Guides
Guides for running AI on a quote desk
Short answer: Start with the pilot guide: write the pass bar before day 0, then score any AI tool on 30 past quotes, 10 graded blind.
People who decide, set up or check an AI quote desk need steps, not a pitch. Vendor pages rarely say how a score was produced or who graded it, which makes tools hard to compare. These guides give the method, with every Staffbox number traced to a measured run or marked as not measured. Send up to three requests your team already answered; drafts come back within 2 business days.
Updated 14 October 2026
Your role decides which page to open first
An owner, a sales desk, an IT provider and a public buyer ask different first questions, so each has a different first page.
- Owner or office manager: how to run an AI pilot.
- Sales desk, answering quotes: prepare your price list.
- IT provider (MSP): the MSP AI offer for clients.
- Georgia state buyer: Georgia AI procurement rules.
The table adds the next steps and two more roles.
| You are | Start here | Then | When you're ready |
|---|---|---|---|
| An owner or office manager deciding whether to try it | How to run an AI pilot | Prepare your price list, then the free 3-request test | Send up to three requests, free |
| Someone on the sales desk who answers quote requests | Prepare your price list | Quote from a price list, the free 3-request test | Send up to three requests, free |
| An IT provider thinking about an AI service for clients | The MSP AI offer | Install and support checklist, security answers | Get the draft partner terms |
| A security or compliance reviewer | Security questionnaire answers | Trust Center, glossary | Get the draft partner terms |
| A Georgia public office | Georgia AI procurement rules | Government page | Send up to three requests, free |
| An engineer who wants the mechanics | Hermes Agent hub | Models and scores, how it works | Talk Hermes with us |
Two step-by-step guides and a glossary cover a pilot from first file to decision
- How to run an AI pilot: write the pass bar before day 0, grade blind, count refusals as misses, and do the break-even arithmetic on one page.
- How to prepare a price list for AI: columns, units, dates and the six ways price lists break a reader.
- We tested on a trade it had never seen: the electrical write-up, 30 of 40 run once, with every miss and what we are doing about it.
- AI quoting glossary: local AI, brain, drafts-only, quote action, scorecard, pass bar and sixteen more, each with an anchor you can link to.
Related pages that behave like guides: quote from a price list, human approval of AI drafts, the accuracy scorecard, AI misses and refusals and what Staffbox does not do.
Thirty past requests, a dated price list and one blind grader are the whole entry fee
Every guide here assumes you can put your hands on three things. If you are missing one, fix that first, because no tool makes up for it.
- Thirty past requests that you already answered, with the answers you sent. Email is fine.
- The price list that was current for those requests, with its date.
- One person who will grade ten answers without knowing which came from the tool.
That is the whole entry fee for a scored pilot. The money question comes after the score. For Staffbox it is drafted as: $595 a month at founding sites (first 100) for as long as the box is live (draft pricing), free 30-day pilot, no onboarding fee; $695 a month plus $1,800 onboarding from unit 101.
Four rules decide whether a score, ours or any vendor's, means anything
Scores appear on many pages. Four rules keep them honest. They apply to everything we publish and to anything you read from another vendor.
| Rule | What it means here | Example |
|---|---|---|
| Who was tested | Every score so far comes from fictional demo companies, graded by machine. No customer has run Staffbox. | Fieldstone IT and Peachtree Cabinet Works are invented. |
| What a miss is | A refusal counts as a miss. A wrong total counts as a miss and is worse. | 6 Oct 2026: 11 misses in 475 answers, 10 declines and 1 routing miss. 0 wrong totals. |
| Sample size | Small sets are small. | The held-out messy sets are 20 and 15 requests. |
| Hardware and speed | Times are for one request at a time on a warm model on a 16 GB Mac mini M4. | Average 2.3 s, 90th percentile 5.3 s on Fieldstone. |
The two results above follow those rules: both come from fictional companies, both count every answer, and each bar carries its date. Where a claim is labelled, goal is where we want to be, and opinion and guess mean what they say.
How we write guides
- Each guide is dated. When a number changes, the date changes.
- Market and competitor facts are labelled "as reported" with a source and a date, and we state the other side's strengths.
- Nothing here is legal, security or financial advice. Pricing is a draft.
- We update guides when a pilot teaches us something. No pilot has run yet, so the guides reflect our test method and not customer experience.
Tell us what is missing: email hello@staffbox.ai.
Index: every guide on this site, by topic
This is the master index, in seven groups. Each topic has a hub page with one line on every guide in it.
Hermes Agent
- Hermes Agent for business
- Install Hermes Agent on a Mac mini
- Hermes Agent on a Mac mini
- Hermes Agent with Ollama
- Best local model for Hermes Agent
- Hermes Agent 64K context
- Is Hermes Agent safe for company data? What we turn off
- Hermes Agent email
- Hermes Agent Obsidian vault
- Hermes Agent managed hosting vs DIY vs IT provider
- Hermes Agent hosting alternative
- Hermes Agent for MSPs
- Hermes Agent skills and MCP
- Non-Chinese model for Hermes Agent
- Hermes Agent updates
- Hermes Agent FAQ
Quoting workflows
Industries
Comparisons
IT providers
Atlanta & Georgia
Questions
Where should an owner start?
Start with the pilot guide, then prepare your price list, then send three past requests for the free 3-request test.
Do the guides depend on Staffbox?
The pilot and price-list guides apply to any AI quote tool. The score method, blind grading and written pass bar are the parts worth copying.
Has anyone run a pilot with these guides?
No. No customer has run Staffbox yet. The guides reflect our own test method.
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.
- Every graded answer: the Hugging Face scorecard.
Or call us: 404 493 6248 · How the free test works