Frequently asked questions

Storage and processing are different questions, so here is each step and where it happens.

Everything we send for AI processing passes through an Australian redaction step first, where Tax File Numbers, Medicare, passport and licence numbers are found and stripped. Redaction is best-effort: in rare cases a file cannot be checked and is processed as uploaded.

Everything we run ourselves is in Australia — storage, identifier removal, recognition, and the container the workpaper is actually built in. What leaves is what the AI model reads: the rest of the document, plus a case summary including the taxpayer's name, goes to the AI provider handling that step. That is a cross-border disclosure under APP 8 and the Privacy Policy says so — naming every provider and the country it processes in — rather than rounding it to "your data stays in Australia".

Payments, email and analytics have their own providers and locations, all listed in the sub-processor table.

No. We use commercial AI services under their no-training terms and do not opt in to sharing client content for model training. We do not use it to train our own models or sell personal information.

Retention depends on the provider and the feature being used. Our main AI services are Fireworks AI for analysis and drafting, and Anthropic for review. The Privacy Policy lists all providers, including alternatives.

Fireworks' open-model inference uses zero data retention (ZDR) by default: prompt and response content is not written to persistent storage. A short-lived memory cache and service metadata, such as token counts, are separate from that commitment. We do not opt in to content logging or use its stored-conversation feature.

Anthropic offers ZDR for eligible services under an agreed arrangement; it does not cover every feature. Its standard API retention is up to 30 days, subject to its published exceptions. Files uploaded for review are held until deleted, and code-execution copies can remain for up to 30 days. We request deletion of input files when the run ends; deletion failures are logged for follow-up. We therefore do not promise that all Anthropic processing is ZDR.

Your firm's source documents and client records are stored separately in Australia under section 8 of the Privacy Policy. An AI provider's ZDR policy does not delete those records.

Some of it, and we can be exact about which.

Stripped, in Australia, before anything is sent. Tax File Numbers, Medicare numbers, passport numbers and licence numbers. Every document and every instruction passes through an automated redaction step on our own Australian service first, on every path — not just the main one. It is best-effort: in rare cases a file cannot be checked and is processed as uploaded.

Also masked in extracted text: bank account numbers and BSBs become codes; phone numbers and email addresses become substitutes. Email addresses are also removed from the source file. Bank details and phone numbers may still appear in a redacted source file sent for AI review. Recognised crypto identifiers are masked too, with the detection limits described in section 2 of the Privacy Policy.

Still present: names, addresses, dates of birth and financial figures. This is data masking and redaction, not full anonymisation: the material sent to the AI can still identify a client.

What contains the rest of the exposure:

• Purpose. Information is sent to provide the AI task your firm requests, such as preparing a workpaper or drafting a client reply.
• No training. We use providers' no-training terms and do not opt in to sharing client content for model training.
• Separate TFN storage. Your firm can hold a TFN on the Australian client record for tax preparation and lodgement; it is removed from AI inputs.
• Named providers. Every processor is listed with its location in the Privacy Policy, so a client's question has a specific answer rather than a reassuring one.

One word we still avoid: de-identified. Stripping identifiers is not de-identification. Because the name and address remain, what we send is still personal information under the Privacy Act 1988 (Cth), and APP 8 applies in full. Partial masking is sometimes marketed as de-identification; if you repeat that word to a client, it should be one that holds up.

They are removed before the document leaves Australia. A payment summary or ATO notice that prints a TFN goes through our Australian redaction service first; the number is detected and stripped, and the document continues on without it.

Your firm can store a TFN on the client record in Australia so your accountant can check it for tax preparation and lodgement. It is masked on screen until deliberately revealed. BeforeMay does not currently lodge returns to the ATO.

The recognised text and AI extraction do not keep a separate TFN value. Where one is detected, the extraction keeps only a note that it was present. The client record and your firm's original document are separate from those AI inputs.

Your own copy of the document is untouched — the removal applies to what we send onward, not to what your firm holds. Please don't type TFNs into free-text notes, where they would be stored as text rather than sitting in a document the redaction step reads.

A page we cannot read is not sent overseas to be read. If the redaction check cannot complete for a file, the workpaper is still built: a photo with no readable text is left out and marked "not in the paper"; other files are processed as uploaded.

TPB(GS) 55/2026 notes at paragraph 26 that where client information involves TFNs, additional obligations under the Privacy (Tax File Number) Rule 2015 apply to you.

The Tax Practitioners Board issued TPB(GS) 55/2026 on 22 July 2026. Two parts of it bear on using any AI tool, including ours.

You need your client's permission. Code item 6 says you must not disclose information relating to a client's affairs to a third party without the client's permission unless you have a legal duty to. The guidance treats any entity other than you and your client as a third party, and says that entering client information into AI tools can be such a disclosure, "depending on how these tools are configured and used" (para 23). We are a third party, and on our configuration your clients' information is disclosed to us and our sub-processors.

Your review is your own. You remain responsible for the accuracy of the work, must apply your own professional judgement, and must not treat AI output as a substitute for your own analysis.

Probably not. Permission is required; a separate AI-specific consent per client generally is not.

Paragraph 24 of the guidance lists the forms permission can take: a signed engagement letter, a signed consent, another communication such as a relevant "fact find" and consent, or — the one most firms already have — a general authority consenting to disclosure to third parties. In the Board's words, "a general authority consenting to disclosure to third parties may also be acceptable."

