Skip to main content
zatersio
A Zatersio product · On-premises

LegalBox: AI that never leaves the building

A device that runs advanced models inside your firm's own network, on hardware you own. Confidential and privileged matter files never touch a public cloud, and no part of a question is sent to a third-party AI provider.

  • GMKtec EVO X2
  • AMD Ryzen AI Max+ 395
  • 128 GB
  • Zero egress

Network connected

Matter files

on your server

LegalBox

models on board

Public cloud

link up

BYTES SENT OUT: 0

Ask

Does the supply agreement let the buyer cancel late deliveries?

LegalBox · answered on device

Connected, and still nothing goes out. Pull the cable to prove it. Illustrative of the product on synthetic data.

Try it: pull the cable and watch the answer come back anyway.

01 · The regulatory reality

Regulators have already said not to paste it into a chatbot

Australian courts and legal regulators have issued guidance cautioning practitioners against putting confidential or privileged client material into public AI tools. The reasoning is straightforward: once the text leaves your network, you no longer control who stores it, where it is stored, or what it is used for. Vendor assurances are contractual, not architectural.

02 · The position most firms are in

Banned it, or used it anyway

Firms have split into two camps. Some banned AI outright: compliant, and slower than the firm across the road. Others use cloud AI quietly and carry a risk nobody has written down. Both camps are responding to the same missing option, which is a way to get the capability without the exposure.

03 · What LegalBox is

The capability, without the exposure

On your premises

The device sits in your office, on your network, under your physical control. Nothing is rented from a data centre you have never seen.

Verifiable egress

Zero network egress is the default, not a setting buried in an admin panel. Your IT team can monitor the segment and confirm it themselves.

Grounded in your own sources

Answers are retrieved from the firm's own precedents, policies and matter material, with citations back to the source so a lawyer can check the work.

Owned, not subscribed

You buy the hardware. There is no per-seat cloud fee, no metered token bill, and no vendor who can change the terms on data you already handed over.

What's inside the box

Six modules, all running on the device. Matter management, questions with citations, redline comparison, tracked-change drafting, firm playbooks and access control.

Matters

Work is organised by matter, the way the firm already thinks. Documents are attached to a matter and stay scoped to it.

Ask

Ask a question across a matter's documents and get an answer with citations back to the clause it came from, so it can be checked rather than trusted.

Compare

Diff two versions of an agreement clause by clause: modified, added, removed and moved. Export the comparison as a redline .docx.

Draft

Instruction, outline approval, drafting, then review. Every AI revision arrives as a tracked change you accept or reject before it reaches the document.

Playbooks

Encode the firm's own positions and preferred clauses once, so drafting and review follow your standards instead of a generic template.

Admin

User and access control stays with the firm. There is no vendor console on the internet holding the keys to your matters.

The device, running

Screens from the demonstration unit. Every document shown is synthetic: fictional parties, fictional agreements, no client material.

LegalBox Matters screen listing active matters with references, document counts and last-updated times.
Matters: work stays organised the way the firm already thinks. Demonstration data.
LegalBox Compare screen showing a clause-by-clause redline between two versions of a supply agreement.
Compare: two versions of an agreement, diffed clause by clause and exportable as a redline .docx.
LegalBox Draft screen in suggesting mode, with AI revisions shown as tracked changes to accept or reject.
Draft: suggesting mode. Every AI revision is a tracked change a lawyer accepts or rejects.
04 · The test
“If your AI vendor can't survive you unplugging the cable, ask what's on the other end of it.”

It is a two-minute test and it does not require a security audit to run. We run it on our own device in front of you.

05 · The hardware

What sits in the comms cabinet

Platform
GMKtec EVO X2
Processor
AMD Ryzen AI Max+ 395
Memory
128 GB unified
Inference
Local. No third-party LLM API
Network
Zero egress by default
Ownership
Hardware bought outright

This is the current demonstration device. Larger firms can be sized up; the architecture does not change.

06 · The walkthrough

Twenty minutes, and your IT team is invited

We send an architecture and security summary ahead of time so your IT team can read it before anyone talks about features. Then we demonstrate the device on synthetic data, and pull the cable while it is running so you can see what happens: nothing.

Book a walkthrough

LegalBox is a working demonstration device today. Demonstrations use synthetic data, never real client material.

LegalBox FAQ

What is LegalBox?

LegalBox is an on-premises AI appliance built by Zatersio for Australian law firms. It is a physical device that sits inside your own network and runs the language models locally on its own hardware. There is no public cloud service behind it and no third-party LLM API call, so confidential and privileged matter files stay on your premises.

Does LegalBox need an internet connection to work?

No. Answering runs entirely on the device, so you can pull the network cable out and it keeps working. That is the test we run during the walkthrough, and it is the same test we suggest you run on any AI vendor: if the product cannot survive you unplugging the cable, something on the other end of it is receiving your data.

What can lawyers actually do with it?

LegalBox ships with six modules: Matters for organising work, Ask for questions answered with citations across a matter's documents, Compare for clause-by-clause redlines exportable as a .docx, Draft for tracked-change drafting a lawyer accepts or rejects, Playbooks for encoding the firm's own positions, and Admin for access control. All of it runs on the device.

Where do the answers come from?

LegalBox retrieves from your firm's own authority sources: the precedents, policies and matter material you point it at. Answers cite what they were drawn from so a lawyer can check the source rather than trust the output. The demonstration device runs on synthetic data, not real client material.

How can our IT team verify that nothing leaves the building?

The device ships with zero network egress as the default and is designed to be inspected rather than taken on trust. Your IT team gets an architecture and security summary before the walkthrough, can put the device on a monitored segment, watch the traffic themselves, and physically disconnect it mid-session while it keeps answering.

Is LegalBox a subscription?

No. The hardware is purchased outright and stays yours: there is no per-seat cloud fee and no metered token bill, because no tokens are being sent anywhere. Each firm's deployment is scoped and quoted after the walkthrough, once we know your data sources and how many people need access.

What hardware does LegalBox run on?

The current demonstration device is a GMKtec EVO X2 with an AMD Ryzen AI Max+ 395 processor and 128 GB of unified memory: enough to run capable models locally at usable speed in a unit that fits in a comms cabinet. Larger firms can be sized up; the architecture does not change.

Does LegalBox satisfy Australian regulatory guidance on AI?

Australian legal regulators have cautioned firms against uploading confidential client material to public AI platforms. LegalBox is built so that material never leaves your premises in the first place, which removes that specific exposure. Your professional and privacy obligations remain yours: we are engineers, not your legal or compliance advisers.

Lakitha 'Lucky' Sahan, founder and lead engineer of Zatersio

Lakitha “Lucky” Sahan

Founder & Lead Engineer — leads the Zatersio engineering team

LinkedIn