02 — Why it matters

Why offices are moving AI off the cloud

Three separate pressures are pushing businesses toward a private AI box, and they tend to show up in this order.

The confidentiality problem cloud AI can’t fully solve

Some material simply can’t leave the building — attorney-client correspondence, patient records, unreleased financials, proprietary client data. Even the “enterprise” tier of a major cloud AI tool, the one that promises not to train on your conversations, still routes every message through that company’s servers, subject to its outages, its policy changes, and its breach history. A private AI box sidesteps the question entirely, because there’s nothing to hand over in the first place.

The subscription math never stops

Per-seat AI subscriptions are cheap-looking until you multiply them. Ten people at roughly $30 a month each is over $3,600 a year — and that number is set by the vendor, who can raise it whenever renewal comes around. A private AI box is typically bought once and covers the entire office at no additional per-login cost, for as long as the hardware runs.

Working without depending on someone else’s uptime

A box on your own network keeps functioning if the office internet drops, if a cloud vendor has an outage, or if a subscription terms page quietly changes overnight. The intelligence lives in the building, not in someone else’s status page.

The question every business eventually asks about a cloud AI tool is: where does what we type actually go? A private AI box is the answer that doesn’t require an asterisk.

03 — Mechanics

How a private AI box actually works

Nothing about it is exotic — it’s the same idea as any office server, applied to AI instead of file storage or email. The box plugs into your existing router like any other device on the network. It arrives with the AI engines and any remote-access app already installed, and switching it on starts a private web app at a friendly local address that anyone on the office Wi-Fi can reach from an ordinary browser — no download required.

A software edition works the same way conceptually, just installed on a Mac, Windows, or Linux machine you already own. The installer checks the machine’s specs and sets itself up in a few minutes; once it’s done, it opens the same office AI in a browser, typically at a local address on that computer.

From there, every request — a chat message, a document upload, a request for a first draft — is handled entirely by the box’s own processor and memory. The model itself is a file that lives on the machine: it was trained on an enormous amount of text before it ever arrived at your office, and that learning is compressed into the model the way a well-read colleague carries knowledge around without needing to look everything up. General reasoning is built in from day one. Anything current, or specific to your business, comes from what your team uploads — or from an optional web-lookup add-on.

Two things are worth being explicit about, because they’re the whole point of the category. First, a normal day of use — chat, document answers, scheduling, drafting — requires zero internet connectivity. Second, the only things that ever reach outside the building are things you deliberately turn on: an optional live web-search add-on, an optional remote-access tunnel, sending a new-teammate invite email, and a small license or update check that never carries any of your content. That last point matters enough that it gets its own section further down.

For a deeper walkthrough of the chat experience specifically — deep-thinking mode, file attachments, offline operation — see how Private AI Chat works, or browse the full docs & help center for the technical detail behind every feature.

04 — Engines

Choosing how much thinking power to run

Inside the box, the “engine” is the actual AI model doing the thinking, and a well-built system offers more than one tier so you’re not always paying the performance cost of the biggest model for a one-line question. Switching engines is normally instant on an appliance, because every tier ships preloaded; on a software install, moving to a bigger engine downloads that model in the background the first time, which can take a while depending on the connection.

That last row matters more than it looks: if a screenshot or a scanned PDF isn’t being read correctly, the near-universal cause is simply being on Swift or Balanced, which are text-only. Switching to Maximum in settings resolves it immediately on a box, since the engine is already sitting on the hardware.

A well-built system typically starts a brand-new setup on the lightest engine, Swift, so it’s usable the moment setup finishes — with an option to move up any time from settings, gated behind a “change engine” permission so not every team member can quietly shift the whole office onto a heavier, slower model.

05 — The economics

Private AI box vs. cloud subscriptions vs. leased appliances

There are, broadly, three ways a business gets AI into daily use: a per-seat cloud subscription, a leased private-AI appliance billed monthly per user, or a private AI box purchased outright. The gap between them widens fast as headcount grows.

A one-time purchase changes the shape of the decision entirely. A subscription renews forever at a price set by the vendor. A leased appliance multiplies that same problem by hardware costs. A private AI box, by contrast, is a fixed cost that covers everyone in the office regardless of how the team grows — see the full pricing breakdown for exact tiers.

06 — What’s inside

What should actually be inside the box

“Private AI box” describes the hardware; what makes one actually useful day to day is the software running on top of it. A serious appliance bundles considerably more than a chat window. Here’s what to look for, and what each piece is for:

• A ChatGPT-style chat assistant as the home screen — start conversations, rename them, and share a specific chat with specific roles, all with a searchable history that never leaves the network.
• A company knowledge base that answers questions straight from your own policies, contracts, and files, citing the exact source document. Uploading, reading, and answering all happen on the box, across PDF, Word, Excel, PowerPoint, plain text, and CSV files up to 25 MB each.
• Team accounts and permissions so every employee gets a login with no per-seat fee, and access can be scoped by role — covered in detail below.
• Shared team chat rooms, where a group works with the AI together, like a group chat with the assistant sitting in the room.
• A document and email generator that produces ready-to-send proposals, quotes, and follow-ups in one click, written in your company’s voice, with room for your own templates.
• A contacts directory and shared calendar that the AI can read from and write to directly — kept in your building’s own timezone, and wired into the rest of the tools.
• Optional web access for the rare question that genuinely needs current information, without exposing anything else. An add-on.
• Optional remote access, so the box’s reach extends to home or the road through an encrypted tunnel, without opening a port on the router. Also an add-on.

