Glossary
AI quoting glossary
Short answer: the 22 terms describe one chain, from request to brain, model, draft, person and scorecard.
Vendors use words like local, open-weight and scorecard loosely, which makes two AI quoting offers hard to compare. Each definition below pins a word to what Staffbox actually does, grouped by where it sits in that chain. Every term has its own anchor, so you can link straight to it. Send up to three requests your team already answered; drafts come back within 2 business days.
Updated 14 October 2026
Your business knowledge lives in a brain you own, not in the model
The model brings language; your notes and price list bring the facts. Without the brain, neither local model we tested scored above 2 of 40 on the Fieldstone IT set.
Brain
The notes, rules and examples the worker reads about your business, kept as a folder of Markdown files (an Obsidian vault) that the customer owns. Without the brain, the 8B model scored 1 of 40 on our Fieldstone IT test and the 27B model 2 of 40. With the brain alone they scored 33 and 39 of 40; with the brain plus the quote action, both scored 40 of 40. Those runs were on the Dell on 30 September 2026.
Price-list row
One line of your price list: an item, its unit and its price. Every priced line in a Staffbox draft cites its row and the date of the list, so a person can check it in seconds. See prepare a price list.
The model runs on a Mac you own, through open software you can read
These terms cover where the work happens and who wrote the parts. The model, the server that runs it and the agent around it all sit on the box in your building.
Local AI
An AI model that runs on a computer you own, in your building, instead of on a vendor's servers. In Staffbox the models run on a Mac mini through Ollama, and the model server listens on 127.0.0.1 only. Cloud AI is off unless the customer adds its own key in writing. See local AI on a Mac mini.
Ollama
Free software that runs open-weight models on your own computer. Staffbox uses it on the Mac mini, with the server on 127.0.0.1.
Open-weight model
A model whose trained weights you can download and run yourself. Qwen (Alibaba) and gpt-oss (OpenAI) are examples. Open-weight is not the same as open source: licences differ, so check the one on the model you choose.
Qwen
Alibaba's family of open-weight models. It is Staffbox's default model class. Buyers who cannot use Alibaba products, such as Georgia state agencies under SS-22-002, can choose gpt-oss-20b instead. See Georgia AI procurement rules.
gpt-oss-20b
OpenAI's open-weight 20-billion-parameter model. On the Dell it scored 39 of 40 on Fieldstone IT, 40 of 40 on Peachtree Cabinet Works and 15 of 15 on a messy set. It needs a 32 GB Mac. Speed on Apple hardware is not measured yet. 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).
Context window
How much text the model can hold in view at once, counted in tokens. Hermes Agent wants at least 64K. Qwen3 8B's native window is 40,960, so the Staffbox installer sets OLLAMA_CONTEXT_LENGTH=65536 on the Ollama server and 65536 in the Hermes profile. All Mac mini scores were measured at that setting; prompts longer than 40,960 tokens have not been tested. See the 64K context page.
Hermes Agent
The open-source (MIT) AI agent from Nous Research that Staffbox is built on. Staffbox is not affiliated with Nous Research. See the Hermes Agent hub and the official repository.
MIT licence
A permissive open-source licence. Staffbox's agent code and Hermes Agent are both MIT. It lets you read, run and change the code.
FileVault
Apple's full-disk encryption on macOS. The Staffbox installer stops if it is off (scripts/install.sh in the public repo). Check it with fdesetup status.
The output is a draft that cites its row, and a person sends it
What comes out of the worker is never a sent email. These terms describe how a draft is built and what it does when it cannot price something.
Quote action
The part of Staffbox that turns a request into a priced draft: it finds the items, looks up each price-list row, does the arithmetic and writes the reply. It flags an item you do not stock instead of substituting the nearest one (by design; 0 wrong totals in 475 test answers).
Unstocked-item flag
The tool says an item is not on the price list and leaves it unpriced. It does not swap in a similar item. Before version 2.4, 6 wrong totals on 5 October 2026 all came from substitution or skipping on such requests.
Refusal
When the tool declines to price a request and asks for more information instead. A refusal counts as a miss on the scorecard. On 6 October 2026, 10 of 11 misses were refusals, which is the safer way to miss. See AI misses and refusals.
Drafts-only
The worker writes draft replies into a folder. It cannot send email or delete files, and Staffbox has no sending code. A person checks and sends every reply. See human approval of AI drafts.
A score on your own past quotes decides whether anything goes live
Nothing is installed on trust. These terms describe how a test is run and read, so a score from us can be checked against your own grading.
Scorecard
The report from a test run. It shows the share of requests priced, the share right with no edits, the minutes spent editing and the 90th-percentile time to a draft. Every graded answer from our runs is public in the scorecard Space. See the accuracy scorecard.
Pass bar
The score a pilot must reach before anything goes live. The customer agrees it before day 0. A pilot that does not meet the bar does not go live. See how to run an AI pilot.
Blind grading
Marking answers without knowing whether the tool or a person wrote them. In the free score, 10 of the 30 past quotes are held back and graded blind by the customer's team.
Held-out test
Test cases the tool was not tuned on. If you fix the tool on a case, that case is no longer a fair test. Our messy set 1 (19 of 20) was seen while fixing; messy set 2 (15 of 15) was unseen.
90th-percentile time
The time within which 90 of every 100 answers arrive. We report it next to the average because a slow case matters more to a person waiting than a fast average. On 6 October 2026 it was 5.3 seconds on Fieldstone IT.
The numbers in these definitions come from fictional demo companies graded by machine. No customer has run Staffbox. Read the guides for the method, and the pilot guide for what to do with a score.
You buy it through an IT provider, at a founding-site price
Two terms cover who installs the box and what the first sites pay.
MSP (managed service provider)
An IT company that runs a client's technology for a monthly fee. Staffbox is sold and supported through MSPs. See the MSP AI offer.
Founding site
One of the first 100 sites. After a free 30-day pilot it is $595 a month at founding sites (first 100) for as long as the box is live (draft pricing), with no onboarding fee. From unit 101 the price is $695 a month plus $1,800 onboarding.
Questions
What is a local AI worker?
A local AI worker is an AI model and a set of tools that run on a computer in your own building. Staffbox runs on a Mac mini and drafts quote replies from your price list. A person sends them.
What does drafts-only mean?
The worker writes drafts into a folder. It cannot send email or delete files, and Staffbox has no sending code. A person checks and sends every reply.
What is a pass bar?
The score a pilot must reach before anything goes live, agreed by the customer before day 0.
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.
- Measured results of 30 Sep, 5 Oct and 6 Oct 2026: evals/results on GitHub.
- Hermes Agent by Nous Research (MIT).
Or call us: 404 493 6248 · How the free test works