A buyer at a mid-tier OEM in Akron sent the same RFQ to four shops on a Tuesday morning. Three of them — including a strong technical fit — responded between Friday afternoon and the following Monday. The fourth got a quote back by end of day Tuesday. The fourth one won the work, at a number that wasn’t the lowest in the pile.
That’s not a story about price. That’s a story about the chair the buyer was in. He had eleven RFQs out that week, an end-of-quarter forecast meeting on Friday, and roughly the same level of patience as anyone with eleven RFQs out and a forecast meeting on Friday.
Your senior estimator knows what walks in the door is rarely fit to quote. The drawing has a missing tolerance. The material callout is ambiguous. The customer’s PO terms haven’t shown up yet. So the estimator does the only sensible thing: they hold the RFQ in a stack, get to it when there’s a clean ninety-minute window, and that window happens, depending on the week, somewhere between forty-eight and ninety-six hours later.
That gap — the gap between RFQ landing in the inbox and a draft response leaving the inbox — is where most mid-market industrial businesses are losing margin without seeing it on any P&L line.
What the workflow actually does
I’ve built a version of this for three different shops now. The architecture is boring and that’s the point: there are no agents, no fine-tuned models, no bespoke ERP integrations. Here’s the loop:
- An RFQ email arrives in a shared inbox (or gets forwarded to a dedicated address).
- A worker reads the email + any attached PDF or DWG, extracts the specs into a structured format (part name, material, dimensions, quantities, tolerances, finish, delivery date, point of contact, terms).
- The same worker flags anything that’s missing or ambiguous — missing tolerances, unclear material grades, conflicting dimensions on different views, undated quantities.
- It generates a draft response in your voice. Two paragraphs. Receipt confirmation, a numbered list of clarifying questions if there are any, an honest preliminary read on lead time and approximate price if the data is clean enough to support one.
- The whole thing — extracted specs, flagged ambiguities, draft email, link to the source PDF — lands in your estimator’s queue with a clearly-labeled “Draft, not sent” status. Nothing goes out without a human pressing the send button.
End-to-end clock time from inbox to draft-in-queue: about three minutes. The estimator’s time to review and send: about ten to fifteen minutes if the RFQ is clean, longer if there’s real engineering to do — but the engineering happens first, on the buyer’s clock, instead of last, after the buyer has already gotten three other quotes back.
The number that moved at the three shops where this is running: their bid-to-award ratio. Not their margin per quote. Not their throughput on quotes per week. The ratio of quotes won. Across the three shops the lift averaged about eight percentage points over a six-month window. That is, frankly, a lot.
What it costs and what it runs on
The whole thing runs on Cloudflare Workers and the Anthropic API. There is no server to maintain, no enterprise license, no per-seat cost. A typical shop’s monthly bill sits between $40 and $120 — most of which is the model usage, not the infrastructure. The build itself is a fixed-fee engagement; the ongoing cost is small enough that it doesn’t really need a line in the annual budget.
You do not need to rip out your ERP. The workflow runs alongside whatever quoting system you currently use; it produces a draft, your estimator owns the actual quote, and the structured spec extraction can be exported into whatever downstream tool you live in. I have shops running this with E2 Shop System, with JobBOSS, with a spreadsheet, and in one case with a paper traveler that gets re-typed at submit time. The workflow doesn’t care.
What it doesn’t do
This is the part the trade publications skip, so I’ll spend a paragraph on it.
It doesn’t quote the RFQ. It drafts a response to the RFQ — sometimes a clarifying-questions response, sometimes a preliminary read with a number, but it never sends a binding quote without an estimator’s eyes. If a model hallucinates a tolerance, the estimator catches it before it leaves the building. The accuracy of the spec extraction on clean industrial PDFs is high; on scanned drawings with handwriting, it drops to the point where the draft mostly just asks for a cleaner copy.
It also doesn’t do work that requires judgment your senior estimator built over twenty-five years. When the buyer is fishing — sending an RFQ they don’t intend to buy from to anchor their target price — your estimator can usually smell it. The model cannot. So we don’t ask it to.
Finally, it doesn’t replace the relationship part of the work. The buyer who awarded that Tuesday quote still talks to your estimator on the phone — the workflow’s job is to make sure that phone call is first in the day, not buried behind the eleven other inboxes the buyer is processing.
How to start without hiring anyone
If you want to test whether this is worth doing inside your shop before you call me or anyone else: pull last quarter’s RFQs out of your inbox, sort them by date received vs. date responded, and look at the median gap. If you’re under twenty-four hours, you’re already in the top quartile of mid-market industrials and the marginal value of this workflow is smaller. If you’re over forty-eight hours, this is probably the single highest-ROI AI build you can do this year — measured in hours of estimator time, in win rate, and in not having to be the person who tells the owner “we lost the Akron job because we got back to them too late.”
If you do the math and the number is interesting, my number’s on the contact page. First conversation is thirty minutes. The most useful outcome is sometimes that you go do this yourself — the architecture is not a secret, and I’d rather you build it than not have it.