A secure AI inside your firm. Only what is needed goes out.

In brief

ITHA sets up a secure AI inside your firm, connected to your documents and apps, with one login per person, a written AI use policy and training for your team. Your documents stay with you: only the extract a task needs goes out, to an AI provider named in your written agreement. No agent deletes or pays, and none sends anything without your team.

The three questions asked first

Where does our data go?Only the useful extract goes out, to the provider named in your agreement.
Who sees what?Each person sees only what you give them access to.
What if an agent gets it wrong?It cannot delete or pay, and sends nothing without your team.

Your environment · your firm’s accounts

Your client’s request · $20,700 ex GST

Request received by email on 07/10/2026: team restructure, 85 staff, three shifts.
12 consulting days over six weeks, from 04/11/2026, as requested.
Rate from your 2026 rate card, page 2: $1,725 ex GST a day.

Proposal total, ex GST:$20,700

Agent · Your firm

Use a personal AI accountBlocked
Delete a documentBlocked
Send without your team, or payBlocked
ITHA access without your approvalBlocked

Example · fictional data.

Your team already uses AI. It does not know your firm.

Without a framework, this is what it looks like.

Everyone on a personal account.
Each person picks an AI and a subscription. Nobody knows who uses it, or for what.
Client data going out.
On a personal account, what your team pastes in can be used to train the provider’s models, and leaves your control.
Generic answers.
The AI knows nothing of your clients, your rates or your documents.
Nothing really automated.
Copy, paste into a chat, then copy the result back into your software. An agent does that work inside your software, with no copying and pasting.
Few regular users.
A few keen people use it every day. The others, rarely.

ITHA sets up a secure AI, connected to your documents and apps.

First the AI, with its rules. Then, if you want, agents connect to it for the repetitive work: your team reviews and decides.First the AI and its rules, then, if you want, agents: your team reviews and decides.

Is your AI set-up really secure?

Business accounts for everyone
Training excluded, in writing
The AI provider known and named
One login per person
A written AI use policy
An answer ready when a client asks

Six points to check, with ITHA or without.

What is included

Choose the AI, or keep yours.
A major provider’s AI under an enterprise licence, or an open-weight model on your own servers.
Connect it to your documents and apps.
It answers from your files, your rates and your contracts.
Access rules.
One login per person. Each person sees only what you give them access to.
An AI use policy.
Written with you: what your team gives the AI, and what it keeps.
Training for your team.
So everyone uses it, within the rules.

A confidentiality agreement before work starts
No training on your data, by contract
Nothing is sent without your team

The journey of one of your files.

The OAIC recommends that organisations keep personal information out of publicly available generative AI.

OAIC

recommends that organisations keep personal information out of publicly available generative AI.

Guidance on commercially available AI products, 21/10/2024

In its words: “do not enter personal information, and particularly sensitive information, into publicly available generative AI tools”. OAIC, guidance on privacy and the use of commercially available AI products, 21/10/2024.

Here is one of your files, from arrival to sending. At each step, where it is and who can read it.

New email

07/10/2026

Your client · Could you send us a proposal?

Your environment

The request arrives and is read where it is

Your documents stay with you.

Your workspace reads the request. Rates and dates are checked in code, with no AI.

Your environment · your accountsOne login per person

Set up in your firm’s accounts, under your access rules; kept separate from other clients.

Read in your environment

Your client’s request · $20,700 ex GST

Request received by email on 07/10/2026: team restructure, 85 staff, three shifts.
12 consulting days over six weeks, from 04/11/2026, as requested.
Rate from your 2026 rate card, page 2: $1,725 ex GST a day.

Proposal total, ex GST:$20,700

Memory, history and working copies stay in your environment.

At this step, nothing leaves your environment.

Read in your environmentYour client’s request · $20,700 ex GST

Example · fictional data

Each installation is kept separate.

An AI agent is a program that reads your documents, runs the checks and prepares the work, then hands it to your team.

Large model, enterprise licence

The useful extract goes out

To draft the proposal: the figures it needs, never the whole file.

The useful extract goes out, encrypted, to a large model under an enterprise licence.The useful extract goes out, encrypted.

