Document requests are one of the most obvious places accountants notice AI.
The work is repetitive. The notes are messy. The same client sends half the items, answers inside an old thread, attaches screenshots instead of statements, and asks what is still missing. Someone has to turn that mess into a clear next message.
AI can help prepare that message.
It should not decide what documents are required, send the client communication, or invent the reason an item is needed.
The useful workflow is a controlled loop: approved missing-item notes in, AI-prepared draft out, human review before anything reaches the client.
The safer document request loop
1. Start with an approved tracker
Do not ask an assistant to “figure out what the client still needs to send” from a broad email history.
Start with a tracker or reviewer note that already names the missing items. The assistant can help organize those notes, but the professional owns whether the list is complete and correct.
Good source material:
- Approved missing-item notes.
- A sanitized checklist.
- Client status labels from a practice-management system.
- Prior reminder language that has already been approved.
- Internal tone rules.
Riskier source material:
- Full private email threads.
- Client portal screenshots.
- Tax documents, payroll records, bank statements, or client files in a tool that has not been approved for that data.
- A vague prompt asking AI to decide what is missing.
2. Sanitize before drafting when possible
For many reminders, the assistant does not need private client details. It may only need item labels, deadline language, tone guidance, and the approved next step.
Instead of pasting a full client thread, provide a cleaned source packet:
| Field | Example |
|---|---|
| Client label | Client A |
| Missing item | January bank statement |
| Status | requested once, not received |
| Approved tone | firm but helpful |
| Human note | do not mention penalties or tax treatment |
That structure makes the AI task narrower and the review easier.
3. Ask for a draft plus reviewer notes
The assistant should prepare two things:
- A client-ready draft for human review.
- A reviewer note that lists assumptions, missing context, and items that should be checked before sending.
This prevents the polished message from hiding uncertainty.
4. Review before sending
The human reviewer should check:
- The missing items match the approved tracker.
- No extra document request was invented.
- No tax, accounting, legal, payroll, audit, compliance, financial, or client advice slipped into the message.
- The tone matches the relationship.
- Dates, portal instructions, and next steps are correct.
- Private details stayed inside the approved data boundary.
5. Keep the send action human-owned
Drafting is not sending.
Even if an assistant can operate near email, a portal, a browser, or a practice-management system, the review gate should remain explicit. The accountant or approved staff member decides what gets sent and when.
A draft-only prompt pattern
Use this as a page example, not as a required process:
Using only the approved missing-item notes below, draft a concise client reminder. Do not add new document requirements, tax advice, accounting advice, legal advice, deadline consequences, or unsupported explanations. Separate the client message from a reviewer note listing assumptions and checks before sending.
The power of the prompt is not the wording alone. It is the workflow around it: approved source, data boundary, output format, prohibited content, and review gate.
Why this makes a good first AI pilot
Document requests are a strong first pilot because the task is narrow and reviewable. The practice can compare drafts against source notes, track edits, measure whether reminders become clearer, and stop if the workflow creates confusion.
The pilot should not begin with live client records in a connected tool. Start smaller: sanitized examples, internal notes, and a human reviewer.
FAQ
Can AI decide which documents to request from a client?
This page does not recommend that. The safer pattern is for a human or approved process to identify missing items, then use AI to prepare a reviewable reminder from those approved notes.
Can AI send the reminder?
This page keeps sending human-owned. Drafting a reminder and sending a client communication are different actions with different risk.
What if the client email has all the context?
That may be exactly why the data boundary matters. A full email thread may include private client information, unrelated facts, or sensitive attachments. Use approved tools and sanitized source packets when possible.