Confidentiality

Where your matter text goes when you hit send

By Michael Nadalin, Founder, Market Lead · 9 September 2026 · 7 min read

You paste four paragraphs of a client's affidavit into a chat window to get a plain English summary. The reply comes back in nine seconds. In those nine seconds that text has been packaged into a request, sent across a network, held in memory on a machine you cannot name, possibly written to a log, possibly retained for a fixed window, and possibly read by a human doing abuse review. None of that is visible to you. The interface shows a typing indicator.

Most discussion of AI and confidentiality stops at the wrong question: will my data be used for training. That is the easy one, and vendors now answer it clearly because it costs them nothing to say no. The harder questions are where the text physically sits, who can lawfully compel access to it, and how long it survives after your session closes. Those are the ones that matter if a client asks you to certify that their commercially sensitive material never left the country.

This is a walk through each hop between your keyboard and the model, what is actually happening at each one, and the specific questions that force a vendor into a straight answer instead of a marketing page.

The nine seconds, hop by hop

A single prompt is not one journey. It is a chain, and each link is a different company, a different jurisdiction and a different retention policy. Firms tend to assess the one they have a contract with and assume the rest inherit its terms. They do not.

The hop that surprises people is the gap between the tool you bought and the model it calls. A great many Australian legal AI products are a well designed interface sitting on top of a model hosted by someone else. Your agreement is with the interface. The model provider has its own terms, its own regions and its own retention window, and you may never have read them.

One prompt, seven stops
Stop 1
Your browser or desktop app, where a draft may sit in local storage before you press send
Stop 2
A content delivery network edge node, often the closest city, which may or may not be Sydney
Stop 3
The vendor's application servers, which hold your account, matter names and prompt history
Stop 4
The vendor's database and backups, usually a different region again from the application
Stop 5
The model provider's API, frequently a different company under different terms
Stop 6
Abuse and safety logging, where a copy is retained for a fixed window regardless of your settings
Stop 7
Analytics, error tracking and support tooling, which quietly captures prompt fragments in stack traces
Each stop is a separate policy question. A single answer about the first one tells you very little about the rest.

What data residency actually means, and what it does not

Residency is a statement about where bytes are stored at rest. It is not a statement about who can reach them. A dataset can sit in a Sydney data centre and still be accessible to an engineer in another country holding valid credentials, and still be subject to a foreign production order if the parent company is incorporated overseas. Storage location and legal control are separate axes, and vendors routinely answer the first when you asked the second.

Under the Australian Privacy Principles, sending personal information to an overseas recipient is a disclosure, and APP 8 makes you responsible for reasonable steps to ensure that recipient handles it consistently with the APPs. That obligation does not disappear because a subcontractor made the actual transfer. If your practice management system calls a US model provider, your firm is the one that disclosed.

Then there is the professional obligation, which sits above the privacy one and is stricter. The duty of confidentiality in the solicitors' conduct rules is not a data protection standard with a reasonableness carve out. It is a duty to the client, and it is not satisfied by pointing at a vendor's compliance page.

What firms assume against what holds
Hosted in Australia
Storage at rest is local. Support access, backups, failover regions and logging may all be elsewhere.
We do not train on your data
Training is one use. Retention, human review and subprocessor access are separate and usually unaddressed.
Enterprise grade encryption
Encryption protects against interception and theft, not against a lawful order served on the party holding the keys.
SOC 2 certified
An audit of stated controls against stated scope. It says nothing about where data lives or who can compel it.
Zero data retention
Often means the model provider's retention only, and often excludes an abuse logging window measured in weeks.
Our vendor is compliant
Their subprocessors are the exposure. Ask for the list and check when it was last updated.
Every row on the left is a real thing a vendor has said. None of them answer the question on the right.

Rank your matters before you rank your tools

Not all matter text carries the same risk, and treating it as though it does produces a policy nobody follows. A blanket ban on AI tools is broken within a fortnight by someone under deadline pressure with a laptop and a personal account. That is worse than a graded rule, because now you have no visibility at all.

The workable approach is to sort your work by what an unwanted disclosure would actually cost the client, then match the tier to the tool. Most firms find the top tier is smaller than they feared, which frees up the rest for sensible use.

Matter sensitivity tiers, highest exposure first
Live criminal, family violence and child protection matters
Disclosure risks physical safety and the material is often subject to statutory publication restrictions. Nothing identifiable leaves the firm.
No external tool
Unannounced transactions and contested commercial matters
Market sensitive or strategically valuable to a counterparty. The value of a leak to someone else is high and immediate.
Sovereign only
Privileged advice and litigation strategy
Privilege is a client right you hold on trust. Even a low probability of third party access is a poor trade for a faster draft.
Sovereign only
Identified client correspondence and pleadings
Confidential but not usually catastrophic. Reasonable on a tool with documented Australian residency and contractual retention limits.
Contracted vendor
De-identified fact patterns and hypotheticals
Strip names, dates, amounts and the distinctive facts, and you are asking a legal question rather than disclosing a matter.
Broad use
Pure research with no client facts
Legislation, case law and procedure carry no client information at all. This is the bulk of most research time.
Broad use
Sort the work first. The tool question becomes much easier once you know which tier you are answering it for.

The questions that get a straight answer

Vendor security pages are written to survive procurement, not to answer a solicitor. The way through is to ask questions that cannot be satisfied with an adjective. Put them in writing, ask for the answers in writing, and keep the reply on file. If a supplier will not commit in writing to something they say on a call, that is your answer.

