AI quoting for fabricators: rate tables yes, drawings no

Stock and repeat parts are lookups; a new custom part is estimator work

Our read of the segment, not measured on a fabricator. Three kinds of request reach a shop's inbox, and only the first two sit close to what Staffbox does today.

• Stock or catalog parts. Standard brackets, plate cut to size, tube by the foot, hardware. The price comes from a table, so this is the nearest fit.
• Repeat parts. "Quote 500 of part 4471 again." The earlier quote acts as the price list. That needs your past quotes in the brain, and it is not measured yet.
• New custom parts. A drawing or a sketch, a material, a finish, a quantity. Machine time, programming and setup are an estimator's judgement, and Staffbox makes no attempt at them.

So the boundary is plain before you send anything. Staffbox does not:

• read drawings, 3D models or tolerance callouts;
• estimate machine time, nesting or material yield;
• connect to a shop-floor or ERP system (the pilot reads one price-list file);
• send quotes. A person approves every draft, and the unit has no sending code.

A draft prices only what the table covers and asks about the rest

Here is how one request breaks apart. The plate, the quantity and the unit are things a rate table can answer. The tolerance note and the missing drawing are not, so that part goes back to a person instead of being priced.

On the fictional test sets, that decline is what actually happened when a fact was missing. Across the 6 Oct runs there were 0 wrong totals in 475 answers; of the 11 misses, 10 were declines that asked the customer and 1 was a routing slip. In our view, a shop would rather get a question back on a part with no material called out than a confident number. On a new trade it had never seen, an electrical distributor (14 Oct, Dell test box), the default Qwen3 8B setup scored 30 of 40 and 2 of 15 on messy requests, with 8 wrong prices in 55, because per-foot units, price levels and trade shorthand are not handled yet; that is why we score your own list first (the electrical result).

The rate table has to hold what now lives in the estimator's head

A draft is only as good as the table behind it. For most shops that means writing down, in plain Markdown, rules an experienced estimator applies without thinking:

• Material price by form, grade and size, with the date of the price.
• Process rates, such as an hourly rate by operation, and the minimum charge per order.
• Quantity breaks and setup charges.
• Finish and outside-processing prices, such as powder coat, plating or heat treat.
• What always goes to a person: a tolerance note, a certification requirement, a drawing revision.

If a price is a formula rather than a row, write the formula out in plain words and include past quotes that used it. The demo tests have the same shape (look up rows, multiply, add options, refuse when a required fact is missing), but formula-driven lists are untested, so the first score will tell you whether yours reproduces.

Thirty past requests show whether your tables reproduce your own quotes

• The material and process tables you quote from, as files, each with its date.
• 30 past requests you would call table-driven, plus three that needed an estimator, so we can show you the draft declining them.
• The rules about what a person always checks.
• The quote you actually sent for each, so the draft is compared with your real figure.

Nothing gets installed in your shop for this. The requests and tables run on our own Mac mini in Atlanta, your team grades 10 held-back answers without knowing which are ours, and the files are deleted within 14 days of sending your results (per our draft data policy). A pass bar is agreed before day 0, and a refusal counts against the score just as a wrong price does. We have not tested a fabricator's tables, so expect the first pass to expose a rule nobody wrote down.

Why a cabinet shop is the easier case: its list looks like a catalog with options on top, which is where our evidence is strongest. A fabricator's list is mostly rates and formulas, which makes it a harder test of arithmetic on the brain's rules. That is a guess, not a result; the cabinet page has the scores we do have.

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: 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 results (0 wrong totals in 475 answers; 10 of 11 misses were declines): evals/results on GitHub.