AI quoting by industry: what is tested, what is not

Three segments have scores, all from invented companies, and the newest one went badly

The offer is written for firms that send at least 8 quotes a week from a price list: distributors, fabricators, cabinet and millwork shops, and IT providers. The IT provider, the cabinet shop and, on 14 October, an electrical distributor and a blind plumbing and HVAC supply house, and on 15 October a blind lumber yard, have been run, and we wrote all five companies, their price lists and their requests. The runs are real; the companies are not.

How we measured: fictional companies, graded by code, Qwen3 8B on a 16 GB Mac mini M4 (Dell RTX 3090 for the 30 Sep notes test). Fifteen requests is a small sample. Source: results files on GitHub; every graded answer is on the Hugging Face scorecard.

Every segment runs the same loop, and a person always sends

A quote desk that works from a price list repeats one loop in every trade: an email arrives, someone finds the rows, multiplies, adds options and writes back. Staffbox does that loop up to the draft and stops. Each priced line cites the price-list row and the date of that list. Where a fact is missing, the draft asks the customer instead of guessing, and in the score a refusal still counts as a miss. This part is built into the software, not a segment result, which is why it holds for the untested trades too.

Staffbox has no sending code and cannot delete anything, so the reply always leaves from your mailbox after a person approves it. Who does what at each stage, for you, your IT provider and Staffbox, is drawn once on the company page.

Three segments are targets, and their pages describe requests, not results

What differs between trades is the shape of the list and the shape of the request. A distributor has price levels, quantity breaks and units of measure. An electrical counter sells wire by the foot or the reel against copper-linked prices; on the electrical list we tested, those two caused 11 of the 23 misses, and they are likely in other supply trades too, though untested (see the electrical result). A plumbing and HVAC counter matches fitting sizes and equipment model numbers. A fabricator works from rates instead of rows. Each target page draws one invented request and marks what the draft has to read in it. Those drawings are our read of the trade, and each one says so.

• Tested on fictional companies. We wrote the company, its price list and the requests, ran them on a 16 GB Mac mini M4 (the electrical and plumbing and HVAC sets on our Dell RTX 3090 test box), and graded by machine. The scores prove the mechanics, not your list.
• Target segment, not tested. The offer names the segment, but no list from it has been run. The page says so in its first lines.
• No page. Print shops and restoration contractors have no page, because we have no evidence and would rather not write one. Lumber yards and building-materials dealers have none either: one invented yard, scored blind once, made 15 wrong prices in 55 (the third blind trade).

Every miss from the fictional runs is listed in public, so you can see how the draft fails before you decide whether your trade is close enough.

Your own 30 past quotes are the test that settles your segment

You send the price list and 30 past quote requests. We run them on our own Mac mini in Atlanta, your team grades 10 held-back answers blind, and we delete the files within 14 days of sending your results (per our draft data policy). The pass bar is agreed before day 0. Where to begin depends on how your quotes are built:

• If your quotes are lookups and arithmetic from a price list, start with the free 30-quote test on your own list. It costs nothing and installs nothing.
• If your quotes start from drawings or a site visit, expect Staffbox to help only after the takeoff.
• If you need entry into an ERP, or want to quote by phone, read Staffbox vs Mercura before you decide.
• If you are an IT provider thinking of offering Staffbox to clients, start with the partner page.

Sources

• Scores, test sets and every graded answer: Staffbox scorecard on Hugging Face and github.com/Staffbox-ai/staffbox (evals/results). Staffbox claims on this page come from our public scorecards (huggingface.co/spaces/staffbox/scorecard) and the open repo (github.com/Staffbox-ai/staffbox).
• 6 Oct 2026 per-set results and the miss breakdown: evals/results on GitHub.