute against the customer's AML, because on a Class 3 build you cannot swap an X7R capacitor for an X5R just because it is in stock. Meanwhile the estimator opens the centroid file, counts placements — 1,842 SMT, 61 through-hole, 14 hand-load parts — calculates line time at your actual placement rate, and adds stencil, first article, program creation, wash, conformal coat if the drawing calls for it, and flying-probe test because the volume does not justify an in-circuit fixture yet.

The workaround everyone pretends is fine: price the top 80% of material spend properly, carry the rest at last quarter's numbers plus a contingency. Every shop does it. It is also why programs quietly run at four points of margin instead of twelve, and you find out on the third build.

It is not slow because your buyer is slow. It is slow because it is single file.

The scrub is serial by construction: one person, one browser, one distributor site at a time, from row 1 to row 640. Hiring does not fix it — put three buyers on the same BOM and you get three formats and a reconciliation meeting. So it stays one person and the calendar does the rest: the BOM eats Monday and Tuesday, line-time math eats Wednesday morning, the program manager builds the quote in Epicor Kinetic that afternoon, you review margin Thursday, it goes Friday.

Then December arrives: OEMs lock next-year pricing while your bare-board fabricators warn about Lunar New Year shutdowns, every pre-buy gets quoted at once, and the desk that handles fourteen a month is looking at thirty-one.

flowchart TD
  A["RFQ lands: Gerbers, 640-line BOM, centroid file"] --> B["Job split four ways"]
  B --> C["Price lines against Digi-Key, Mouser, Arrow, Avnet, TTI"]
  B --> D["Check lifecycle, PCN and EOL against the customer AML"]
  B --> E["Count placements and hand-load parts from the centroid file"]
  B --> F["Read the fab drawing for panel size, finish and DFM flags"]
  C --> G["Merged cost sheet with one exceptions list"]
  D --> G
  E --> G
  F --> G
  G --> H["Estimator works the exceptions and sets margin"]

What Agent Teams changed about that single file

Agent Teams arrived as a research preview alongside Claude Opus 4.6, and the plain description is the useful one: instead of one assistant working through your job in order, several work on different parts at the same time and merge what they found. You can also hand over the whole packet at once — full BOM, fab drawing, assembly notes, the AML — instead of feeding it in chunks and hoping it remembers row 40 by row 600.

The difference from what you tried in 2024 is not that the answers are smarter. It is that the work stopped being a queue. One part of the job prices lines 1 through 160 while another prices 161 through 320, a third runs lifecycle and change-notice checks, and a fourth reads the drawing and centroid file for placement counts and process flags. They finish at roughly the same time, and the merge is the piece that earns its keep: one cost sheet and one exceptions list instead of 640 lines nobody has time to read.

What arrives on your estimator's screen is not a quote. It is the boring 90% done with the interesting 10% circled: four EOL parts, one connector only available through a broker, nine lines where the AML alternate beats the specified manufacturer, and a note that the 0.4 mm pitch BGA means an X-ray step the drawing does not mention.

Tuesday, 8:40 a.m., on the quoting desk

The packet came in Monday at 4 p.m. from a medical device OEM — Class 3 workmanship, ISO 13485 flow-down, 500-piece annual usage. By 8:40 Tuesday the merged cost sheet is sitting there. Your buyer does not price 380 lines; she works 31 exceptions. She calls TTI about the connector, gets a 22-week lead time, and pulls the AML alternate instead. She flags the two NCNR lines so the program manager puts them in the terms section rather than eating them.

Your estimator spot-checks the placement count against the centroid file himself — eight minutes on a sample of reference designators — then applies your real line rate, not a nameplate number. The NPI engineer agrees the 0.4 mm BGA needs X-ray and that the panel as drawn will not run on your rail width, and writes two lines of customer feedback that is frankly the thing that wins the job. The quote goes out at 2 p.m., 22 hours after the RFQ landed, with a DFM note the other three bidders did not send.

The arithmetic: what the extra ten quotes are worth

Assumptions, all illustrative — put your own numbers in the boxes. A 60-person shop, three SMT lines, 26 RFQs a month, quoting 14. Loaded cost of the desk is $86 an hour. Average won program is $71,000 in first-year revenue at 12% gross margin. Win rate on quotes you chase is 24%; assume the marginal quotes you used to decline convert at a worse 15%, because they are the ones you know least.

Two honest caveats. This only works if you have line time to build what you win; quoting faster into a full plant creates late deliveries. And the desk-hours saving is not a headcount cut — it is a buyer with time to negotiate your top twenty spend lines, worth more in most shops than the throughput.

The four things that stay on a person's desk

Do not let the merged cost sheet make commitments. NCNR terms, broker buys on allocated parts, and any purchase above your controller's threshold need a human signature, every time. Second, DFM judgement stays with your process engineer — a machine can tell you the drawing calls for a 0.4 mm pitch device; it cannot tell you this customer's paste release will fight your stencil shop.

Third, margin is yours. The number that wins a Class 3 medical program comes from what you know about that customer's history of engineering changes, not a spreadsheet. Fourth, and this bites shops that skip it: if the packet holds ITAR-controlled technical data, it does not go into a general-purpose tool without export-control review first. Gerbers for a defense program are controlled technical data.

Start Monday with one thing: take the last three RFQs you declined, re-run them as a split job, and time it. If the exceptions list is short and correct, you have your answer. If it is long and wrong, you have learned what your AML data quality really looks like.

Frequently asked questions

Will it price parts correctly, or is it just guessing?

It reads distributor pricing the way your buyer does and reports what it found with the source next to the line. The failure mode is not invented prices — it is stale ones, and quantity-break mistakes where standard pack quantity is not honoured. Insist every priced line carries the distributor, quantity break, pack quantity and the date pulled, then spot-check ten lines per quote. Four minutes.

What about customer BOMs that are a mess?

Most are. Merged cells, four ways of writing the same manufacturer, reference designators crammed in one cell, and a "notes" column holding the only mention of a required deviation. Handing over the whole file at once helps here most, because the note in row 612 gets matched to the part in row 40. A BOM with no manufacturer part numbers is still a phone call to the customer's engineer.

Do I need to change ERP systems for this?

No. The output is a cost sheet you paste into Epicor Kinetic, Cetec, Global Shop or wherever your quote template lives. Rewiring your ERP is a six-month project; changing how the scrub gets done is a two-week habit change on one desk.

A note on the phone while all this is happening

The other half of a quoting week is the phone: the sourcing engineer chasing Tuesday's RFQ, the customer asking about a rush build, the distributor rep returning a lead-time call. CallSphere builds AI voice and chat agents that answer the line and website chat around the clock, capture who called and which program they are asking about, and book the callback. It does not scrub your BOM. It does mean the buyer who calls at 6:15 p.m. gets an answer instead of voicemail.