45 Part Numbers, a Five-Day RFQ Window, One Estimator: Agent Teams Turns Print Review From Serial Into Parallel

6:50 Monday, one zip file, 45 part numbers, bids due Friday at five

The estimator gets in before anybody else because the phone does not ring yet. On the screen is a single email from a contract manufacturer's commodity buyer: annual re-quote, 45 part numbers, prints and STEP models attached, pricing due Friday 5:00 p.m. Eastern, quantity breaks at 25 / 100 / 500. Some are parts you have run for six years. Some are new. Two are assemblies with an outside anodize operation, and one print is a scan of a scan from 1998 with a handwritten revision in the margin.

By Wednesday afternoon the estimator has worked through 26 of them. On Thursday the foreman pulls him onto the floor for four hours because a fixture on the Doosan is not repeating. Friday at 4:40 the shop sends pricing on 26 line items and a polite note about the other 19. The buyer, who has three other shops doing the same thing, awards those 19 elsewhere. Not because your price was wrong — because you never gave one.

That is not a capacity problem or a skill problem. It is a sequencing problem. The re-quote is slow for exactly one reason: one person can only read one print at a time.

Why a 45-line re-quote has always been serial work

Walk through what happens on a single line item and it becomes obvious. Open the PDF. Read the title block for material, revision level and finish. Find the tightest tolerance and decide whether it drives an operation or just a gage. Look for the geometric callouts — true position with maximum material condition on a bolt pattern means a different fixture and a different inspection plan than a plus-or-minus hole. Check the notes for a first article requirement per AS9102, for a special process that has to go to a NADCAP-approved vendor, for a customer packaging standard. Then size the stock, check the bar diameter against what your Citizen can pass, guess the setups, estimate cycle time, add inspection minutes, and enter it in the estimating worksheet in JobBOSS² or Paperless Parts.

Twenty-five to forty-five minutes per line, honestly done. Nothing about that requires the same brain to do all of it, but until this year there was no way to split it. Junior help does not exist in the estimating room of a shop this size.

So shops triage. You bid the parts you already run, you bid the easy milled parts, and you quietly let the awkward ones go. Everybody in this trade does it and nobody puts it in the quality manual.

What Agent Teams changed about that Monday

Agent Teams arrived as a research preview alongside Claude Opus 4.6, and the plain description is this: instead of one assistant working through a long job one item at a time, several work on different pieces of the same job simultaneously and their results are merged into one deliverable. Opus 4.6 also reads far more material in one sitting than last year's version could — in shop terms, you can hand it the whole RFQ package, every print, the customer's supplier quality manual and last year's pricing file, all at once, without cutting it into pieces first.

For a job shop, parallel agents do not make the estimating better; they make the calendar different, because a five-day RFQ window stops being a limit on how many line items you can price.

The 2025 version of this idea was one assistant grinding through a list, which helped per-print but did not change what fit inside Friday. The 2026 version splits the list.

flowchart TD
  A["Zip file: 45 prints and STEP models, bids due Friday 5pm"] --> B["Job split into four part families"]
  B --> C["Turned parts: 14 prints"]
  B --> D["Milled parts: 19 prints"]
  B --> E["Assemblies and weldments: 7 prints"]
  B --> F["Parts needing outside plating or anodize: 5 prints"]
  C --> G["One estimating worksheet, one row per part number"]
  D --> G
  E --> G
  F --> G
  G --> H["Estimator prices, marks 6 no-bids, signs and sends"]

Tuesday morning at the estimating desk

Monday 7:10 a.m. The estimator drops the whole zip in, along with two things he would never have bothered to attach before: the customer's supplier quality requirements document, and the CSV of what you quoted this customer last year. He asks for one worksheet, one row per part number, with material, revision, tightest tolerance, geometric callouts, thread specs, outside processes, whether AS9102 first article is called out, stock size from your standard bar list, and last year's price if the part number appears in the history file.