Pay attention to what happens when you cancel. Deletion on termination is frequently promised for the application database and silently excluded for backups, which may roll for months. A firm that closes an account believing its matter text is gone, when a restorable copy persists offshore, has a problem it does not know about.

Six questions to send before you sign
1
Name every region
In which specific data centre regions is our content stored at rest, processed in memory, backed up, and replicated for failover. List all four, not just the first.
2
List every subprocessor
Which third parties receive our content, including model providers, hosting, error tracking, analytics and support tooling. When was this list last changed and how are we notified of additions.
3
State the retention clock
How long is prompt and output content retained, in days, in each system including abuse and safety logs. Is that window contractual or a policy you can change without notice.
4
Identify who can read it
Which staff roles can access customer content, from which countries, under what approval, and is that access logged and available to us on request.
5
Describe the compulsion path
Which corporate entity holds the data, in which jurisdiction, and what is your process and notification commitment if served with a foreign order for customer content.
6
Define deletion
On termination, what is deleted and when, and specifically when do backups and log copies expire. Will you certify deletion in writing.
Send them as a numbered list and ask for a numbered reply. Vague answers to specific questions are themselves informative.

The de-identification trap

The common workaround is to strip the names and proceed. It is a reasonable instinct and it works for a lot of routine work, but it fails in a specific and predictable way: legal facts are identifying even without names. A dispute over a particular development site in a particular council area, with a particular date and a particular sum, identifies the parties to anyone who follows that sector, and often to anyone who reads the local paper.

The test is not whether you removed the proper nouns. It is whether an informed reader in that industry could work out who this is. For most commercial and property matters the answer is yes, from three or four facts, well before you have said anything you would recognise as a name.

There is also a workflow problem. De-identification only holds if it happens every time under time pressure, and it does not. The paste that goes wrong is the one at 6pm on a Friday before a Monday hearing. Any policy that depends on manual redaction discipline should be treated as a policy that will be breached, and designed around accordingly.

What is left in a typical 'de-identified' paste
Names and direct identifiers removed
Dates, amounts and locations left intact
Distinctive procedural history left intact
Industry and transaction specifics left intact
Names and direct identifiers removed: 25
Dates, amounts and locations left intact: 30
Distinctive procedural history left intact: 20
Industry and transaction specifics left intact: 25
An illustrative breakdown of what survives a quick redaction pass. The removed names are rarely the identifying part.

A position you can put in an engagement letter

The end point of this is not a ban and not a shrug. It is a written position your firm can state to a client without hedging: these categories of work never leave Australian infrastructure, these categories may be processed by named suppliers under contractual retention limits, and here is the list of those suppliers. A client who asks is entitled to that answer, and increasingly they will ask, particularly government and institutional clients who have their own residency obligations flowing down to you.

Practically, that means picking tools where residency is a documented architectural fact rather than a support ticket answer, keeping the subprocessor list current because it changes without fanfare, and reviewing the position when a vendor changes model providers, which they do. It also means logging which tool was used on which matter, so that if a question arises two years later you can answer it from records rather than memory.

Legal Brain runs on Australian infrastructure for exactly this reason, and publishes where matter text sits and how long it is kept, but the point holds whichever tool you choose: if you cannot describe the path your client's text takes, in writing, you are not in a position to give the assurance your conduct rules already require of you.

Frequently asked questions

Does putting client information into an AI tool breach legal professional privilege?

Privilege attaches to confidential communications made for the dominant purpose of legal advice or litigation, so the risk is loss of the confidentiality that underpins it rather than an automatic breach. Disclosure to a supplier who processes material under confidentiality obligations, purely to assist you in giving advice, is generally treated differently from disclosure to the world. The exposure grows where the supplier retains the content, where staff can read it, and where a foreign authority could compel production. Document the arrangement so you can describe it if it is ever challenged.

Is it enough that my AI vendor stores data in Sydney?

No. Storage at rest in Australia is one control among several. Ask separately about where processing happens, where backups and failover copies go, which countries support staff access from, and which corporate entity holds the data, because that entity's jurisdiction determines who can lawfully compel it. A Sydney region operated by an overseas parent still sits inside that parent's legal system.

What do the Australian Privacy Principles require before sending client data overseas?

APP 8 requires you to take reasonable steps to ensure an overseas recipient does not breach the APPs, and under section 16C of the Privacy Act you can remain accountable for their acts. In practice that means contractual commitments about handling, retention and onward disclosure, plus a record of the assessment you made. Note that your professional duty of confidentiality is separate and stricter, and satisfying APP 8 does not discharge it.

Can I just remove the names before pasting a matter into an AI tool?

It helps but it is not sufficient on its own. Dates, amounts, locations, industry detail and procedural history frequently identify the parties to anyone familiar with the sector, even with every proper noun removed. Ask whether an informed reader could work out who this is, and treat manual redaction as something that will eventually be skipped under deadline pressure.

Check your understanding

Two quick questions

No score is stored. Pick an answer to see why it is right.

  1. 1A vendor tells you your firm's data is stored in an Australian data centre. What does that statement leave unanswered?

  2. 2Which of these is the strongest reason not to rely on manually removing names before pasting matter text into a general AI tool?

Research Australian law without handing over client data

Legal Brain searches Australian legislation and case law, shows you the source behind every answer, and anonymises client-identifying detail before anything reaches a model.

Request early access