The slow part of the season is not the return. It is wrangling the client's pile of documents into an organized, complete set: sorting the W-2 from the three 1099s, labeling each one, and figuring out what is still missing before you can even start. That sorting and listing is exactly the kind of work AI is good at. The trap is that the same tool that builds you a clean index in seconds will also decide a deduction is allowed, compute the return as if it were final, and answer the client's tax question, none of which is its call. The return goes out over your signature, not the model's. This is a plain look at where AI genuinely helps with document intake, where it does not, and a set of prompts you can paste in and test on your own cases.
One ground rule before any of it: these are taxpayer records with real PII. Keep Social Security numbers, account numbers, and other identifiers out of a public chatbot, test with de-identified documents, and for real client data use a private tool that keeps it inside your own systems. More on that at the end.
The honest picture
AI is genuinely useful for turning a pile of documents into an organized, complete set:
- What AI does well today: take a batch of client documents, sort them into a standard index (W-2, 1099-INT, 1099-NEC, 1098, K-1, and so on), label each one, transcribe the figures off them, and build a missing-items list by comparing what arrived against what the return needs. It compresses the intake, not the judgment.
- What AI does not do: prepare the return. It cannot decide a tax position, compute the tax owed as a final number, choose a filing position, or advise the client. The preparer owns every position and the signature. AI organizes and summarizes the documents the client provided and flags what is missing. The preparer decides every position and signs.
The right way to think about it: AI is a fast clerk that sorts the pile and tells you what is not in it, not a preparer. The organized set is a draft you review. The return is yours.
The line: it will decide a tax position if you let it
The specific failure to watch for is helpfulness, and it shows up four ways:
- Deciding deductibility or a filing position. Hand AI a stack and it will tell you a home-office expense is deductible or that the client should file head of household. Whether something is deductible, and which position is correct, is a judgment with rules and client facts behind it, not a document lookup. Keep it on organizing and flagging, never on deciding.
- Computing tax as final. It will happily total income and hand you a tax owed. A number it computed is not a prepared return, and treating it as final is how an error reaches a filing. Let your software and your review compute the return.
- Inventing a figure. If it misreads a box on a 1099 or cannot read a scan cleanly, it may fill the figure with a plausible guess rather than flag it. A transcribed number that is not actually on the document is a defect.
- Giving the client tax advice. It will answer "can I deduct this?" in a confident paragraph. That answer is yours to give, not the model's.
The fix is one discipline applied without exception: AI organizes and summarizes the documents the client provided and flags what is missing or unclear. The preparer decides every position, computes the return, and signs.
The setup that keeps the return with the preparer
Three habits make AI safe for intake, and the prompts below build them in:
- Give it a standard index. Tell it the document types and the index you sort into, and to use only those labels. A model handed your index sorts to it; a model left to invent categories produces a mess you have to remap.
- Make it flag, not decide. Require it to transcribe figures exactly, list what is missing, and surface anything ambiguous for you, while deciding no position and computing no final number. You want an organized set that tells you where to look, not a return.
- Keep the data yours. De-identified documents for testing in a public tool. Real client data only in a tool that stays inside your own systems.
How to test it on your own work
Do not trust a polished demo, including this one. Pull two or three of your own recent client sets, de-identified, a timer, and the prompts below. Rate each output 1 to 5 on usefulness and accuracy, check whether the transcribed figures match the documents exactly, and confirm its missing-items list catches what you would have caught. Compare the time against how you intake today. Keep what wins.
Paste-ready prompts
Copy these as written. Bracketed text is what you swap per client.
Test 1: Sort and label a batch into a standard index (text model)
I am pasting a batch of a client's tax documents. Sort and label only the
documents I paste into this standard index: [paste your index, e.g. W-2,
1099-INT, 1099-DIV, 1099-NEC, 1098, K-1, brokerage statement, other].
Rules:
- Use only the labels in my index. If a document does not fit, label it
"needs review" and say why. Do not invent a label.
- List what is present, one line per document, with its label.
- Decide no tax position and compute no final number.
- Flag missing documents and anything ambiguous for me. Do not guess.
Documents: [paste]
Watch for: did it sort to your index without inventing labels, and did it flag the document it could not place instead of forcing it?
Test 2: Build a missing-items checklist (text model)
Here is last year's return summary for this client: [paste the summary]. Here
are the documents that have arrived this year: [paste the labeled set]. Compare
the two and build a missing-items / follow-up checklist: which items appeared on
last year's return that I have not yet received this year, and any new items
that may need a document I do not have. Decide no tax position and give no
advice. Only list what to follow up on and why, so I can decide.
Watch for: does it catch the 1099 that showed up last year but not this year, without inventing items the client never had?
Test 3: Transcribe figures into a worksheet, exactly (text model)
Transcribe the key figures from the documents I paste into a worksheet, exactly
as they appear. Rules:
- Read only what is on the documents. Transcribe figures exactly, box by box.
- Do no math. Compute no totals and no tax. Decide no tax position.
- If a figure is unclear, unreadable, or could be one of two boxes, do not
guess: flag it under "unclear, needs my eyes" and quote what you see.
For each figure, note the document and the box it came from.
Documents: [paste]
Watch for: do the transcribed numbers match the documents exactly, and did it flag the smudged box instead of filling in a plausible figure?
Test 4: Fabrication and overreach audit (text model)
Review this organized set and worksheet against the documents I provided. For
each problem, quote the exact line.
1. Any tax position stated or implied (deductibility, a filing status, a
treatment) rather than just organizing the documents.
2. Any computed total or tax owed presented as a number.
3. Any figure that is not traceable to a document I provided.
Do not fix anything. Only flag. Treat any decided position, any computed total,
and any untraceable figure as a defect.
Documents I provided: [paste]
Organized set and worksheet: [paste]
Watch for: does it catch the line where it quietly decided a deduction or totaled the income? Run it on a set you already worked.
What success looks like, and where it could go
If your own testing shows real time savings, the next step is a small pilot: run a week of client intakes through the prompts and measure the hours against your usual setup. If that holds up, the natural next step is a simple agent, running on your firm's own cloud, that you use in plain language. The most useful version intakes and labels the documents to your standard index, transcribes figures exactly, builds the missing-items list against last year's return, and cites the source document behind every entry, so nothing on the worksheet is untraceable. The preparer reviews and signs. Because it runs in your own systems, the client's PII never leaves for a public chatbot.
The principle holds the whole way through: AI organizes the pile and tells you what is missing. It does not decide a position, compute the return, or advise the client. The preparer owns every position and the signature. Used this way it speeds the intake without ever putting a tax call in front of the client that you did not make.
This is general information about workflow tools, not tax, legal, or accounting advice.
Want a straight answer for your tax practice?
I build practical AI and custom software for businesses, on Google Cloud. If you want a second set of eyes on how AI could fit your document-intake workflow, or on a tool you are considering buying, tell me what you are working with. No pitch, just a straight answer.