AI document assistant
The AI document assistant is included in every plan. It is off until a firm owner or admin turns it on, and it can be turned off at any time.
What it does
- Names and tags client uploads. When a client uploads a file, the assistant suggests the form type (W-2, 1099-NEC, 1098-T, K-1, brokerage statement and so on), the tax year and the issuer, and a clean name such as "2025 W-2 - Acme Corp.pdf". Staff see "Suggested by AI" and click Accept name or Dismiss. An accepted name can be undone.
- Matches uploads to open requests. If the client has an open request, the assistant suggests which item the file answers. Staff click Confirm match and the file is attached to that item and marked done.
- Drafts a checklist from a prior-year return. On a client's AI tab, pick last year's return (PDF). The assistant lists the documents to expect next season, with issuer names. You edit the list and send it as a normal client request.
Everything runs in the background. Nothing is renamed without a click, unless an admin turns on "Rename files automatically".
Exactly what is sent
| When | What is sent | What is not sent |
|---|---|---|
| A client uploads a file | The text of the PDF (first 15 pages; scans and photos are sent as page images), and the labels of the client's open request items | Client name, email, notes, other documents |
| You draft a checklist | The prior-year return you picked | Anything else on the client |
- Before text is sent, numbers that look like a Social Security number or an EIN are masked.
- Files over the page limit, password-protected PDFs and file types the assistant can't read (Word, Excel, ZIP) are skipped.
- Only uploads made by clients are processed. Staff uploads are not.
To whom
The Claude API of Anthropic, PBC, a US company. Processing happens in the United States. Anthropic acts as our service provider under contract.
Training and retention
- No training. Anthropic does not train its models on data sent through its API.
- We keep only the result: the suggested name, form type, tax year, issuer and a confidence score, plus token counts for usage. We never store the document text that was sent.
- Anthropic keeps API inputs and outputs only for its standard short retention window, to provide and secure the service, and then deletes them.
- Every AI action is written to your firm's audit log (who turned it on and when, each document processed, the model used and the token counts).
Why your clients don't need to sign a consent
The assistant is used only to perform your firm's work on that client's return, by a US provider bound by contract, with no training and no retention beyond processing. Under Treas. Reg. 301.7216-2(d) this is a disclosure to an auxiliary service provider, which needs no taxpayer consent. You still decide: it stays off until you turn it on. If your firm's policy or a client agreement says otherwise, leave it off.
Turning it on
- Go to Settings > Integrations > AI document assistant (Claude).
- Read what is sent, tick the box and click Turn on the AI document assistant.
- Choose the features you want. Each one has its own switch.
Turning it off
Go to Settings > Integrations > AI document assistant (Claude) and click Turn off the AI document assistant. From that moment nothing more is sent. Suggestions already made stay visible, and you can still accept, undo or dismiss them. Turning a single feature off works the same way.
Limits and usage
- Each firm can process up to 500 documents a day (server setting). Files over the limit are skipped and show as "Skipped".
- Admins see the documents processed, input and output tokens and an estimated cost for the last 30 days on the settings page.
- AI can be wrong. Treat every suggestion as a draft: check the name and the match before you confirm.