This page describes how the NGTAX.AI platform uses artificial-intelligence ("AI") models, as designed for its initial release; some of the controls described are in development. It is a summary for transparency and does not modify any Service Agreement.
1. Which models we use
We use large language models provided by Anthropic, PBC — Claude Sonnet and Claude Opus — accessed directly through the Anthropic API. We assign each role to a pinned model version, and escalate to a person when a response cannot be validated or the task requires a decision. The exact model version used for each judgment is recorded. Model assignments may change as models change.
2. What models do and do not do
Models are used to classify transactions, read documents, draft questions, identify potential issues, propose tax treatments with supporting authority, and draft resolution options for a professional to choose among.
Models do not: compute any amount reported on a financial statement or tax return (all amounts are computed by deterministic software); choose a filing position; approve, sign or transmit a return; or contact any taxing authority.
3. Human oversight
Every return is signed as paid preparer by a US certified public accountant: the taxpayer's own CPA, or a CPA engaged through NGTAX.AI. The company's officer signs as taxpayer. No return is approved, and no finding is resolved, without the decision of a named person, which is recorded with that person's identity and the time.
4. Records
For each model judgment we record the model version, an identifier and hash of the instructions used, a digest of the input, the validated output, and the time. These records support review, reproduction of the return, and response to inquiries.
5. Client data
Client return information is used only to prepare that client's return and as otherwise permitted by Internal Revenue Code §7216 and its regulations, including with any consent they require. We do not use client return information to train AI models. Client data is sent to a model provider only under commercial terms we have approved for that purpose, and the platform is designed so that a firm can disable model processing for any client. Our development and evaluation use synthetic data.
6. Limitations
AI models can produce incorrect, incomplete or inconsistent output. Our controls — schema validation, authority checks, confidence thresholds, deterministic computation, independent review and professional sign-off — are designed to reduce, not eliminate, the effect of such errors. Content in uploaded documents and transaction descriptions is treated as data and not as instructions to the models.