If your standard engagement letter carries such an authority, that may be enough. Check your own template against your PI insurer's and your association's guidance. If you would rather be explicit, we publish a clause you can adapt — ask us for the disclosure pack.

The guidance also recommends (this part is best practice, not a requirement) telling the client to whom and where the disclosure is made, where data is stored, and whether AI is used. The table at the top of this section is written so you can answer all four.

The permission sits between you and your client. We are not a party to that engagement and cannot obtain consent on your behalf. What we can do — and what this page is — is state accurately what happens to the data, so that what you tell your client is right.

We ask an authorised person at your firm to confirm they have read this once, during setup. Not per client.

Yes, and this is the part most firms underestimate.

Paragraph 16 of the guidance says you should verify and review AI-generated content throughout the workflow, and that each of these steps should be documented to support ss 30 and 40 of the Tax Agent Services (Code of Professional Conduct) Determination 2024 — record-keeping, and a documented system of quality management.

The platform keeps that record as a by-product of how the work gets done:

• every point the AI raised, and what you did with each one
• who changed what, and when, actor-stamped
• every version of the workpaper, with restore
• the reasoning behind each flagged item, so you can contest it

The review is still yours to perform. What you don't have to do is assemble the evidence that you performed it.

Almost all of it now is.

Storage, identifier removal, text recognition and building the workpaper all run on our own infrastructure in Sydney. They used to run offshore; they were moved deliberately, and the sub-processor table is the record of it rather than a promise.

The one step that still crosses the border is the AI model itself. The workpaper is built inside our own Sydney container; what that container sends out is the model call — the redacted document text the model reads and the answers it writes back — to whichever provider handles that step, in the country the Privacy Policy names for it.

Running the model in an Australian region is possible for some providers and costs materially more per workpaper, so we do not do it today. If onshore-only processing is a hard requirement for your practice, tell us — we track that, and enough of it changes the decision.

The Privacy Policy is the authority. Section 5 covers AI processing and the human-review boundary, section 6 lists every sub-processor with its location, section 7 covers cross-border disclosure, section 8 covers retention, and section 13 is written specifically for registered tax practitioners.

No. We prepare the workpaper; you review every figure and lodge through your existing software. Nothing is sent to the ATO automatically, ever.

A real .xlsx workbook with live formulas — not a PDF, not a locked file, not a macro-enabled workbook. File it in your document management like any workpaper you'd prepare by hand.

You can get a general AI assistant to talk about a return. Four things it will not do, and each of them is the part that takes the time.

• Read the client's actual paperwork. Not a pasted excerpt — the whole bank statement, the photographed receipts, the trial balance. Recognition runs before analysis, so a scan is read rather than described.
• Tie every figure to where it came from. Each number in the workbook traces back to the page of the source document it was read off, so a bad read is visible in one click instead of argued about.
• Check itself before it hands you the file. The build runs its own validation and a blind review pass before you see the workbook, and what it cannot fix it flags rather than shipping quietly. A chat window has no notion of failing.
• Leave the record. The evidence that you reviewed it is assembled as a by-product of the work. A chat transcript is not that record.

There is also the part that is yours either way: whatever you paste into ChatGPT or Copilot, you have made that cross-border disclosure yourself, to whichever provider sits behind it. Here, identifiers are stripped in Australia first and every provider is named with the country it processes in.

It gets flagged, not lodged. Low-confidence extractions go to Needs review and are never auto-approved. Every figure traces back to its source document, so a bad read is visible in one click. You sign off before anything leaves the platform.

They are available by invitation while we work on them, and we say so rather than burying it in fine print. Individual and sole-trader returns are where the output is strongest today; company and trust workpapers are earlier and need more of your review. If you want access, ask.

Receipts, invoices, bank and brokerage statements, trial balances and general ledgers, as PDF or image, up to 25MB each. Scanned documents are read by our recognition pipeline before analysis.

The workbook is ours today. Firm-template output is on the roadmap, not shipped. If your practice has a house template you cannot deviate from, tell us before you sign up rather than after.

Nobody browses them. Documents are isolated per firm by database row-level security keyed to firm membership, storage is private with short-lived signed URLs issued only after we verify membership, and access is logged. Support access to diagnose a problem happens with your knowledge.

A portal invitation is bound to one email address and has to be explicitly accepted — visiting a link is not consent, and there is no anonymous upload path. Clients upload from their own signed-in account.

Yes. Deleting moves a case to trash for 30 days, then it is removed. Retention for everything else is in section 8 of the Privacy Policy.

Yes, and we recommend turning it on during setup.

New practices start with A$99 free, no card, so you can judge the output on your own files. After that you pay from one balance in Australian dollars. There is no monthly fee: an individual return costs A$12, A$30 or A$45 depending on how much it carries. Company and trust papers are in beta and priced on their own. BeforeMay Verify, in early access, is paid from the same balance; its prices are listed in Verify.

If you'd rather commit monthly, a plan of A$299, A$999 or A$1,999 a month adds A$328.90, A$1,198.80 or A$2,598.70 to your balance — 10%, 20% or 30% more than you pay, so the bigger the plan, the better the deal. Unused plan balance stays for 12 months; top-ups never expire. See pricing for the current rates. All prices include GST.

Turn on auto top-up and the balance refills when it runs low, up to a monthly limit you set. If auto top-up is off, new work waits until you top up in Plan & balance.

No lock-in contract. Your data is yours; the deliverable is a standard Excel file you already have a copy of.