On 15 July the Prime Minister made "what did you train on" a national question. He was answering it for novelists. Nobody in Canberra is answering it for you.
The short answer. No, VibeCFO does not train AI on your accounting data. EVA runs on frontier models that do not train on customer inputs, and under our Xero partnership we are contractually prohibited from training on client data. She does not learn from your ledger. She writes a SQL query against your own dedicated database and returns the figure that is already there.
Anthony Albanese stood up at the University of Sydney and put artificial intelligence in the same bracket as civil aviation in the 1920s and genetics in the 1990s. A new Office of AI now sits inside the Department of the Prime Minister and Cabinet, effective immediately. It will coordinate a set of mandatory Australian AI Standards, which the government says it will legislate early next year. Data centres get real obligations: underwrite your own power, do not push up household bills, keep water use down.
Significant. Also, if you run a business or a firm, strangely quiet.
The sharpest language in the speech was about training data.
The government has again ruled out a text and data mining exception. That is the carve-out that would let AI companies scrape Australian books, music, art and journalism to build models without permission or payment. Albanese's framing was blunt. Unless the creator keeps control over how the work is used and what it is worth, anything less is theft.
Be precise about the scope, because it matters.
He was talking about creative work. Novelists, musicians, artists, journalists. He was not talking about your ledger, your client records or your customer list. Those sit under privacy law and existing regulators, and they were not in the speech. There are no fines here either. There is no legislation yet, so there is nothing yet to be fined under.
Strip the sentence back and here is what is underneath it.
Nobody should be feeding your work into a model without you controlling whether that happens and what it is worth.
Read that again with a general ledger in place of a manuscript. It reads identically. The law treats the two differently. Your gut does not.
And once you ask that question about business data rather than creative work, the answers in the market get uncomfortable.
From 17 August 2026, Atlassian begins using customer content from Jira and Confluence to train its AI products. Whether you can opt out at all depends on which subscription tier you are on. It reverses a position the company's own documentation still states, that customer data belongs to the customer. Figma has spent the past year defending a claim that it trained on millions of customer design files after years of assuring users it would not. Figma denies it and the matter has gone to arbitration.
Notice the shape of both. Nothing was stolen from outside. It was repurposed from inside, quietly, by default, by a vendor who had already been trusted with it.
That is the version of the theft question that lands on your desk. Not whether an AI lab can train on Tim Winton. Whether your finance software is doing something with your numbers that you did not agree to.
Panel (a) is the fear. Your ledger, your client records and your pricing go into a corpus, get retained, and turn up later in somebody else's answer.
Panel (b) is how VibeCFO is built. The difference is not a promise. It is an absence. There is no arrow.
Those are different things, and the difference is the whole architecture.
Every business gets its own database. Not a shared table with your name in a column. A dedicated database, hosted on AWS in Sydney for Australia and New Zealand, encrypted in transit and at rest. Your data is never pooled with anyone else's.
Every client has its own credentials. Inside a firm, each clientID has a unique database username and password, with row-level security underneath. There is no master key. Client 5 cannot surface in Client 10's answer, because there is no route by which it could. Isolation is enforced where the data lives, not asked of the model and hoped for.
She writes SQL, she does not invent numbers. Ask EVA a question and she composes a query against your database. The figure comes back from your rows, the same rows your accountant works from. Pin a visual and it becomes a stored SQL command, no longer touched by AI at all. This is why hallucinated numbers are not a category of risk here. She is not generating the number. She is fetching it.
Nothing trains on it. EVA runs on frontier models from Anthropic and OpenAI on AWS, and those models do not train on your data. Under the Xero partnership we are contractually prohibited from training on client data. Two independent reasons, one outcome. The same holds when an outside agent asks: The VibeCFO MCP queries your data in place, per call, and every call is logged.
So when EVA is sharp, in depth and accurate the way a good CFO is, know where that comes from. It comes from purpose-built ETL that pulls row-level data. Xero invoice lines. Timesheets row by row. Stock on hand. Across financial and operational systems, because operational data is what explains financial performance. Every figure traceable to the source row.
Not from having eaten anybody's books. A model that has swallowed a million ledgers still cannot tell you why your gross margin moved last month. One that can read all of yours, properly, can.
None of that is a compliance posture we adopted last Wednesday because the Prime Minister gave a speech. It is how the thing has been built since 2018. The full detail is on our security page.
Read the announcement closely and one group is almost absent. The ordinary Australian business owner.
No productivity targets for SMEs. No adoption incentives. Nothing practical for the cafe, the tradie, the accounting firm or the manufacturer trying to work out what AI does for their numbers on a Tuesday afternoon.
That is a description of the gap, not a complaint. Governments set frameworks. They do not run your reporting, chase your aged debtors, or tell you whether this month's margins can carry a new hire.
No. The frontier models EVA runs on do not train on customer data, and under the VibeCFO Xero partnership we are contractually prohibited from training on client data. EVA does not learn from your ledger. She writes a SQL query against your own dedicated database and returns the figure that is already there.
On AWS in Sydney for Australian and New Zealand customers, with dedicated regional infrastructure for the UK, US and Canada. Data is encrypted in transit with TLS 1.3 and at rest with AES-256, in a database dedicated to your business alone. How we protect your data.
No. Every business has a dedicated database and every clientID has its own database credentials plus row-level security. There is no master key, so there is no route by which one client's data could surface in another's answer. Isolation is enforced at the data layer, not asked of the model.
EVA does not generate numbers. She writes SQL against your database and returns what is in your rows, the same rows your accountant works from. Pin a visual and it becomes a stored SQL command that AI no longer touches. If a figure looks wrong, it is the ETL refresh, not a hallucination. See how EVA works.
The Australian Government established an Office of AI within the Department of the Prime Minister and Cabinet and ruled out a text and data mining exception, meaning AI companies get no copyright carve-out to train on Australian books, music, art or journalism. The announcement covered creative work, not business or accounting data, and introduced no fines.
That is the real question, and it is why architecture beats policy. A promise can be updated on a Tuesday. A missing arrow cannot. VibeCFO has no training path to switch on, because the contributory arrow is absent by construction rather than by policy promise.
Connected data is becoming normal. Secure, trusted CFO action is not.
The Prime Minister asked what the country's creative work is being trained on. Ask the smaller, sharper version about your own numbers, and ask it of every vendor holding them. Ask where the database sits. Ask who holds the credentials. Ask whether the answer is generated or queried. Ask what happens on the day the terms of service change.
Then check the diagram again and count the arrows.
Out of the box. Governed and secure. No dashboard to decode, and no AI to train.
VibeCFO is an Australian agentic AI CFO for small businesses and accounting firms, protecting accounting-firm data since 2018. Meet EVA at vibecfo.ai, or read exactly how your data is handled at vibecfo.ai/security.
Connect your data, ask EVA a question, and watch every figure trace back to your own dedicated database. No pooled data, no training, no arrow.
Start your 14 daysUnderstanding the balance between using AI capabilities effectively while maintaining data privacy and security.
Connecting data is not the same as understanding a business. The real challenge is not access. It is interpretation.