For nothing to go out, everything can run on your servers. See the three ways to set up

What goes out

Encrypted in transit

team restructure 85 staff three shifts
12 consulting days six weeks 04/11/2026
$1,725 ex GST a day

$20,700

Under an enterprise licence, never a personal account: no training on your data, by contract.

Large model, enterprise licence

Provider named in your written agreement.

Your team

The send waits

The draft comes back to your environment. Nothing is sent without your team.

Proposal prepared · Your client

Read the proposal

Thank you for your request. We propose 12 consulting days over six weeks, from 04/11/2026, as you requested.

At $1,725 ex GST a day, the proposal comes to $20,700 ex GST. Sources: your request, your closest past engagement, your 2026 rate card.

Decision pending

Send the proposal
Edit first

Your team reviews and decides, file by file.

Every action is logged: who, what, when. See the control panel

Your client’s request, from arrival to sending. Example · fictional data.

In your accounts or on your servers, the same protections.

Where your agents work is set in the feasibility check, from your software, your data and your security needs.

The choice is made in the feasibility check, from your data and your security needs.

On your servers

Where the agents work: On your own server
What goes out: The useful extract, to a large model under an enterprise licence
Who maintains it: ITHA, with the access you grant
The right choice if: Your server can carry the load, checked in the feasibility check

In your firm’s accountsThe example’s

Where the agents work: In your firm’s accounts, under your access rules
What goes out: The useful extract, to a large model under an enterprise licence
Who maintains it: ITHA
The right choice if: You would rather not run a server of your own

On your servers, nothing goes out

Where the agents work: On a dedicated machine, on your own servers
What goes out: Nothingan open-weight model also writes
Who maintains it: ITHA, with the access you grant
The right choice if: Nothing must go out, after a trial on your documents. Writing quality drops and hardware costs more.

What the OAIC says

“Deploying a system locally or on premises is likely to be more privacy-preserving as it limits the risks of third party access to the data.” OAIC, guidance on privacy and the use of commercially available AI products, 21/10/2024.

Each person sees only what you give them access to.

The agent shows each person only what they are allowed to see.

6 logins, one per person. ITHA: access to your data only with your approval.

What does the new hire’s contract say?

Management account

Employment contract · Full-time, ongoing

Your firm and the new hire · 08/10/2026

Clause 1. Position. Project Consultant.
Clause 4. Pay. Salary and superannuation, as set out in Schedule 1.
Clause 5. Probation. Six months.

A consultant’s account

Outside your access

Employment contracts: management only.

Example · fictional data.

What no setting reopens.

Security does not rely on the AI behaving well. Four actions stay closed to every agent, and each block is tested before go-live.

Permissions · Your firm

Use a personal AI account, outside the agreementBlocked
Delete a documentBlocked
Send without your team, or payBlocked
Give ITHA access without your approvalBlocked

Tests on 10/09/2026, before go-live

Data flowsChecked
Access rights6 of 6
Blocked actions4 of 4
Backup restoreRestored

Setting locked.

Example · fictional data.

Everything you have just seen, in writing.

Each set-up has its own sheet: where data is, who has access, what is blocked, what was tested. Here is the example’s.One sheet per set-up, three commitments in your agreement.

The 4 blocked actions, tried on 10/09/2026: all blocked.

Example · fictional data

Security sheet · Your firm

Set up on 14/09/2026 · version 1

Environment: Set up in your firm’s accounts, under your access rules; kept separate from other clients.; The journey
AI: Reading: your workspace, in your environment. Sums: in code, with no AI. Writing: a large model under an enterprise licence, provider named in your written agreement.; The journey
What goes out: The extract the task needs, encrypted in transit, under an enterprise licence with the provider. Training on this data: excluded by contract. Data retention: checked before work starts.; The journey The set-up
Retention: Action log; working copies; backups kept off the server.Periods set by you
Authorised people: 6 logins, one per person; employment contracts: management only; ITHA: access to your data only with your approval, logged.; Access
Blocked actions: Personal AI accounts, outside the agreement; deletion by an agent; payment by an agent, or sending without your team; ITHA access without your approval.; The blocks
Tests passed: 10/09/2026: data flows checked; access rights 6 of 6; read-only access checked; blocked actions 4 of 4; backup restored.; The blocks

Security sheet for your set-up. Example · fictional data.