A box that only offers chat is really just a smaller, slower version of a cloud tool. The point of a private AI box is that it replaces several separate subscriptions — chat, a Q&A tool, a document generator, scheduling — with one machine your business already owns. See the full feature list for how these pieces work together on a single login system.

07 — Team & permissions

Team accounts and who can see what

Because a private AI box has no per-seat fee, adding people costs nothing extra — the real work is deciding who can see and do what, which is where a properly built permissions system earns its keep.

Adding and removing people

An admin adds a person by name and email; that person receives a temporary password by email and picks their own the first time they log in (if the box happens to be offline and can’t send that email, it simply shows the admin the temporary password to pass along directly). From the same team screen, an admin can promote someone to Administrator, reset a forgotten password, or disable someone’s access instantly — disabling turns off access without deleting their chats or files, since there’s deliberately no permanent-delete button hiding in the app.

The roles that come built in

Every workspace starts with two roles: Administrator, with full access to everything including billing and settings, and Team member, the default for new people, who can use chat and see whatever’s been explicitly shared with them. Beyond those two, most systems offer ready-made add-on roles — a Manager who runs the team and knowledge base but not billing or full admin, a Contributor who uses chat and builds out the knowledge base, a Basic user limited to chat and reading shared knowledge, and a Guest restricted to chat only. Fully custom roles can usually be built by ticking exactly the permissions wanted, and a person can hold more than one role at once.

08 — Reachability

Reaching the box, on-site and remotely

On the office network, a box is normally reachable in three interchangeable ways: a custom web address you claim once (something like yourcompany.getprivateofficeai.com, with automatic HTTPS), its friendly local network name (like yourbox.local), or its raw network address (like 192.168.1.50). A software install typically answers at a local address on that specific computer instead, such as localhost:8080.

Using it away from the office

Reaching the box from outside the building is usually a separate, optional add-on rather than something baked in by default — for good reason. Rather than opening a port on the office router (which exposes the box to the public internet), a private AI box typically reaches remote devices through an encrypted tunnel built on a private-networking layer such as Tailscale. Turning it on means signing into a free account on that networking layer once and approving the box, after which the panel shows a private address — something like yourbox.tailnet.ts.net — that only devices signed into the same private network can ever reach.

From there, any device — a laptop at home, a phone on the road — installs the same free companion app, signs into the identical account, and opens that private address to use the office AI exactly as if sitting at a desk in the building. Turning the add-on off (or letting it lapse) simply stops outside access within about an hour; the office network keeps working normally the entire time, since remote access is additive rather than something the rest of the system depends on.

See Remote Access for the feature overview, or the docs for the exact setup steps on both the appliance and the software edition.

09 — Ongoing costs

Add-ons, billing, updates & the Care Plan

The base purchase is one-time, but a handful of genuinely optional monthly add-ons exist for teams that want them — each cancelable anytime, with no minimum term.

A newly purchased add-on can take up to roughly an hour to activate on a box that’s already running; most systems offer a manual “check now” option, or a simple restart, to apply it immediately instead of waiting. One license generally runs one box, so replacing the hardware means transferring the license via a quick support request rather than buying a new one.

Updates, without the disruption

Every box or software license typically includes a full year of free software and security updates from the date of purchase, downloaded, verified, and installed automatically in the background — checked a few times a day, with no action needed. A well-built system waits until the box is idle before restarting its own app to apply an update, so a conversation already in progress is never cut off; it’s the software restarting, not the hardware rebooting. After the first year, an active Care Plan keeps new updates and features coming — but critically, the box keeps working forever either way, simply continuing to run whatever version it already has if the plan lapses.

10 — Reliability

Backups, power loss, and troubleshooting

A box worth trusting with confidential material should also be boring in the best way: it backs itself up without being asked, and it recovers from the obvious failure modes on its own.

Automatic backups

A well-built box backs itself up roughly once a day — at most once every 24 hours, and only while someone’s actually using it — keeping something like the 14 most recent snapshots stored on the box itself. Nothing rides up to the cloud for this; a mistake stays recoverable without introducing a privacy trade-off to fix it. Rolling back to an earlier snapshot is normally a quick support request rather than something to fumble through alone.

Power loss and restarts

If the power cuts or the box restarts, it should start the office AI back up automatically the moment it boots — no manual restart of any service required.

