Held in Sydney, kept from AI, never used to train

In Sydney, Australia. The servers, the database, the encrypted store for Xero tokens and the logs all run in the AWS Sydney region. Data is encrypted in transit and at rest. The models the assistant uses also run in Australia, on Amazon Bedrock, so an explanation request does not leave the country either.

Two services sit outside that. Payments are handled by Stripe, which holds your billing name, email and card details and may process them outside Australia; it never receives your accounting or payroll data. This public website is served by Netlify, which counts page views from its server logs outside Australia. No script, no cookie, no customer data.

Tax file numbers, bank account numbers, salary figures, employee names and raw organisation identifiers never reach any AI provider.

Every figure Toby states is calculated by our own software, which contains no AI at all. AI is used afterwards, to explain a result or a provision in plain language. Before anything reaches a model, it is stripped back to what the explanation needs: an account code, a label, a figure, a section reference. Where an explanation has to refer back to a specific finding, the identifier is replaced with a scrambled code unique to your organisation. The model cannot reverse it and it means nothing outside your account.

Toby works in two layers. The calculation layer is our own software, contains no AI, and produces every figure. The explanation layer uses AI to put a result into plain language. It never produces a figure and never sees personal information; when an explanation mentions a figure, it is the one the calculation layer produced.

The rates, thresholds and dates the calculation layer applies are held in versioned schedules, not written into the software. Each schedule names its source, such as the ATO, a state revenue office or the FairWork Commission, and the dates it applies from and to. A person, acting as our Compliance Lead, checks each schedule against its source and attests it. The attestation records who attested and when, with a fingerprint of the schedule, and a changed schedule has to be attested again before Toby will use it.

The Accuracy page lists the worked examples published by the ATO and state revenue offices that Toby reproduces, the schedule behind each and when it was attested. How it works covers the method in more detail.

Nothing you send to Toby is used to train a model, ours or anyone else's. The assistant uses models we call, not models we train, and the provider's terms do not permit training on requests. Your ledger, your payroll, your questions and your answers are yours. The only record Toby keeps of them is your own audit trail.

Toby can also remember, per organisation and without personal information, which checks were useful, so the assistant picks up where you left off. An owner can switch it off, which deletes the record.

Sign-in is through a managed identity provider; Toby stores no passwords. Each organisation has its own account, and every read and write is scoped to it. Within that account an owner can invite and remove members, connect and disconnect software, manage billing and switch the de-identified record on or off; a member can run the tools and read the results. Scheduled audits are recorded against the account, with the schedule version and source each answer carries.

TobyAI staff do not read customer data in the ordinary course. Operational logs record hashed identifiers, not the values behind them. Where support needs to look at a specific record, it is with your permission.

A Xero connection is read-only, authorised by you in Xero, and the token is held encrypted in Sydney. Toby reads the accounting and payroll data the checks need and writes nothing back to your file. You can revoke the connection in Xero at any time, and it stops that moment. Exports from MYOB, QuickBooks and payroll systems are uploaded by you, read once, and the checks run over them the same way.

An owner can close the account at any time. For 30 days after closure the account stays read-only so you can export your records. After that we delete or de-identify your data within 60 days, keeping only what the law requires us to retain, such as tax and billing records, and de-identified computation and audit logs. The Privacy Policy sets out the detail.

Each promise above corresponds to something in the product.

• Residency: the infrastructure is defined in code that names the Sydney region, and the assistant's model routing is bound to a register of measured Sydney models.
• Personal details kept from AI: automatic checks in the build refuse AI provider code in the calculation layer, and refuse a raw identifier or a person's name in any text a model can read.
• Attested figures: every schedule carries its source, its date and a signed fingerprint; the engine refuses a schedule whose fingerprint does not match.
• Independent review: every change to the code is reviewed by a model from a different family than the one that wrote it before it can merge, and that review is not treated as compliance sign-off.
• Audit trail: each scheduled audit records the schedule version, the source and the working, so an answer can be reconstructed later.

This page describes the product as it ships today. The Privacy Policy covers this site, the free Health Check, News subscriptions and the product.

To access, correct or delete your information, or to ask how any of this works, write to hello@tobyai.io. A person reads it. TobyAI handles personal information in accordance with the Privacy Act 1988 (Cth). The status page reports the live health of the service.