TRANSPARENCYAI PRODUCT
MODEL SPEC
Last updated: April 21, 2026
This document describes how Lodgey's AI assistant works, what it can and cannot do, and how we handle your data. It is aligned with the Australian Government AI Transparency guidance and the Voluntary AI Safety Standard.
01.System Overview
ProductLodgey — AI-powered tax assistant for Australian property investors
PurposeOrganise tax evidence, surface potential deductions, prepare information for tax returns, and generate EOFY workbooks
OperatorLodgey Pty Ltd, Melbourne VIC, Australia
AI TypeGenerative AI (large language model) accessed via commercial API
Model ProviderAnthropic, PBC (San Francisco, US)
Default ModelClaude Sonnet 4.6
BYOK ModelsClaude Opus 4.6 / 4.7 (when user provides their own Anthropic API key)
Fine-tuningNone — Lodgey uses the model as-is via API, with no custom training or fine-tuning
02.Intended Use
Lodgey is designed to help Australian property investors:
- Upload and classify tax documents (bank CSVs, loan statements, settlement docs, depreciation schedules)
- Track rental income, deductions, and cost-base items across properties
- Calculate negative gearing, capital gains tax, and depreciation (Division 40/43)
- Prepare PAYG variation packs and EOFY tax summaries
- Organise an evidence pack that a registered tax agent can use to lodge
- Surface ATO key dates, thresholds, and compliance requirements
Important: Lodgey produces general information only — not professional tax, financial, or legal advice. All outputs must be verified with a TPB-registered tax agent before lodging with the ATO.
03.Out-of-Scope Uses
Lodgey is not designed or permitted to:
- Provide tax agent services or BAS services as defined under the Tax Agent Services Act 2009 (Cth)
- Provide personal or general financial product advice under the Corporations Act 2001 (Cth)
- Recommend whether to buy, sell, refinance, or restructure property or loans
- Advise on superannuation, SMSF, insurance, shares, managed funds, or crypto
- Sign off on a tax position for ATO lodgement
- Guarantee a specific refund, saving, or ATO outcome
- Provide credit assistance under the National Consumer Credit Protection Act 2009 (Cth)
- Provide legal advice
When users ask about these topics, Lodgey is instructed to decline and redirect to the TPB public register or an AFSL-licensed adviser.
04.How Lodgey Produces Responses
When you send a message, the following happens:
- PII redaction: A server-side redactor masks Australian TFNs (mod-11 check), ABNs (mod-89), Medicare numbers (mod-10), BSB + bank account numbers, and credit card numbers (Luhn) before any data leaves our servers. Your saved copy retains the originals.
- Context assembly: Your message, relevant conversation history, project data, and domain knowledge files (tagged by financial year) are assembled into a prompt.
- Model inference: The prompt is sent to the AI model provider via their commercial API. The provider processes it and returns a response. No conversation data is stored on the provider's infrastructure.
- Tool execution: The model may invoke Lodgey-defined tools (read/write project data, manage evidence, search the web, run calculations in a sandboxed environment) and opt-in integrations (Gmail, Drive, Xero).
- Response delivery: The response is streamed back to you. All interactions are logged in our database for audit and contestability.
05.Data Handling & Infrastructure
DatabaseNeon Postgres — Sydney AU region (ap-southeast-2, AWS). Encrypted at rest, TLS in transit.
HostingVercel — edge compute and serverless functions (US).
AI inferenceAnthropic Messages API (US) — chat messages sent after server-side PII redaction. No session storage on provider infrastructure.
AI observabilityLangfuse (EU, Frankfurt) — session traces for debugging and quality monitoring. GDPR, SOC 2, ISO 27001 compliant.
Product analyticsPostHog (cookie-consent gated); Vercel Analytics (privacy-focused, no cookies).
AuthenticationNeon Auth — session-based, HttpOnly cookie.
Training data useNone. Our commercial API agreement prohibits the provider from using API inputs to train models.
Data retentionConversation history is retained in our database for the life of your account. You can request deletion at any time.
BYOK isolationIf you provide your own API key, your data is sent directly under your own agreement with the model provider. Your key is encrypted at rest in our database.
06.Known Limitations
- Hallucinations: Large language models can generate plausible-sounding but incorrect information, including fabricated case references, incorrect tax thresholds, or misattributed ATO rulings.
- Arithmetic errors: While Lodgey uses a sandboxed code execution environment for calculations, the model can make logical or arithmetic errors in its reasoning.
- Outdated training data: The model's training data has a knowledge cutoff. Tax thresholds, rates, and ATO guidance change frequently. Lodgey uses live tools (web search, ATO/RBA/ABS APIs) to mitigate this, but gaps can occur.
- Document misinterpretation: OCR and document parsing can misread figures, dates, or categories — particularly with poor-quality scans or handwritten annotations.
- Omissions: Lodgey may fail to surface relevant deductions, compliance requirements, or edge cases that a qualified tax professional would identify.
- No demographic bias evaluation: We have not conducted formal bias testing across demographic groups. The model's outputs may reflect biases present in its training data.
07.Human Oversight
Human-in-the-loopNo. Lodgey's responses reach you without human review.
Post-hoc reviewEvery AI message includes a flag button. Flagged outputs are queued for human review by our team.
ContestabilityYou can dispute any output via the in-app feedback form or by emailing hello@lodgey.io.
MonitoringWe test Lodgey against a curated set of Australian tax scenarios and publish material changes to its behaviour.
08.Regulatory Alignment
- Tax Agent Services Act 2009 (Cth): Lodgey is not registered with the TPB. It is a software tool under TPB(I) 09/2011 Example 5 (non-customised software with tax calculators). Outputs are general information, not tax agent services.
- Corporations Act 2001 (Cth): Lodgey does not hold an AFSL. It does not provide financial product advice (s 766B). It does not consider your personal objectives, financial situation, or needs.
- Privacy Act 1988 (Cth): APP 8.2(b) consent collected at signup for overseas data disclosure. PII redacted before transmission. See Privacy Policy and Privacy Impact Assessment.
- Australian Consumer Law: Lodgey does not make representations about guaranteed outcomes. Calculators and estimates are clearly labelled as such.
- TPB(I) D62/2026 (draft): While this guidance primarily binds registered tax practitioners, Lodgey's consent-based model (explicit opt-in for data sharing, per-session integration permissions) aligns with the TPB's expectation of client-informed AI use.
- DISR Voluntary AI Safety Standard (Sep 2024): This Model Spec, together with our Terms, Privacy Policy, and PIA, addresses the Standard's 10 guardrails including accountability, transparency, contestability, and human oversight.
09.Agentic Capabilities
Lodgey can autonomously execute multi-step actions on your behalf. These include:
- Searching your Gmail for tax-relevant receipts and statements (opt-in per session)
- Reading and writing files in Google Drive and Sheets (opt-in per session)
- Filing evidence against deduction categories in your evidence pack
- Running calculations in a sandboxed code execution environment
- Searching the web for current ATO thresholds and rulings
- Updating your tax situation record with new facts from conversations
Safeguards: External integrations (Gmail, Drive, Xero) require explicit opt-in per session and are revocable at any time. Destructive actions (deleting evidence, overwriting data) require confirmation via an in-chat prompt before execution. All tool invocations are logged for audit.
10.Feedback & Contact
If you have questions about this Model Spec, want to report an issue, or need to exercise your rights under the Privacy Act:
- Email: hello@lodgey.io
- In-app: Use the flag button on any AI message or the feedback form in the chat menu