The sheet, in technical terms

Encrypted in transit: the extract is encrypted on its way to the provider.
Backups off the server: copies kept outside your environment, restore tested before go-live.
Open-weight model: an AI model whose weights, the trained model itself, are published, so it can run entirely on your own servers.
Every action logged: a log of who did what, and when.
Provider named: the AI provider is named in your written agreement, with your written approval.

Three commitments, written into your agreement.

Three commitments, written into your agreement

Agreement extract · Your firm and ITHA

1.Each provider named.
The AI provider and every other party that handles your data are named in the agreement; none is added without your written approval.
2.Tested before go-live.
Each test and its result, written on your sheet.
3.If you leave.
Data, history, automations and documentation are handed back, ready to reinstall. ITHA then deletes its copies, backups included when they expire.

Example · fictional data.

Questions

What business owners ask first.

Security does not rely on the AI behaving well. It is built to hold when the AI gets it wrong.

Does the AI provider see our data, or train on it?Does the AI provider see our data?

It receives only the extract a task needs, encrypted, under an enterprise licence, never a personal account. It does not train on your data, under the enterprise terms, and its data retention is checked before work starts. Documents, memory and history stay with you.

Who sees what?

Each person has their own login and sees only what you give them access to. The AI answers each person within their rights: an employment contract stays with management.

What if an agent deletes a document or goes beyond its rights?What if an agent goes beyond its rights?

An agent has only the rights it is given, like a new employee: the documents opened for its task, within the rights of the person who asks, read-only by default. It cannot delete anything, and sending or paying waits for your team. Every action is logged, backups are kept off the server and restore is tested.

Does ITHA see our data?

ITHA maintains the set-up without reading your data. Any access to your data is requested, approved by you and logged. A confidentiality agreement is signed before work starts, from the feasibility check.

We already have an enterprise licence. What does ITHA add?Already on an enterprise licence: why ITHA?

The licence protects your data at the provider. It does not connect the AI to your documents or apps, set who sees what, or write your AI use policy. ITHA sets all of that up, with the licence you already have.

Does this work with the AI we already use?And the AI we already use?

Yes, if it is under an enterprise licence: ITHA keeps it and connects it to your documents. If not, the feasibility check compares the options on your own files.

Can our data go overseas?

Depending on the AI provider chosen, yes: the useful extract may be processed outside Australia, by the provider named in your written agreement. For nothing to go out, everything can run on your servers.

Is an AI on our own servers, with no outside provider, enough?Is an AI on our own servers enough?

For reading and searching, often. For writing, the feasibility check tries it on your documents beside a large model, and compares quality and cost.

AI on your own servers: an open-weight model, one whose weights (the trained model itself) are published, run on your own machine.

Will the AI you install go out of date?

No, the model can be changed, through your licence or on your servers. Open-weight models improve quickly; for writing, the gap is narrowing. ITHA tests each new model on your documents before adopting it: a trial, a comparison, a switch only if it does better.

What about the Privacy Act?

ITHA uses the Australian Privacy Principles of the Privacy Act 1988 as its reference. A written agreement sets out what is handled, how and by whom, and names the AI provider, with your written approval.

What happens if there is a data breach?

ITHA tells you as soon as it knows, with what is affected and what is being done. Each installation is kept separate.

What if ITHA stops trading?

Everything is documented and backed up off the server, so another provider can take over. Handover terms are written into the contract.

How do we take everything back?

Data, history, automations and documentation are handed over, ready to reinstall, within the time set in the contract. ITHA then deletes its copies, backups included when they expire.

A first conversation about your AI and your data.

Thirty minutes to look at the AI you use, where your documents are, and who should see them.

Thirty free minutes, at a time that suits you. You leave with a reasoned view.

Length: 30 minutes, at a time that suits you
Cost: Free, no documents needed
Outcome: A reasoned view: automate, or not

Email us

Mélanie Revellat: melanie.revellat@itha.io
Romain Danielant: romain.danielant@itha.io

Then, if you decide: the feasibility check, which sets where your agents work, then a fixed-price proposal. Set-up time depends on your security needs and on the data connected.