The most common issues, in practice

• Can’t reach the box on the network — confirm the device is on the same Wi-Fi, then try its local name or its raw network address; a restart of the box clears most one-off issues.
• Remote Access won’t turn on — confirm the add-on is actually active, and that the private-networking companion app is installed, running, and signed in.
• An add-on isn’t showing as active yet — give it a few minutes, use a “check now” option if one exists, or restart the box.
• Images or screenshots aren’t being read — almost always means the engine is set to Swift or Balanced; switch to Maximum, the only tier that reads images.
• Forgotten admin password — a proper system supports a local, one-time recovery code generated from the account portal, so the vendor never sees or holds the actual password.

11 — Privacy & security

What actually leaves the building — and what never does

This is the section worth reading closely before buying anything calling itself “private,” because the honest answer is rarely “absolutely nothing, ever.” A trustworthy vendor will tell you exactly what small signals its box does send, rather than making a blanket claim that falls apart under a network trace.

In a well-built system: documents, questions, and answers are stored and processed on the box, in the building, never sent to the vendor or to any cloud AI company. The box is typically firewalled so only the office network — and a private remote-access tunnel, if one is enabled — can reach it. What the box does send home is usually limited to two small, non-content signals: a tiny license-and-version check every few minutes, and roughly once an hour, a slightly fuller health check covering things like which engine is selected and whether it’s running, disk space, total memory, uptime, and the operating system — the kind of information a support team needs to help remotely, not anything that resembles your business data. Separately, claiming a custom web address means registering that address alongside the box’s local network address, purely so the address resolves correctly.

Because everything lives on the box, the data on it is unambiguously the buyer’s — exporting it, or wiping the box entirely before repurposing or returning it, should be a straightforward request to the vendor’s support team, not a negotiation. Full detail on any specific product’s approach is worth reading directly — see, for example, Security and Privacy.

12 — Who needs one

Who a private AI box is actually built for

Any business handling confidential client material benefits immediately, but a few industries feel the need before anyone has to explain it to them.

13 — Buying guide

How to choose a private AI box

Not every product calling itself a “private AI appliance” is actually built for an office to run day to day. A few questions separate a real one from a repurposed mini-PC with a model installed on it:

Does it scale with team size, or does it charge per person?

If a “private” appliance still bills per seat, it hasn’t actually solved the subscription problem — it’s just moved it onto different hardware. Look for a flat, one-time price that covers the whole office.

How much memory does it actually carry?

Memory determines which engine tier a box can run comfortably — the difference between Swift, Balanced, and the image-reading Maximum engine covered above. A box aimed at 15–40 people should carry meaningfully more memory than one built for a five-person office.

Does it answer from your own documents, or only from general knowledge?

General AI knowledge is useful, but the actual daily value comes from a system that can answer “what’s our refund policy” or “what did we quote this client” by reading your own files — with the source cited, not invented.

Is the permissions system real, or just a single admin login?

A one-login-for-everyone setup isn’t fit for an office with HR files, client-specific matters, or anything sensitive. Look for genuine role-based access — the kind described above, where a folder with no role set is visible office-wide by default, so privacy is a deliberate choice rather than an accident waiting to happen.

What happens to updates after the first year?

Most reputable boxes include a year of free software and security updates. After that, confirm whether the hardware keeps working on its existing version indefinitely (it should) and what an optional care plan actually adds.

Can it be reached remotely without exposing it to the internet?

A box that requires opening a port on the office router to be used remotely is a security liability by design. Look for an encrypted tunnel model instead — see how remote access is handled without exposing the appliance publicly.

14 — Setup

What setup day actually looks like

On a well-built appliance, setup is a plug-in-and-log-in process rather than an IT project. Connect the included power and network cable, then open a browser on any device on the same Wi-Fi and go to the address on the setup card. The AI engines and remote-access app are already installed — there’s nothing to download. A software edition follows the same shape: download the installer, let it check the machine, and it opens the same office AI in a browser a few minutes later.

The very first screen typically asks for, in order: a license key (software only — an appliance already has one built in), a company or office name to brand the workspace, an optional custom web address so nobody has to remember an IP, and finally the owner’s name, email, and a password of at least eight characters. The system usually starts on the lightest engine, Swift, so it’s usable immediately, with the option to move up later from settings.

From there, adding the team is a matter of entering a name and email per person — each new teammate receives a temporary password and sets their own on first login. Administrators can promote, reset, or disable any account instantly, which matters more than it sounds: the moment someone leaves the company, their access to every chat, file, and knowledge folder should end immediately, not at the next license renewal.

A box worth buying also backs itself up automatically — daily, stored locally, with no cloud dependency for recovery — and restarts on its own if the power cuts out. For the exact click-by-click steps on every setting mentioned in this guide, the full documentation and FAQ cover it in more depth than any single article can. Ready to see it running with real data rather than a slide deck? Book a live demo, or go straight to building a box for your team.

15 — FAQ

Frequently asked questions about private AI boxes