How AI is used with my info

This description matches the code in this repository. It has not had professional review.

Darian reviews. AI never files.

Anthropic's standard API retention applies to prompts and outputs sent from Ascend. Zero data retention has not been confirmed for this firm.

Document reading

What is sent. The file itself (PDF, PNG, JPEG, GIF, or WebP), the file name, and today's date, when the file is 8 MB or smaller. Files over 8 MB, and other types such as HEIC, are stored for a person to review and are not sent to the model.

Who receives it. Anthropic Messages API, model claude-sonnet-4-6.

What comes back. A JSON result: document type, tax year, a short summary, up to six extracted fields, a deductibility label (yes, maybe, no, or na), a short note, and a flag. The prompt tells the model not to put a Social Security number in the result. A matching W-2, 1099, or 1098 can also save one fact in client_memory.

Limits. 25 document reads per signed-in user per day.

Human review. Darian reviews. AI never files.

Assistant chat

What is sent. The recent conversation (up to the last 14 messages), up to 40 facts already saved in client_memory, savings goals on the file, and a Live Tax Bill planning snap when the portal sends one. The document file is not attached to this chat.

Who receives it. Anthropic Messages API, model claude-sonnet-4-6.

What comes back. A planning reply. The assistant can save one short fact (Social Security numbers and long account-number strings are blocked), ask for a document, create a task for Darian, or post a note in the message thread.

Limits. Daily message caps in the current code: 40 on the free tier, 200 on Plus or Sentinel, 400 on Pro or Summit. A signed-out portal call is refused.

Human review. Darian reviews. AI never files.

Plaid classification

What is sent. Up to 40 bank transactions that do not yet have an AI category. Each item sent is the name, merchant, amount, and Plaid category. The bank access token and account numbers are not in that payload.

Who receives it. Anthropic Messages API, model claude-haiku-4-5-20251001.

What comes back. A category label, whether it looks like business spending, a possible-deduction hint, and a confidence number. Those fields are saved on the transaction.

Limits. 40 transactions per classification call. Plaid's environment defaults to sandbox when PLAID_ENV is unset. A sandbox connection is not a live production bank link.

Human review. Darian reviews. AI never files.

Advisor phrasing

What is sent. Titles and details that deterministic detectors already computed from bank transactions and savings goals (subscriptions, fees, goal pace, spending changes). Dollar figures, counts, and dates are computed before the model sees them. Only that title and detail text is sent for warmer wording.

Who receives it. Anthropic Messages API, model claude-haiku-4-5-20251001. If the call fails, the deterministic wording is kept.

What comes back. Rewritten title and detail text. Dollar figures are checked so a figure is not dropped. The result is written to the client feed.

Limits. Planning language only. The advisor consent check exists in code and defaults off. This page does not turn that check on.

Human review. Darian reviews. AI never files.

TaxCopilot (admin only)

What is sent. Clients do not use this feature. A client session that asks for it receives no firm-chat answer. The server replies that TaxCopilot did not run because it needs a live admin session. When an admin runs it, the request can include the question, pasted letter text (up to 12,000 characters), a briefing, and a server-loaded client file: name, plan tier, return stage, saved facts, document summaries, open tasks, and recent messages.

Who receives it. Anthropic Messages API, model claude-sonnet-4-6.

What comes back. A preparer-facing draft. It does not file a return and it does not message a client by itself.

Limits. Admin session required. Fail-closed: without that session, TaxCopilot does not answer.

Human review. Darian reviews. AI never files.

Public site chat

What is sent. The visitor's recent messages on the public website (up to the last 10, each truncated). No client file, no documents, and no bank data are attached.

Who receives it. Anthropic Messages API, model claude-sonnet-4-6.

What comes back. A general tax answer for a website visitor, with a suggestion to book or create an account when that fits.

Limits. 10 messages per minute per IP, and 50 per IP per day. The assistant is told not to ask for Social Security numbers or account numbers.

Human review. Darian reviews. AI never files.

Facts saved about you

Facts the assistant or document reading saves are stored in client_memory. In the signed-in portal, What Ascend knows lists each fact and deletes that one row when you remove it. Deleting a fact does not delete an uploaded document or a filed return. Ascend has not confirmed that Anthropic deletes its own API copy.