Can ChatGPT answer an IRS notice for a client?
Pasting a CP2000 into a general AI chat assistant and asking for a response is an obvious first experiment. The result reads well. Whether it is right is a separate question, and for a notice that proposes extra tax, right is the only thing that matters. This guide sets out what the job takes, where a chat assistant helps, where it breaks, and when it is a perfectly reasonable choice.
What does answering a notice really take?
A CP2000 comes from the IRS underreporter program. It lists income or payments that third parties such as employers, banks, brokers or clients reported to the IRS and that do not match the filed return, and it proposes a change to the tax. It is a proposal, not a bill. A good response does five things in order:
- Lists every item on the notice with its amount.
- Finds where the filed return reports each one, or confirms that it does not.
- Decides a position per item: agree, disagree, or agree in part.
- Shows the comparison as a numbered schedule, to the cent.
- Writes a short letter that points to the schedule and the evidence.
Only step 5 is writing. Steps 1 to 4 are reading, locating and arithmetic, and that is where most errors enter. A state notice of proposed assessment follows a similar pattern, with its own forms and rules.
Where does a chat assistant help?
- Drafting the prose. A letter opening, a polite close and a clear sentence per position are quick for a language model and easy for you to check.
- Explaining a notice. Asking what a section of the notice means, or how to structure a response, is a fair use, as long as you verify anything it says about a rule.
- Tidying your own draft. Shortening a paragraph or fixing a tone is low risk, because the facts came from you.
Where does it break down?
Each problem below is about the approach, not about one product.
- Arithmetic. Unless the assistant runs code for the sums, a language model produces numbers as text rather than calculating them. It can add a column wrongly or round a figure without saying so, and the answer looks just as confident. A schedule that is off by a cent still does not add up, and that weakens the whole letter.
- Dropped pages and items. A long notice with a full return attached is a lot of text for one chat, and the tool may summarize or skip part of it. An item on page 4 that never makes it into the answer leaves no trace.
- No source for each figure. Unless you ask for the page and line every time and check it, you cannot tell whether a figure was read from the return or produced from nothing.
- Cited authority that is not there. A model may name a code section or procedure that does not exist, or describe a real one wrongly. Cite only what the notice itself prints or what you have looked up.
- Different answers on different runs. Ask twice and you may get two schedules. That is awkward when a client or a supervising partner asks why the letter says what it says.
- No second check by default. Nothing reviews the answer again or lists what stayed uncertain unless you ask, and then the same tool is checking its own work.
- Client data. Where the text goes and how it is kept depends on the account and settings you use. Check before you paste.
How do a chat assistant and a purpose-built tool compare?
| What you need | General AI chat assistant | NoticeBrief |
|---|---|---|
| Matching each notice item to the return | Depends on your prompt and the file | Each item looked for on the return and quoted with its page; gaps listed |
| Arithmetic | You check it yourself | Computed by code to the cent, not by a model |
| Where a figure came from | Only if you ask, and then you verify it | Code checks every quote against its page |
| Output | Mostly text you copy into your own documents | Letter (Word), Exhibit A (PDF and CSV), result (JSON) in one packet |
| Second check | You | A second, stronger AI model, with leftovers listed first |
| Notice types | Anything you can describe | CP2000, CP504 and state notices of proposed assessment get a position; other kinds get Exhibit A and a list to check |
When is a chat assistant enough?
- You have one short notice with one or two items and the arithmetic is trivial.
- You only want help with wording, and you supply the figures and positions.
- Your firm already has a policy for AI tools and the data in this file is allowed under it.
In these cases a chat assistant plus a spreadsheet for the sums is a fair, cheap workflow. For the manual method step by step, read how to respond to a CP2000 notice and the response letter checklist.
How to use a chat assistant safely
- Remove names and identification numbers you do not need.
- Do the sums outside the chat and paste only the results.
- Ask for the page and line of every figure, then open the page and look.
- Count the items on the notice yourself and confirm each appears in the answer.
- Delete any authority you cannot find in the notice or in a source you trust.
- Run the finished letter against a checklist before you sign it.
How NoticeBrief does this step
NoticeBrief does steps 1, 2 and 4, proposes a position per item for step 3, and drafts step 5. You upload the notice and the filed return, and a payment record if one matters, as PDFs with a text layer, Word files or text files, up to 6 files and 80 pages in total. Claude reads the pages and quotes each figure with its page. Code then checks every quote against that page, parses the figures and computes Exhibit A to the cent. No model adds or compares a figure. The letter is drafted from the statuses code computed, and you decide every position.
A second, stronger AI model reviews every result, fixes what it can, and lists what is left for you to decide. Results are model output checked by code and by a second model; they are not verified by a person. You get a packet with the letter, Exhibit A and the open items first, plus the letter as a Word file, the schedule as a CSV and a JSON result.
NoticeBrief does not read scanned paper notices, does not send or file anything and does not give tax or legal advice. You add the date, the taxpayer's identification and your signature, and you read every position. One notice answered is 29 credits, with no subscription. Sign up to try it.
Questions
- Is it a problem to paste a client's notice into a chat assistant?
- It can be. Before you paste anything, check your engagement letter, your firm's data policy and the rules that apply to preparers handling client information. If you are not sure, redact names and identification numbers first, or do not use the tool for that file.
- Can a chat assistant at least draft the cover letter?
- Yes, drafting prose is what it is good at. Give it your own facts and positions, and keep the figures out of its hands. Read every sentence, because it may add a claim you never made.
- Does a better prompt fix the arithmetic problem?
- A careful prompt helps with tone and structure but does not make the sums reliable. Some assistants can run code for arithmetic when asked; unless you can see that it did, compute every sum and difference outside the chat, in a spreadsheet or with code, and paste the results in.
Answer the notice with every figure compared.
Sign in, upload the notice and the return, download the packet.