AI quoting for distributors: price levels, breaks and units

Every distributor request hides five things the draft has to read

Take one ordinary email from a trade account. Before a single line can be priced, the draft has to pin down the part, the quantity, the unit, the account's price level, and whatever the customer left out. The drawing below marks each of those on a made-up request. It is a picture of the job, not a result.

Requests at a distribution counter come in a handful of shapes. This table is our read of the segment, not something measured on a distributor.

Price levels, quantity breaks and units set the total, so they belong in the brain

In our tests the notes mattered far more than the model. On the Fieldstone IT test set (GPU test box, 30 Sep 2026), Qwen3 8B with no notes scored 1 of 40, with notes 33 of 40, and with notes plus the quote action 40 of 40; the 16 GB Mac mini matched the 40 of 40 on 6 Oct 2026. For a distributor, that means the effort goes into writing your pricing rules down where the draft can read them. These six usually decide it:

• Price levels. List, contractor, wholesale and customer-specific columns. The draft must know which level this account gets. Not handled yet: in the 14 Oct electrical test it quoted list price to contractor accounts even with the level written in the notes.
• Quantity breaks. A price that changes at 10, 100 or a pallet. Totals depend on picking the right break, and the brain should say whether a break applies per line or per order.
• Units of measure. Each, box, case, foot. A request for "a box" against a per-each price is a classic wrong total, so the case pack belongs on the row. Not handled yet either: the 14 Oct electrical test priced 300 wire nuts as 300 boxes.
• Superseded and discontinued part numbers. The list should say what replaced what, or the draft should flag the part.
• Effective dates. Staffbox cites the date of the list it used. If you reprice monthly, a draft that cites an old date is easy to spot.
• Freight and minimums. Surcharges and minimum order values are rules, and rules live in the brain, in plain Markdown.

None of these six has been tried on a real distributor list. Finding the ones that break on yours is what the free score is for.

The draft stops where your price list stops: no stock, no ERP, no sending

When the list does not answer a question, the draft is built to say so rather than fill the gap. In the 6 October runs on the fictional companies, 10 of the 11 misses were declines that asked the customer, and none priced a wrong total. For a distributor, the edges are predictable:

• Stock and ship dates. A price list does not know what is on the shelf. Those questions go to a person.
• ERP entry. The pilot reads one price-list file and writes drafts into a folder. It does not post into your ERP. If ERP entry is the job, read Staffbox vs Mercura.
• Reading specifications or drawings. We make no claim there.
• Sending. The unit cannot send or delete anything. Speed of reply is bounded by the person who approves.

Every miss is listed in what the AI got wrong, in public.

Thirty of your past quotes, graded blind by your desk, show whether it fits

The demo scores below show that lookup, totals and declines work on invented lists. Machine-graded, one request at a time with the model warm, on a 16 GB Mac mini M4 running Qwen3 8B. They say nothing about your list.

On that unseen electrical distributor, run once with no tuning, per-foot units and contractor price levels broke it; both are likely common on a distributor list too, though untested on one (see the electrical result).

Fifteen requests in one set is a small sample, and the generated requests were written by a local model, not by customers. Your own quotes replace all of that. For the score, send:

• One price-list file, the one your desk really uses. If you keep two, tell us which wins.
• 30 past quote requests as they arrived, with the quote you actually sent. Pick a mix, not only the easy ones.
• A note on any quote where you gave a special price, and why. We would rather learn the rule than guess it.
• Three requests you think are hard, ideally one that crosses a quantity break and one asked in cases against a per-each price.
• Names of the 10 you want held back. Your desk grades those blind, so you do not have to trust our grading.

It is free and nothing is installed: the files run on our own Mac mini in Atlanta and are deleted within 14 days of sending your results (per our draft data policy). The pass bar is written down before day 0, and a refusal is scored as a miss.

Related: electrical supply houses (wire, breakers and copper-linked lists) · plumbing and HVAC supply (fittings, equipment models and cross-references) · all industries · RFQ automation · quote from a price list · the free 30-quote test.

Sources

• How we measured: two fictional companies (Fieldstone IT, Peachtree Cabinet Works), graded by code, Qwen3 8B on a 16 GB Mac mini M4 (Dell RTX 3090 for the 30 Sep runs), 30 Sep to 6 Oct 2026. No customer has run Staffbox yet.
• Scores, test sets and every graded answer: browse every graded answer on this site, the 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).
• For a dated comparison with a hosted distributor quoting platform, see Staffbox vs Mercura.