He goes and does the Monday production meeting. When he comes back the worksheet exists — 45 rows, with 11 rows flagged in a column called "needs a human", which is the column that actually earns its keep. Row 22: revision on the print does not match the revision on the purchase order line. Row 31: the note calls for passivation per AMS 2700 Method 1, which you send out. Row 38: the 1998 scan, unreadable in three places. Row 40: tolerance stack on a datum-referenced bore that wants a real conversation with the customer's engineer.

Tuesday is spent on those 11 rows and on pricing all 45. The tricky ones get the estimator's full attention because he did not burn Monday and Tuesday reading title blocks. Wednesday the quote goes out — two days before the deadline, which in this trade occasionally gets you a phone call from the buyer asking a question, and a phone call from a buyer is worth more than a sharp number.

The arithmetic: count bids submitted, not hours saved

Hours saved is the wrong measure here because the estimator's week was full either way. The measure is how many line items you actually got a price on. Assumptions below are illustrative — put your own hit rate and average order value in.

Against that, the run cost a few dollars of model use plus roughly six hours of estimator review on the flagged rows. Cut the marginal hit rate in half to be conservative and the extra bids are still worth about ten thousand dollars a year of work you were handing to the shop across town. The 45 rows are also a record of what this customer's parts demand, which makes next year's re-quote faster again.

One caution on the arithmetic. Bidding more is only good if you are not bidding badly. If the extra 19 come back at prices that cannot be run, you have bought yourself a year of unprofitable spindle time. Which is why the flag column, and the estimator's signature, are not optional.

The prints that still go to a person first

Anything with geometric dimensioning and tolerancing that drives fixturing is a human read. A profile-of-a-surface callout referenced to three datums is a fixture design conversation, not a data extraction task. Anything where the print revision and the purchase order revision disagree goes back to the buyer before it gets priced, every time. Old scans with handwriting in the margin — the situation is better than it was two years ago but a wrong material callout read off a faded 1998 print will follow you through a whole production run.

Cycle time estimating stays human at this shop size. The assistant can tell you the part has 14 holes, two counterbores and a 3/8-16 tapped hole; it cannot tell you that your operator does that family in one op on the pallet changer because you built the fixture two years ago. That knowledge lives on your floor and it is the reason your number beats the shop across town.

And anything with export-controlled technical data on it stays inside whatever boundary your quality manager has written down. Speed is not a reason to move that line.

Start with the RFQ you are already planning to partly ignore

Do not pilot this on a three-line job. Take the next re-quote with more than 20 part numbers, run the split, and compare the flagged rows against your own reading of the same prints on ten of them. If the flags catch what you would have caught, you have a tool. If they miss a first-article note, you have learned exactly where the human column has to sit. Either outcome is worth one Monday.

Frequently asked questions

Does this work on prints, or does it need the model files?

It reads the PDF prints, including the title block and the notes, and having the STEP models alongside makes the feature counts more reliable. Neither replaces a real look at the geometry when the part is complicated. Treat the output as a well-organized first pass by a sharp new hire, not as a quote.

We use Paperless Parts already. Is this the same thing?

Quoting platforms handle the workflow, the customer portal and the pricing math, and several do their own feature recognition. What is new in 2026 is handing an entire mixed RFQ package — prints, a supplier quality manual, last year's pricing, purchase order terms — to several agents at once and getting one merged worksheet back.

What if the customer's supplier quality manual contradicts the print?

That is a real and common finding, and it should land in the flag column rather than being silently resolved. The correct answer in a shop is a clarification email to the buyer before pricing, with the manual section and the print note quoted side by side. Getting both documents read in the same pass is the actual gain here.

How many parts is too many for one batch?

The constraint is not the number of prints — it is how many flagged rows your estimator can work through before the deadline. If a 90-line RFQ comes back with 30 flags and you have two days, you are still triaging. Plan the human hours first.

Quoting faster means more buyers calling back with questions, often while everybody is on the floor. CallSphere builds AI voice and chat agents that answer the shop line and web chat 24/7, take the part number and the question, and book a callback with your estimator instead of letting it go to a voicemail box nobody clears until Thursday. The pricing is still yours. The point is that a buyer with a question at 4:45 on a Friday reaches a live answer.