Honesty page
What Staffbox does not do
Short answer: No AI phone agent, no sending, no attachments, no certification, no customer yet. It drafts replies to written requests; a person sends each.
People reach this page looking for an AI that answers the phone, sends email or runs an office. A site that lists only features leaves you to discover the gaps in a sales call. So here are the limits, split into the deliberate ones and the honest gaps. Send up to three requests; drafts come back within 2 business days.
Updated 14 October 2026
The software's limits are deliberate: it cannot send, delete or take a call
| Thing | Status | Why, or what would change it |
|---|---|---|
| Answer the phone or take voice calls | No. The AI worker has no phone or voice channel (you can still call us). The company phone, 404 493 6248, is answered by a person. | Not offered, and we make no claim there. |
| Send email | No. There is no sending code and no mail credentials. A person sends every reply. | By design. See human approval. |
| Delete or change your files | No. In the quote setup it can write drafts and a log line, and has no permission to send or delete. | Set at install. |
| Run commands, browse the web, write code | Off in the default install. The installer turns off nine tool groups plus image generation: terminal, code execution, browser, computer use, web, connections, delegation, cron jobs and text-to-speech. | Stays off unless the install is deliberately changed. |
| Use a cloud AI service | Off by default. Models run on the box. A cloud provider is used only if you ask in writing and supply your own key. | Your choice, your key, your bill. |
| Train a model on your data | No one trains a model on customer data. | Stated in the security sheet. |
| Read attachments, PDFs or drawings | No claim. The intake reads the plain-text part of an email. | Not tested. See RFQ email intake. |
| Follow up on quotes, match purchase orders, draft invoices, close the month | Not built, and no tests exist for them. Back-of-house work is listed as planned on the how it works page. | Needs a feature and a scorecard before we say more. |
| Connect to your ERP or CRM as a standard feature | Per-site work, planned. No standard connector is claimed. | Decided per site. |
| Price wire by the foot with reel, coil and cut rules | Not handled yet (found in the 14 Oct electrical test). It read an ordered length such as "500 feet" as a reel size and refused 7 stocked items. | Product work, not shipped. Scored on a blind plumbing & HVAC set when it is. |
| Apply contractor or account price levels | Not handled yet (found in the 14 Oct electrical test). It quoted list price to contractor accounts; 4 wrong totals. | Product work, not shipped. On the blind plumbing & HVAC set (14 Oct) a fix still made 11 wrong prices in 55 where account levels, the refrigerant surcharge and quantity breaks interact (the blind result). |
| Match trade shorthand ("12-2 nm", "sp 20a", "2p 30a") | Not handled yet (found in the 14 Oct electrical test). It refused 7 messy requests it could not match to a row. | Product work, not shipped: trade synonyms per part family. |
| Respect stock and special-order flags on the list | Not handled yet (found in the 14 Oct electrical test). It priced 3 special-order items instead of declining them. | Product work, not shipped. |
The point of the list is that none of these are quiet. If a feature is not on the page that describes what Staffbox does, assume it is not there.
The company's gaps are real: no customer, no certification, no entity yet
| Gap | Where it stands | What would close it |
|---|---|---|
| A customer | No customer install and no paying customer yet. Every score on this site is from fictional test companies. | The first founding site. |
| A certification | No SOC 2 or other certification. | No date is published. |
| Insurance | No cyber or errors-and-omissions insurance is confirmed. | Open. |
| A company entity | Staffbox, Inc. is in formation in Delaware and is not yet incorporated. We plan to incorporate when the first customer signs; until then the founder operates it. | The first signature. |
| Mac Studio numbers | Mac Studio is planned and not yet measured. A 27B-class model scored 40 and 39 of 40 on our GPU test box, which is other hardware. | A measurement on a Studio. |
| gpt-oss speed on Apple hardware | Not measured. The model scored well on our GPU test box, and needs a 32 GB Mac. | A run on a 32 GB Mac mini. |
The team is six people, staff and contractors, across product, sales, engineering, design, marketing and research. Only the founder administers systems that hold customer data. We say this because a page that lists limits should not hide how small we are.
Every test is from fictional companies, mostly on one 16 GB Mac mini
- Every number is from our own tests on five fictional companies: an IT provider, a cabinet shop, an electrical distributor, a plumbing and HVAC supply house and a lumber yard. Every Mac mini number is from one 16 GB M4; some model comparisons ran on a Dell GPU workstation. See the scorecard.
- On the unseen electrical distributor, run once with no tuning, it scored 30 of 40 and 2 of 15 on messy requests, with 8 wrong prices in 55 (results file). On a blind plumbing & HVAC distributor it scored 26 of 40 and 3 of 15 with 11 wrong prices in 55, and failed our no-wrong-prices bar (the blind result). That is why we score your own list before anything is installed.
- It does not price a new trade out of the box. On a third blind trade, lumber and building supply, it made 15 wrong prices in 55, and 16 with a fix that had reached 0 on the two trades it was built against. Each site gets its notes built on its own price list and scored on its own past quotes before go-live (the third blind trade).
- Fifteen unseen messy requests is a small sample. 360 generated requests add volume, but they are a generated regression set (requests written by a model, answers checked by code).
- The worker handles one request at a time. Concurrent requests queue.
- Misses exist, and we have counted them: misses and refusals.
What it does do fits in one paragraph
In one paragraph: Staffbox reads a quote request from a folder, finds the items on your price list, does the sums in code, and writes a draft that cites the row and the list's date. It sorts email into queues from a rule you write. It answers policy questions from your notes. A person checks and sends. The whole path is in how to quote from a price list, and the product page is AI quoting software. If it is not what you need, we would rather say so on the phone than sell you a pilot: 404 493 6248.
If you came looking for something else
Some people reach this site searching for an AI that answers the phone, an AI that sends your emails, or an AI that runs a whole office. Staffbox is none of those today. It drafts replies to written requests from your own documents, mainly quotes from a price list, and a person checks and sends every one. This page lists what it does not do, with the reason and what would have to change. It is meant to save you a call if you need something else, and to show that the limits are deliberate where they are, and honest where they are gaps.
Questions
Can Staffbox answer my phone?
No. The AI worker has no phone or voice channel (you can still call us). The company phone line is answered by a person. If you need an AI phone agent, Staffbox is not that product.
Does Staffbox have SOC 2?
No. There is no SOC 2 or other certification yet, and no cyber or errors-and-omissions insurance is confirmed. The Trust Center lists every control and its status.
Does Staffbox have customers?
Not yet. There is no customer install and no paying customer. All published scores come from fictional test companies. Every Mac mini number is from one 16 GB M4; some model comparisons ran on a Dell GPU workstation.
Sources
- Product and company status: Staffbox's own records, 13 Oct 2026; Trust Center at trust.staffbox.ai.
- Default tool set and mailbox behaviour: github.com/Staffbox-ai/staffbox, read 13 Oct 2026.
Or call us: 404 493 6248 · How the free test works